1 s2.0 S037851732300234X Main

International Journal of Pharmaceutics 636 (2023) 122814
Contents lists available at ScienceDirect
International Journal of Pharmaceutics

journal homepage: www.elsevier.com/locate/ijpharm
Implementation of a fully integrated CM direct compression and coating

process at a commercial pharmaceutical facility – Part 2: PAT and RTD
results for normal operational conditions batches
Juan G. Rosas a, *, 1, Peter Brush b, Bruce Thompson b, Charles Miller b, 2, Paul Overton c, 1,
Neil Tugby c, 1, Daria Stoliarskaia a, 1, Samantha Hurley d, Manoharan Ramasamy b, Stephen
L. Conway c
a
MSD, Pharmaceutical Technical Operations PAT, UK
b
Merck & Co. Inc, Analytical Chemistry in Development and Supply PAT, United States
c
MSD, Pharmaceutical Technical Operations, UK
d
Merck & Co. Inc, Pharmaceutical Commercialization Technology, United States
A R T I C L E I N F O A B S T R A C T
Keywords: This is the second of two articles detailing the continuous manufacturing (CM) development and implementation
Continuous manufacturing activities for an marketed product which have been realized in novel, qualified equipment, using validated
Near-Infrared spectroscopy control strategy elements to enable manufacture of batches under current good manufacturing practices (cGMP)
Partial least squares regression
and compliant with data integrity principles. Here, the application of process analytical technologies (PAT) and
Real time release testing
Residence time distribution
automation tools on batches produced under normal operational conditions is reviewed. The results from resi
dence time distribution (RTD) models for predicting API concentration, in-line near infrared (NIR) testing of
blend uniformity (BU) and at-line NIR spectroscopy analysis of core tablet concentration and tablet identity for
real-time release testing (RTRT) are discussed. The influences of process equipment and design choices on NIR
and RTD model variability, as well as the use of the PAT tools for monitoring the evolving properties under
standing of CM process development, such as overcoming flow instabilities, is described. Results demonstrate
that the RTD and NIR models developed and validated are robust to operating conditions and are critical for
assuring steady state control of the continuous manufacturing process. Finally, the NIR and RTD model lifecycle,
including procedures for necessary and normal model upgrades in a cGMP production environment, are
presented.
1. Introduction and operational philosophy associated with a product, it can be used to

realize lower-cost and smaller footprint manufacturing with more rapid
Since the first oral solid dosage (OSD) form new drug application development. The Food and Drug Administration (FDA) is supporting
(NDA) for manufacturing purely via continuous manufacturing product and encouraging the adoption of innovative manufacturing technologies
was approved in 2015 (Badman et al., 2019; Gregory et al., 2017; Nasr including CM by using a scientific and risk-based approach and they
et al., 2017; Wahlich, 2021), the pharmaceutical industry has acceler recommend pharmaceutical companies have early and active discus
ated evaluation of the CM approach for drug product development and sions with the agency during CM implementation.
manufacture. As is well-known, CM offers many advantages over Nowadays many pharmaceutical companies and contract manufac
traditional batch manufacture, including production flexibility, turers are evaluating CM strategies, in particular examining the devel
manufacturing efficiency and improved product quality assurance. opment and manufacture of OSDs using such technology (Vanhoorne
Furthermore, if the CM system is designed to address the specific risks and Vervaet, 2020). Likewise, equipment manufacturers have begun to
* Corresponding author at: Organon, Shotton Lane, Cramlington NE23 3JU, UK.
E-mail address: [email protected] (J.G. Rosas).
1
Current affiliation Organon & Co.
2
current affiliation Aspen Technologies.
https://doi.org/10.1016/j.ijpharm.2023.122814
Received 29 October 2022; Received in revised form 2 March 2023; Accepted 5 March 2023
Available online 12 March 2023
0378-5173/© 2023 Elsevier B.V. All rights reserved.
J.G. Rosas et al. International Journal of Pharmaceutics 636 (2023) 122814
provide commercially available integrated solutions, such as the and GEA (Holman et al., 2021). The aim of the study was to demonstrate
continuous tableting line from GEA Pharma with their ConsiGma™ the robustness of the CDC-50 system running continuously under rele
coaters, and now numerous integrated continuous tableting lines are vant process conditions for over 120 h.
available on the market such as those supplied by Glatt, Lödige or L.B. To be successful in the implementation of NIR spectroscopy in CM for
Bohle among others (Vanhoorne and Vervaet, 2020). Alternatively, RTRT is key to define three main components about NIR methods:
many pharmaceutical companies have also developed in-house solu • The NIR method, which includes key attributes that enable the
tions, integrating equipment from various suppliers into CM lines NIRS measurement of the analyte of interest and are not expected to
(García-Muñoz et al., 2018). However, most of these CM lines are mainly change over product lifecycle such as the equipment and spectrometer
focused on R&D activities and it would be important to accelerate the type (i.e., NIR), sample measurement interface (i.e., diffuse trans
implementation of CM lines in commercial manufacturing environment mission), data collection parameters (number of scans, resolution),
for the benefit of patients. calibration algorithm (i.e., Partial least squares, PLS) and model vali
MSD has installed and qualified the first CM line in a GMP facility dation criteria (EMA, 2014).
based on GEA Pharm Systems ConsiGma™ direct compression (CDC-50) • The NIR model, that is the mathematical engine within the NIR
and ConsiGma™ coater technologies for a legacy product that is method that describes how the NIR spectra are related to the analyte
currently being manufactured by a traditional batch process. In this property of interest (EMA, 2014; FDA, 2015b; USP38, 2015). Model
case, the batch process is supported by the first commercial application attributes include calibration set composition, NIR spectral wavelengths
real time release testing (RTRT) using at-line NIR analysis. A key utilized, spectral pre-treatment algorithms, and the number of principal
advantage of having a CM process being built upon the foundation of components (rank) of the model. Hence, these NIR model attributes may
using PAT to monitor blend uniformity (BU) and to perform automated be updated over the lifecycle of a product to include new variabilities
NIR testing of core tablets and residence time distribution (RTD), is to that could influence the NIR model that were not captured in the original
eliminate the requirement for manual at-line testing. A key to this suc model calibration/validation set. The model is validated against pre
cessful outcome has been the participation of a cross-functional team of defined performance criteria.
development and process engineers; operations and supply chain staff; • Lastly, the lifecycle management program, which is already part of
analytical, PAT, and statistical/modelling specialists; equipment and company’s quality system, is used to provide a detailed and standardized
software vendors; along with inclusion of representatives from quality listing of activities and documentation required for model verification
and regulatory functions early in the design process. and update activities.
This is the second of two articles detailing the development and The proposed approach to manage the NIR method can be viewed as
implementation of CM line in a supply manufacturing site to enable a performance-based approach to the NIR model portion of the method,
manufacture of batches under good manufacturing practices (GMP) and utilizing an analytical target profile (ATP). While “performance-based
compliant with data integrity principles. This CM line has been evalu approach” is a new term utilized in ICH Q12 (ICH, 2019a, b), it is not a
ated for an alternative manufacturing process for a marketed product. new concept. The idea of describing analytical methods based on their
Part 1 of this series described the development of the integrated control performance criteria has been around for at least a decade and was
strategy meeting MSD’s needs for batch size flexibility and real-time acknowledged in the FDA-EMA parallel assessment pilot Q&A (FDA-
monitoring, automated rejection/diversion and release testing for a EMA, 2013) that an ATP “can be acceptable as a qualifier of the expected
CM process converted from the approved batch process, and the inte method performance”. In a performance-based approach, updates to the
gration with existing commercial infrastructure. In this article the NIR model are performed within the pharmaceutical quality system
application of PAT and automation tools for normal operational condi (PQS) without regulatory reporting as long as the performance criteria
tions (NOC) batches is described. are maintained. Changes to the NIR method (technology, configuration,
The proposed control strategy includes full integration of the CDC-50 etc.) or the model’s performance criteria would initiate regulatory
and ConSigma coaters with existing infrastructure at a commercial fa reporting.
cility and differs from other CM lines reported in the literature (Van The advantage of using a performance-based approach on the NIR
hoorne and Vervaet, 2020). To facilitate future commercial control model is that it allows the manufacturer to rapidly deploy model updates
strategy capabilities beyond the needs of the initial CM product, BU and assure continuity of supply. It also allows the analytical measure
monitoring by NIRS was applied for development of the RTD process ment to continually adapt such that the most appropriate measurement
model and NOC batches. is being used throughout the product and process lifecycle. Simplifying
In the pharmaceutical industry NIRS is being used extensively as a the regulatory reporting lowers the regulatory barriers for use of PAT
qualitative and quantitative PAT tool due to its well-known advantages technologies and supports continual improvement efforts. Continued
(e.g., fast technique, safe, non-destructive, sample preparation is not performance of the PAT method and quality of the manufactured
required, etc.). In the last decade, several studies have been performed product are assured through achieving the predetermined model per
applying spectroscopic techniques in CM such as NIR and Raman spec formance criteria and monitoring within the PQS. In summary, the
troscopy for BU upstream of the press (Järvinen et al., 2013; Palmer performance-based approach defined to manage the NIR model and
et al., 2020; Shi et al., 2016; Vanarase et al., 2010; Vanarase et al., 2013; method ensures that the validation criteria for the initial validation and
Vargas et al., 2017; Vargas et al., 2018, Sierra-Vega et al., 2019) or in the any subsequent re-validation efforts will be used to verify method per
tablet press frame (De Leersnyder et al., 2018; De Leersnyder et al., formance. The NIR performance criteria also serve as the verification
2019; Pauli et al., 2019) and NIR for API content uniformity (Järvinen and revalidation criteria for the model. The initial model and any model
et al., 2013; Palmer et al., 2020). Although, the concept of RTD to model updates are required to meet these criteria.
CM processes in the chemical industries is well-known since the middle The objective of this article is to describe the full PAT implementa
of the last century (Levenspiel, 1993), the use of RTD in CM by the tion with CM line allowing automatic prediction of API concentration by
pharmaceutical industry has been relatively recent reported (Holman RTD and NIR models enabling real-time product release. We also show
et al., 2021; Hurley et al., 2022; Palmer et al., 2020; Vanhoorne and how in-line NIR testing of the formulated blend can support RTD process
Vervaet, 2020) and its application is increasing. Different control stra model development, as well as provide alternative monitoring to
tegies for CM have been discussed in the literature where different ap maintain control of tablet assay. The incorporation of these elements
proaches to assess quality of CM batches are described (Almaya et al., into a commercially viable process is described, as well as results from
2017) and different statistical methods applied in CM have been initial CM production batches manufactured at normal operational
described in the literature (Cárdenas et al., 2020). One recent study conditions.
using a GEA CDC-50 has been reported in collaboration between MSD
2
2. Materials and methods hierarchy and automation sequences to ensure the final product quality
of the product manufactured by CM.
2.1. Pharmaceutical formulation
2.3.1. In-line NIR spectroscopy
A film-coated tablet product manufactured from a traditional batch Toward the assessment of in-line BU, the blend NIR spectra were
direct compression process was used as business case for the use of CM as collected in reflection mode by using a diode array spectrometer Sen
an alternative manufacturing process to the batch process. The formu troPAT FO (Sentronic, Dresden, Germany); for simplicity, this is called
lation comprises 32 wt% active pharmaceutical ingredient (API) sup blend NIR from now on. The covered spectral range was 1100 – 2100 nm
plied in tablets of 25 mg, 50 mg and 100 mg strengths. The formulation with a spectral resolution of approximately of 12 nm. To ensure spectra
contains five other excipients (referred to here as EX1 to EX5). The core are representative of the blend flow the SentroProbe DR MS7 with seven
tablet target weights were 100, 200 and 400 mg for the three strengths inline windows (Sentronic, Dresden, Germany) was used, which is fully
with a targeted label claim (%LC) of 250 mg/g API/tablet (LC = 100%). integrated with the GEA supplied Lighthouse ProbeTM. The system is
All three film coated tablets strengths share a common blend formula configured to gather ten average scans in less than 1 s and provide a
tion and are compressed into tablets that are weight multiples of each single sample spectrum every 5 s during manufacturing.
other. The film coated tablets are differentiated by weight, size, The NIR probe and housing are located immediately above the tablet
debossing, and film coating colour. press: between press and the blender outlet as shown on Fig. 1a. The
probe housing is capable of automated probe cleaning and reference
2.2. CM line for continuous direct compression and coating checks during manufacturing operation. The entire blend NIR system is
fully integrated with the CDC-50 and the existing PAT management
The implementation of CM line is part of a broader strategy that system (PMS) (SIPAT, Siemens). Care was taken in the design of the
promotes developmental and production flexibility, reduced lead times system to ensure the optimal presentation of the blend to the NIR probe,
and greater responsiveness to changing market demands for medicines. where the powder level is controlled by a powder dosing valve (PDV)
Hence, the main advantage of CM over batch manufacturing in the (Fig. 1b). Close to the PDV is located the material inlet hopper vibrator
commercial environment is the ability to produce drug products with a which starts vibrating when the hopper is 5% filled and stops when
flexible batch size. To demonstrate this, MSD worked with GEA to hopper is filled at 95% to facilitate the blend flows and avoid bridging of
design, install and qualify a CDC-50 and ConSigma coating line (GEA powder above the press inlet. High frequency NIR Spectra are recorded
Pharm, Germany) allowing continuous direct compression and coating only when sufficient product stream is detected via the level sensor
of an oral solid dosage drug product. This system is configured to located in the dropline between the blender outlet and the press
combine unit operation steps for material dispensing and loss in weight (Fig. 1c). Therefore, by automatically pausing the PAT collection
feeding (LIWF) of up to six individual components, continuous blending method, no measurements are taken when product is unavailable, and
and compression and coating in a single integrated process. Similar the probe is exposed to air. Following material level being re-
systems have been widely described in detail in the literature (Van Snick established, the PAT method is automatically restarted, triggering
et al., 2017a; Van Snick et al., 2017b; Vanhoorne and Vervaet, 2020). probe cleaning and reference checks prior to collection of data.
The installation at the Cramlington facility was the first commercial Although the blend NIR is not included in the commercial control
implementation of the CDC-50 and Consigma coaters in a commercial strategy, this system was used as proof of concept to provide real-time
manufacturing environment operating under GMP conditions with a measurements of blend properties as an alternative to RTD to estimate
fully integrated control strategy, and automated sequencing of all unit blend potency (API % LC), as well as to support the development of the
operations from powder dispensing to coated tablet packaging in bulk RTD model and for non-quality critical monitoring of routine operation
containers. Details on the integration of CM line into GMP facility and on NOC batches.
control strategy were discussed in the part 1 of this series.
We incorporate an RTD process model for real-time monitoring of 2.3.2. At-line tablet analyser coupled NIR spectroscopy
API concentration to support rejection of potentially non-conforming The real-time release testing on core tablets is performed by the
material and NIR for real time release testing, thus completely elimi TANDEM IIIA which consists of a Sotax Auto Test4 (AT4) tablet tester
nating laboratory tests for both process control and release. The pro (Sotax AG, Switzerland) integrated with a Bruker FT-NIR spectropho
posed control strategy exploits the benefits of CM and is comprised of a tometer (Bruker Optik, Ettlingen, Germany), both are located next to the
hierarchy of control elements that ensure drug product attributes CDC-50 as shown on Fig. 2a. For simplicity, this is called tablet analyser
approved for the batch process are maintained in the CM process. Based from now on. The AT4 unit includes the tablet transfer system along
on the product formulation and process risk assessment, the proposed with the components to measure tablet weight, thickness, diameter, and
approach is based upon the use of LIWF controls combined with controls hardness. The NIR unit contains the NIR spectrometer and the custom
based on blend composition estimates derived from an RTD model, to ized elements for tablet transport, tablet positioning and ejection sensors
provide robust in-process control of API concentration in the core tab and a cassette for tablet storage which allows full traceability of tablets
lets. The LIWF controls are the primary controls ensuring that the drug being evaluated by NIR (Fig. 2 b-e respectively).
product is always within specifications for API composition, whilst the The tablet analyser is fully integrated and automated to the CDC-50
RTD process model serves both as an additional control predicting API system to measure weight, thickness, hardness, diameter, NIR identifi
concentration in the drug product and a means to perform material cation and API concentration of core tablets. Successful implementation
tracing to facilitate automatic rejection of potentially non-conforming of the tablet analyser required the development of a customized vacuum
material. Realtime product disposition decisions based upon process transfer and deduster system that successfully removed loose powder
parameters and automated at-line tablet testing are also performed. from the core tablets while gently transporting them to the tablet ana
lyser entry port by using customized deduster brushes (Fig. 2f).
2.3. Integrated process analytical technology (PAT) The FT-NIR spectrophotometer is equipped with a RT-InGaAs de
tector and a quartz halogen lamp and we refer to this as tablet NIR from
The CM line installed in MSD has integrated multiple PAT tools to now on. Each NIR spectrum is collected using diffuse transmission and
ensure highest product quality. The PAT components implemented the average of sixty-four scans with a resolution of 32 cm− 1 over the
include NIR spectroscopic methods and an RTD mathematical model spectral range of 5785 – 12480 cm− 1. Customized sample nests (holders)
which are used to measure and track the blend and tablet API concen were designed for each tablet potency to minimize stray-light and ensure
tration. These different methods are integrated with the control system correct NIR spectra gathering. The system was designed to use easily
3
Fig. 1. Blend NIR system setup in the CM line, a) NIR probe location, b) PDV and calibration ring location, and c) level sensor location.
Fig. 2. A) Tablet analyser integrated to the CDC-50 platform. b) customized table nest, c) positioning sensors, d) tablet ejection sensors, e) tablet cassette and f)
dedusting brushes.
changed accessories for each potency such as brushes, funneles, sample each with its own functionality:
nest, vials, and cassettes to minimize change time between tablet
strengths. An instrument cart was custom designed to provide a stable i) Siemens WinCC supervisory control and data acquisition
base for the tablet analyser while also facilitating the cleaning and (SCADA) system,
movement of the instrument when required. The tablet analyser system ii) Automation layer (Siemens programmable logic controllers
is integrated to the CDC-50 and PMS and can be operated in a completely (PLCs) and dedicated controllers),
automated and touchless mode by the CM operators. iii) PMS to support management and integration of the PAT in
struments on the CM line and,
2.3.3. PAT and automation overview iv) The two PAT Instruments: Blend NIR and tablet analyser.
Fig. 3 shows the high level of the system architecture and commu
nication links for the PAT instruments and CM line. As shown, the The PMS communicates with the SCADA and automation layer
automation of the entire CM line consists of four main building blocks through standardized interfaces (PAT and DPC interface). The tablet
4
Fig. 3. High-level PAT/automation system architecture and communication.
press (automation layer) communicates directly with the tablet analy i) Two out of average potency predictions from SPC are points more
ser. A historian software interface is established to connect the auto than 2σ from centreline will detect Mean shift or variation increase
mation to the existing plant historian data archiving system. Note that and
the CDC-50 does not directly control the blend NIR instrument; instead, ii) Eight average potency predictions in a row, more than 1σ either side
the CDC-50 sends orders (i.e., initialize, run instrument performance of centreline will detect increase in process variation.
qualification (PQ), start/pause/stop data collection) to the PMS which
then in turn controls the blend NIR system via the PMS collector station. The integration of the tablet analyser and the PMS has been done in a
The coordination between NIR measurement and probe position is way that ensure collection, contextualization and time alignment of
driven by the PMS, this also includes moving of the probe for cleaning, tablet metadata and results such as sample and reference spectra, NIR
sampling, and reference check. The tablet analyser communicates to the predictions and tablet physical testing data. Furthermore, care was
CDC-50 through an OPC-DA server. In normal operation the CDC-50 taken to ensure compliance with data integrity principles. The PMS
pushes a recipe to the PMS, which then triggers a data collection pro report generation capability was used to support GMP batch reporting.
cess to the tablet analyser. All data gathered is stored and archived in the The tablet NIR system control, and execution of NIR calibration models
PMS. is managed by the tablet analyser software (OPUS by Bruker) but is
Two of the main responsibilities of the automation layer are: triggered by the PMS.
1) the API concentration prediction by the RTD model and

2) the in-process monitoring control of statistical process control (SPC) 2.4. Continuous manufacturing process
alarms used to minimize or prevent adverse impact on product
quality and minimize non-value-added work caused by unnecessarily A schematic Process Flow Diagram of the CM process presented in
investigating inherent process variation. The two Nelson rules this article is shown in Fig. 4. Firstly, at reception in the warehouse, raw
(Nelson, 1984) used for SPC in the API NIR prediction monitoring in materials are checked for identity using existing procedures with hand-
CM are: held Raman spectrometer. For each development and NOC batches
produced using the CM line, the major excipients; EX1, EX2, and API are
Fig. 4. Process flow diagram for continuous manufacturing process of selected legacy product.
5
charged into intermediate bulk containers (IBCs). The remaining minor Table 1
excipients; EX3, EX4, and EX5 are charged to smaller stainless-steel Summary of conditions for NOC batches.
containers for gravity feed to LIWFs. The raw materials in IBCs are Batch Tablet Mass line flow # Sampling Batch
pneumatically conveyed to LIWFs. We consider the controlled, precise, strength (mg) rate (kg/h) points duration (h)
and quantitative discharge of material from the LIWFs to be the point Batch 100 75 12 2.4
that material dispensed into the CM process. The top-up valves located 1
in the LIWF are activated periodically to ensure a constant mass flow in Batch 100 75 11 2.3
the system. The API and excipients are transferred via LIWFs and 2
Batch 25 21 11 2.5
conveyed to the first blender at constant mass flow rates that are 3
determined by the total system mass flow and the composition of the Batch 25 21 12 2.5
formulation. The material then falls by gravity into a second blender 4
where lubricants (EX4 and EX5) are added from additional LIW feeders. Batch 25 21 8 1.4
5
The resulting blend from second blender is continuously fed by gravity
Batch 25 21 29 6.6
through a powder dosing valve into a single-sided rotary tablet press. 6
Tablets are compressed through a traditional compression process and Batch 25 21 17 3.8
then conveyed to one of two duplexed ConsiGma ™ coating pans for film 7
coating. Batch 25 21 14 1.3
8
Finally, the film coated tablets, once discharged from each coating Batch 100 75 10 3.5
pan, are cooled in a cooling drum, and transferred via the distribution 9
arm into drums for storage, shipping, and packaging operations.
Critical quality attributes of the blend are controlled through appli
cation of RTD which models the LIWF, blender and press using idealized automatically calculated by the PMS, and each batch recipe is aligned
mixing units, with parameters that are designed for the actual equip with the acceptable range shown in Table 2.
ment. Once the fitting parameters are established and validated, live In the event of an out-of-control limit result is generated:
flowrate input from each of the LIWFs is then used at high frequency to
both predict the progress of API material in the system and activate i) an alarm is displayed on the CDC-50 system human machine inter
automated rejection of potentially non-conforming material. The RTD face (HMI) display and the disposition of that portion is managed
process model predicts API concentration of core tablets exiting the after the coating process. This alarm is produced for each individual
tablet press over time. NIR spectroscopy is utilized to monitor the crit result generated outside of the internal control limits and for mean
ical quality and performance attributes (API concentration and Identity) results outside of mean limits as specified in the recipe.
of core tablets by using the tablet analyser. This is used to analyse a side- ii) The press and upstream equipment are stopped immediately upon
stream of tablets, periodically drawn from the press to both control the receipt of the result generated by the tablet analyser for weight and is
press operations via a feedback loop using the core tablet weight and investigated by the Operator as defined under site procedures.
hardness measurements and provide critical quality attribute data to
facilitate tablet release and to provide disposition decision post coating. Regarding NIR testing, three core tablets per sampling location are
More details about the control strategy approach including RTD and NIR evaluated for API concentration using the tablet NIR (no less than 30
for real time release can be found in part 1. individual tablet results per batch). Although, the identity testing by NIR
is performed in each tablet, these identity results are not part of the IPC
2.4.1. Normal operational conditions (NOC) batches
As part of process development activities and prior to manufacture of
process performance qualification (PPQ) batches, several NOC batches Table 2
were manufactured in the same fashion as expected in commercial Summary of IPCs and CQAs for the product manufactured using continuous
manufacturing process.
manufacturing scale. The purpose of these runs was mainly to perform a
shakedown of integrated automation systems (i.e., enterprise resource Test variable Acceptance criteria
management, MES, PMS and data historian) and challenge PAT methods Process step
including NIR instrument functioning as well as robustness of NIR and Loss-in-weight RTD API prediction 85 – 115%
RTD models. In total, nine batches were manufactured, three batches for feeding, blending,
compression
the 100 mg strength and six batches for the 25 mg strength. These
Compression Average tablet 25 mg: Target 100 mg, weight
batches were manufactured using the same API and raw materials weight (in-process 95–105 mg
sources as currently used for routine batch process manufacture under test) 50 mg: Target 200 mg, weight
full GMP conditions. Table 1 shows a summary of NOC batches is pre 190–210 mg
sented; note that for each tablet strength, batches size and sampling 100 mg: Target 400 mg, weight
380–420 mg
points vary. This was intended to challenge the system and simulate Individual tablet 25 mg: Min 4.5 kP
different scenarios in manufacturing. The data gathered and presented hardness (in-process 50 mg: Min 6.3 kP
in this article are focused only on PAT and RTD results. test) 25 mg: Min 8.9 kP
NIR API prediction Individual tablet API
(in-process test) concentration 75.0 – 125.0% of
2.4.1.1. In-process controls (IPC). The weight of at least one hundred target concentration
tablets and hardness of fifty tablet units were measured by the AT4 unit Uniformity of dosage The batch complies if the number
of the tablet analyser on each batch and at approximately 15 minute units (release test) of tablets outside of 85 – 115.0%
intervals (minimum) from the start of compression to the end of the of label claim is less than or equal
to the acceptance limit.
batch. The sampling interval and number of core tablets sampled are
Identification by NIR The presence of API is confirmed
dependent on batch size. However, it should be noted that it could take (release test) if the sample spectrum passes the
up to 10 minutes to complete a full IPC testing cycle, during which time criterion set forth in the library
the press is still running, and tablets are being transferred to the coaters. validation.
As part of the release process the mean weight of ten individual tablets is Assay by NIR (release 95.0 – 105.0% of label claim
test)
6
strategy. However, in the event of an out of specification result for API Table 4
concentration (outside of acceptable range of 75.0 – 125.0 % for indi Summary of variables and ranges included in the experimental designs.
vidual tablets) or identity (fail threshold), an alarm is generated and Potency API sources Tablet weight (mg) Hardness (kP)
displayed on the CDC-50 HMI, following by the immediately stoppage of
25 mg 2 92.5 – 107.5 5–9
the press and upstream equipment, which will require investigation by 50 mg 2 185 – 215 7 – 15
the operator as defined under site procedures. 100 mg 2 380 – 420 11 – 20
2.4.2. Development batches

The NIR model calibration designs were split into two experimental by calculating the root mean square (RMS) of the analyte band in a
matrices: Design-1 included variability around both API and excipient derivative spectrum (used to minimize influence of physical properties
concentration while Design-2 included only variability in API concen on measurement). The main objective for this qualitative method is
tration from 60 to 140% of label claim (excipients were maintained at similar to the classical moving block standard deviation (MBSD) (Hailey
target concentration ratio). The experimental designs utilized to et al., 1996; Sekulic et al., 1996), which is to evaluate spectral variation
manufacture tablets being used to build both, the tablet and blend NIR across the batch that can be linked to blend uniformity.
models, which summarised in Table 3. The size of the moving block used in this paper is optimized based on
Concerning NIR assay models, the tablet properties included in the the blend mean residence time (MRT) – which varies based on the flow
experiments were API concentration, API source, excipient concentra rate of the batch. For example, a flow rate of 21 kg/h (25 mg dose)
tion, tablet weight and tablet hardness. Table 4 summarises the variables would have an MRT of 226 s while a flow rate of 75 kg/h (100 mg dose)
and ranges included within the tablet NIR calibration designs for each would have an MRT of 114 s.
tablet potency. Moving blocks from 0.25MRT (less smoothing / high sensitivity to
Due to the scale independent nature of continuous processing, and analyte spike) to 3MRT (more smoothing / lower sensitivity to analyte
the fact that model development activities were conducted using the spike) were evaluated. A moving block of 0.25MRT was selected based
same instrument and process equipment that will be used for produc on discussions with engineers in favour of enhancing the sensitivity of
tion, it was possible to eliminate production scale, site location and in the measurement and thus minimizing the likelihood of missing a spike
strument to instrument variability from the list of attributes that were in API that may impact content uniformity.
varied in the experiments. Furthermore, minor changes in tablet prop To illustrate the moving block calculation, let’s consider the MRT of
erties caused by raw material variability and/or process variations were a blend is 240 s (0.25MRT = 60 s) and 1 spectrum is collected every 5 s.
incorporated into the calibration set throughout model development The moving block size calculation would be as follows:
activities. Hence, by incorporating data from these activities, various (
1spectrum
)
potential contributors to NIR method error including processing varia Movingblock = 60sx = 12(spectra) (1)
5s
tions, excipient variability, spectrometer drift and tablet orientation
(randomized throughout data collection) are implicitly included into the To assess BU, the NIR spectral data was pre-treated using the first
model calibration and validation sets. Furthermore, robustness to derivative Savitzky-Golay with a 15-point window and second poly
moisture was not specifically studied as part of these model develop nomial order. The magnitude of the API band is measured by calculating
ment activities as tablets are produced in a temperature and humidity- the root mean square (Equation (2)) of the derivative values (y) for each
controlled production suite and analysed in the same area within mi spectrum in the spectral range 1664 – 1734 cm− 1, which is the main
nutes of compression. In addition, alarms are present for temperature region where API absorbs. For each block, a moving average and moving
and/or humidity excursions beyond normal operating conditions. For standard deviation is calculated from the RMS result from each spec
the low probability of occurrence that these controls fail, outlier di trum. For each moving block (mb), the %RSD (Equation (3)) is
agnostics are in place for every spectrum collected and will flag spectra calculated.
that are outside the variability captured in the calibration data set. √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
1∑ 2
RMS = y (2)
n i i
2.5. NIR methods development and validation
Where, n number of measurements, yi each value
2.5.1. In-line qualitative estimation of BU by using moving block root mean ( )
σmb
squares relative standard deviation (%RSD). %RSD = x100 (3)
xmb
A principle of NIR spectroscopy is that the intensity of an absorption
band belonging to the analyte is proportional to the concentration of The moving block %RSD will be compared to the 99.9% limit, which
analyte in the matrix. Measuring the analyte band intensity in each was established at 3.0 % and calculated using the McKay formula
blend spectrum and calculating the variation of the band intensities (McKay, 1932). This limit was selected based on experience in previous
across a given time period (moving block) indicates the consistency of development experiments and considering outcome from batches
that analyte in the blend. For this work, the band intensity is measured
Table 3
Summary of experimental designs used to build tablet and blend NIR models.
NIR model Design Runs Tablet Throughput (Kg/ API API EX1 EX2 EX3 EX4 EX5
strength h) LC (%) (%w/w) (%w/w) (%w/w) (%w/ (%w/ (%w/
(mg) w) w) w)
Assay by tablet Design- 5 25, 50, 100 21, 42, 50 65 – 135 20.9 – 23.1 – 24.8 – 1.6 – 2.4 3.0 – 3.6 0.8 – 1.2
NIR 1 43.4 38.0 37.1
Design- 9 25, 50, 100 21, 42, 50 60 – 140 19.3 – 25.1 – 25.1 – 1.6 – 2.4 2.4 – 3.6 0.8 – 1.2
2 44.9 36.8 36.8
BU by blend NIR Design- 6 100 50, 90 109 – 35.0 – 22.9 – 22.9 – 1.5 – 2.5 2.2 – 3.7 0.7 – 1.2
1 155 49.8 31.0 34.0
Design- 7 25 15 100 – 32.1 – 22.9 – 22.9 – 1.4 – 2.0 2.2 – 3.3 0.7 – 1.2
2 155 49.8 38.1 30.9
7
operated at different flow rates and, thus, different moving block sizes. analysed and then converted to a sample weight using the blend density.
This limit of 3.0% RSD was chosen as a worst case scenario – so it was the As a result, sampling masses of 95 mg, 167 mg and 257 mg were
lowest calculated limit when using the block sizes from the low to high calculated for mass line flow rates of 10, 50 and 100 kg/h, respectively
throughput. A disturbance is defined as N consecutive values above the (which span the mass flows used in this work). For comparison, the
3.0% limit where N is the block size. A batch is considered consistently tablet weight targets for the 25 mg, 50 mg and 100 mg tablet strengths
blended if there are no disturbances based on the aforementioned are 100 mg, 200 mg and 400 mg, respectively. Consequently, blend
moving block results. sampling masses per spectrum are on the same order as the unit dose
masses for the tablet strengths, and thus each collected spectrum
2.5.2. In-line quantitative blend potency estimation reasonably approximates a unit dose.
Toward supporting RTD process validation and manufacturing of
NOC batches, two fit-for-purpose partial least squares (PLS) models – for 2.5.3. Tablet NIR assay model
simplicity, this is so-called blend model 1 and blend model 2 from now A PLS regression was used to construct a separate NIR calibration
on – have been built using average for high performance liquid chro model for each strength. The calibration models were validated with
matography (HPLC) assay results on a composite sample of core tablets independent samples that reflect the commercial scale. For CM, the
collected at press level during experimental runs shown in Table 3. spectral data used for model calibration and validation were collected
The aim of these NIR models was to demonstrate that BU can be with the same equipment used for commercial production.
quantitatively assessed by using reflectance NIR spectra from blend and The optimum number of PLS latent variables was selected based on a
matching with HPLC results of active core tablets manufactured minutes minimization of cross validation error (expressed as the root mean
later. This approach provides estimation of the API % concentration in squared error of cross validation or RMSECV). According to ICH Q2(R1)
the blend between the outlet of the second blender and would suggest (ICH, 2005), accuracy expresses the agreeability between reference and
that the process is in a state of control. Table 5 shows the figure of merits predicted values. Therefore, RMSEC and RMSECV are used to report the
for the blend NIR models. The blend NIR model 1 was used to support accuracy of the selected calibration. The agreement between the NIR
preliminary RTD development runs and model 2 (model 1 updated) was predicted concentration and HPLC reference values for the independent
employed on the NOC batches. Regarding the NOC batches, the API validation data set is reported through the root mean squared error of
predictions were obtained off-line to identify the best approach for prediction (RMSEP). For all three tablet strengths, it was determined
future manufacturing. that the best spectral pre-processing was first derivative Savitzky-Golay
For the purposes of model development and validation, it is impor 17-point over full spectral range followed by multiplicative scatter
tant that each blend NIR spectrum used in the analyses represents correction (MSC) which was restricted to the spectral range used for
approximately a single unit dose of material (~1-3X). An estimation of calibration. The spectral region used in PLS model development was
the mass of blend sample analysed by the system is a function of probe selected to contain the absorbance band at approximately 8800 cm-1
type (spot size and number of spots) acquisition parameters (integration which was assigned to the API of interest. Table 6 shows the figures of
time and number of scans to average), mass line flow rate, powder blend merit for the tablet NIR models for the 25 mg, 50 mg and 100 mg po
density and NIR penetration depth. An NIR penetration depth of 1 mm tencies. In spite of the fact that models were built for the three potencies,
was used in the calculations as this represents a realistic estimation of only results for 25 mg and 100 mg batches are presented in this article.
the value as reported in literature based on the blend properties and NIR
scan range used for this measurement (Berntsson et al., 1999). These 2.5.3.1. Outlier diagnostics. The outlier detection performance of the
parameters were used to calculate the volume of blend formulation models is evaluated by calculating the ‘in-space’ metric Mahalanobis
distance (MD) and the out of space’ metric F-probability (Fprob) for all
samples within the independent validation data sets. Action limits for
Table 5 outlier diagnostic mechanisms are established based on the samples in
Figure of merit for blend NIR models used to estimate assay in the blend prior the calibration set to flag spectra as outliers.
compression. An ‘in-space’ metric is used to determine if the acquired NIR spec
Parameter Blend NIR model 1 Blend NIR model 2 trum is within the calibration sub-space, but outside the range of the
description calibration set spectra. This metric is employed to determine if the
Calibration Samples in 337 997 spectrum being analysed represents an interpolation or extrapolation of
set calibration set the variable space defined by the calibration model. This statistic is
Spectral pre- Savitzky-Golay first Savitzky-Golay first
useful for the detection of outliers at the modelling or prediction stage.
treatment derivative, 35 points derivative, 13 points
and 2nd polynomial and 2nd polynomial
MD expresses the distance of a given spectrum from the mean of the
order followed by SNV order followed by SNV calibration set spectra. Outlier assessment is done by comparing the MD
and mean centring and mean centring value of a single measured spectrum to a threshold MD value, which is
Model spectral 1551–1870 1300 – 1870 calculated from the MD values of the calibration spectra. If a measured
range (nm)
sample’s spectrum has an MD value within the threshold, then the
Calibration 79 – 154 60 – 152
range (HPLC sample is reliably analysed.
%LC) An ‘out of space’ metric is used to determine the extent to which the
Model rank 2 3 acquired spectrum is outside of the model sub-space. It is a measure of
Total X- 95.2 94.0 the spectral variation that is not considered by the model. The residual
variance
explained (%)
for a given sample and a given variable is computed as the difference
Total Y- 98.4 98.6 between the measured value and the model-fitted (or projected or pre
variance dicted) value of the variable on the same sample. The Fprob values in
explained (%) dicates the probability of the existence of an outlier in the distribution of
RMSEC (%LC) 2.6 2.6
all F-values, which are calculated directly from the spectral residuals.
Validation Prediction 89 – 148 66 – 147
set range (HPLC The larger the F-value of the analysed sample, the more likely it is an
%LC) outlier and consequently provides a larger Fprob value (close to 1). Both
Number of 157 265 in-space and out-of-space outlier diagnostic data is generated for every
samples tablet analysed via NIR for assay and, along with API concentration
RMSEP (% LC) 2.5 2.5
8
Table 6 calculated tablet volume measured by the tablet NIR is 50% – 80%. In
Figures of merit of tablet NIR models for API 25 mg, 50 mg and 100 mg tablet addition, the validated HPLC reference method effectively analyses an
strengths. entire tablet. Since the tablet NIR model correlates well with the HPLC
Parameter 25 mg strength 50 mg strength 100 mg strength results across a wide sample population, it can be stated that the NIR
Description tablet sampling efficiency is sufficient to meet the intended use of the
Calibration method.
Number of 93 93 97
samples 2.5.4. Identification (ID) NIR model
API 148.9 – 349.0 150.5 – 348.6 149.9–423.5
The tablet NIR identity model was developed for core tablets man
calibration
range (mg/ ufactured as 25 mg, 50 mg or 100 mg tablet strengths which nominally
g) weigh 100 mg, 200 mg or 400 mg, respectively. A single NIR spectral
API 58.9 – 154.1 57.5 – 152.5 57.4 – 141.2 library was utilized to provide end process identity release testing for all
calibration three tablet strengths.
range (%LC)
Spectral pre- 1st Derivative, full 1st Derivative, full 1st Derivative, full
Since the same tablets were used for both the NIR assay method and
treatment range, range, Savitzky- range, Savitzky- NIR identity method, the range of API for the calibration set, and for the
Golay 17-point Golay 17-point positive controls in the validation set, is at least 70% to 130 %LC. The
Savitzky–Golay 17- smooth followed smooth followed validation set included placebo tablets as negative controls. The vali
point smooth by multiplicative by multiplicative
dation of the identification library was performed by using the inde
followed by scatter correction* scatter correction*
multiplicative pendent positive control and negative control datasets.
scatter correction* The library was constructed using twenty-seven samples for the three
Spectral range 9519.6 – 8593.9 9411.6 – 8779.0 9535.0 – 8701.9 potencies. Table 7 shows the figures of merit of the NIR identity model.
(cm− 1) The spectral range of 8440 to 9304 cm-1 was used to distinguish be
Model rank 4 4 4
tween active tablets and the negative controls used in the validation data
Validation
Number of 33 33 36 set. The spectral responses over this region for unknown samples in the
samples validation set are compared to those stored in the spectral library, and
API validation 162.0 – 349.7 162.0 – 349.7 150.7 – 353.8 identification is performed to determine if the sample is identified as an
range (mg/
active core tablet or not.
g)
API validation 59.1 – 142.1 59.1 – 142.1 63.3 – 148.9
range (%LC) 2.6. Residence time distribution (RTD)
Outlier
diagnostic The RTD of ideal reactors, such as a plug-flow reactor (PFR) or a
thresholds
Mahalanobis 0.21 0.19 0.21
continuously stirred tank reactor (CSTR), are well established in litera
distance ture (Fogler, 2016; Hurley et al., 2022) and described in part 1 of this
(MD) series.
F-Probability 0.999 0.999 0.999 During development, the RTD model was optimized using blend NIR
(Fprob)
and tablet HPLC data across a matrix of process conditions including
* MSC is restricted to the spectral ranged used for modelling. mass line flow rate, blender speed and incorporating raw material
variability.
prediction, is considered part of the model output. A schema of response Hurley et al (Hurley et al., 2022) described the development and
for different scenarios with high outlier metrics is displayed on Fig. 5. validation for RTD model employed in this CM process. In this article the
The outlier diagnostic thresholds for the tablet NIR models are main results of RTD model applied to NOC batches are shown. In
summarised in Table 6. Based on these results, outlier diagnostic per addition, it is proposed our approach for RTD model lifecycle
formance for all models was considered acceptable for their intended management.
use.
3. Results and discussion
2.5.3.2. Considerations on mass analysed by transmission NIR. The vol
ume of tablet analysed by diffuse transmission spectrometry is greater 3.1. Tablet NIR assay method validation
than that of diffuse reflectance spectrometry based on photon migration
theory in highly scattering media. While the sampling volume in a non- The acceptance criteria for specificity, accuracy, linearity and range,
turbid media can be easily estimated using the NIR beam dimensions measurement precision, method precision and robustness, as well as the
and given pathlength, diffuse transmission sampling volume of phar obtained results are summarised in Table 8.
maceutical tablets cannot be readily calculated from first principles Specificity of the tablet NIR assay method was evaluated by
using optical and sample properties. Scattering phenomena in physically comparing the NIR spectrum from a placebo 100 mg core tablet (same
complex samples such as pharmaceutical tablets are difficult to model excipient ratio but with 0% API) in comparison to pure API spectrum
and are highly dependent on variables such intra-tablet morphology, along with active tablets of nominally 60%,100% and 140% LC. The pre-
pore volume, pore size distribution, particle surface area, surface treated spectra plotted in Fig. 6 clearly illustrate the spectral regions
roughness, etc. Scientific literature based upon time resolved spectros specific to the API within the NIR range utilized by the NIR model. More
copy studies show that multiple scattering occurs within a tablet as the precisely, the band at approximately 8800 cm− 1 becomes more intense
NIR radiation is transmitted through the tablet. The study by Johansson with increasing concentration of API in the tablet. This absorption band
(Johansson et al., 2002) showed that the optical pathlength in a tablet is was assigned as the second overtone of the aromatic fundamental – C–H
55 – 70 times greater than the physical pathlength (3.5 mm physical stretching vibration. Since the only compound with an aryl moiety in the
pathlength resulted in a 200 – 250 mm optical pathlength). The effective formulation is API, this observation confirms specificity for the analyte.
sampling volume of core tablets was approximated through geometric Broader regions of the spectrum were also be considered to account for
calculations including the sample holder aperture used on the specific spectral variations potentially originating from optical thickness and/or
NIR instrument and tablet dimensions. Based on these assumptions the physical properties of the tablets. Since the three tablet strengths are
dose multiples and the particular tablet strength used to evaluate
9
Fig. 5. Schematic representation of outlier diagnostic metrics used to support tablet assay NIR results and response to outlier alarms during a CM process.
Table 8 shows that the regression models for all three potencies are
Table 7
satisfactory for intended use.
Figures of merit for Identity NIR model.
Accuracy was evaluated by calculating the RMSEP and bias for all
Method Parameter Value samples within the independent validation data sets and as detailed in
component
Table 8; all three tablet NIR models showed acceptable accuracy as
Library model Identification algorithm Euclidean distance compared to performance criteria and bias was not statistically different
Spectral pre-treatment 2nd derivative Savitzky-Golay 9-
from zero (at α = 0.05).
points over full range, followed by
vector normalization on the spectral
Measurement precision was demonstrated by the replicate analysis
range of an individual active core tablet (one tablet for each potency) using
Spectral range 8439.6 – 9303.6 cm-1 only one sample presentation. Ten separate spectra were collected using
Library samples 27 a single core tablet without moving the tablet in between scans. This was
%LC range in library 58.9 – 142.1
repeated for each of the three potencies and results shown in Table 8
Validation samples 111
(positive and negative demonstrate measurement precision was deemed acceptable as
controls) compared to performance criteria.
Euclidean Significance level 99% Method precision was verified by the replicate analysis of an indi
distance Distance threshold 0.512
vidual active core tablet (one tablet for each potency) using ten sample
Mean distance 0.203
# Spectra outside 0
presentations. After completion of the first NIR spectral data collection,
the tablet was manually returned to the entrance port of the tablet NIR.
This process was repeated ten times and as shown in Table 8 method
specificity should have minimal to no bearing on the specificity of the precision results were satisfactory as compared to performance criteria.
NIR method, the specificity analysis was limited to 100 mg active core Robustness was evaluated by imposing deliberate and systematic
tablet spectra. variations. These outcomes offered additional guidance to determine the
Linearity and range were determined by the minimum and maximum limits for application of the models. The robustness with respect to tablet
label claim samples that were included within the validation data set. weight and hardness was evaluated by measuring active core tablets
The range for the three NIR models included all eleven concentration from experimental design batches that were compressed to have a range
levels from the calibration design nominally spanning 60 to 140 %LC. of physical properties. These NIR spectra were included in the
10
Table 8
Validation parameters for the tablet NIR assay methods.
Validation criteria Parameter Acceptance Strength
criteria
25 mg 50 mg 100 mg
Linearity Range (HPLC % LC) at least 70% to 59.1 – 142.1 66.2 – 139.8 63.3 – 148.9
130% LC
Number of API levels At least 5 level 11 11 11
Slope (HPLC vs NIR) Between 0.90 1.00 0.98 0.98
and 1.10
R2 ≥ 0.95 1.00 0.99 1.00
Accuracy RMSEP (% LC) ≤ 2.5 % LC 0.9 1.6 1.1
Bias (%LC) N/A 0.05 0.1 0.3
Measurement precision Average (% LC) N/A 99.4 100.0 101.9
(Repeatability) Standard deviation (% LC) N/A 0.06 0.2 0.4
% RSD RSD ≤ 1.0% 0.06 0.2 0.4
Method precision Average (% LC) N/A 101.3 99.8 101.2
(Reproducibility) Standard deviation (% LC) N/A 0.3 0.4 0.4
% RSD RSD ≤ 2.5% 0.3 0.4 0.4
Robustness (API source, mass line Average absolute residual (%LC) by API source (A, B) ≤ 2.5% LC 0.5 (A), 1.1 1.3 (A), 1.1 0.8 (A), 0.9
flow rate, tablet age) (B) (B) (B)
Average absolute residual (%LC) by mass flow (A = 15Kg/h, B = 21 ≤ 2.5% LC 1.2 (A), 0.7 2.1 (A), 0.4 0.9 (E), 0.3
Kg/h, C = 25 Kg/h, D = 42 Kg/h, E = 50 Kg/h, F = 75 Kg/h, G = 90 (B) (C), 1.1 (D) (F), 0.6 (G)
Kg/h)
Average absolute residual (%LC) by tablet age (A = 3-10 min, B = 3 h, ≤ 2.5% LC 0.7 (A), 1.5 1.2 (A), 1.7 0.9 (A), 0.6
C = 6 h, D = 43 h, E = 45 h, F = 73 h) (E), 1.3 (F) (B) (C), 1.5 (D)
Over whole robustness Pass/Fail Pass Pass Pass Pass
potencies. Tablet age is defined as the approximate elapsed time from

the time the tablet was originally compressed (Time = 0) to the time the
NIR spectrum for the tablet was collected. Whereas samples used for
calibration and validation sets have a tablet age of 3 – 10 min, as part of
this testing the tablet age times from 3 h up to 73 h were investigated
and as highlighted in Table 8, results demonstrated be robust to tablet
age.
Note that tablet misplacement is often studied in regard to NIR tablet
methods. However, the tablet NIR system has built-in sensors that
monitor the tablet position and confirm proper tablet placement.
Therefore, using system automation, tablets that are not placed properly
in the sample nest are rejected prior to NIR spectral measurement and a
replacement tablet is ordered. Moreover, for the more than four hundred
tablets analysed as part of this model development and validation, there
were zero tablets identified as spectral outliers, indicating that no tablets
were misplaced. The robustness of the tablet NIR methods with respect
to tablet debossing was not specifically evaluated as the orientation of
the tablets (i.e., company logo facing the transmission detector or the
light source) was deliberately not controlled during NIR spectral
acquisition. Thus, this variability was included in the calibration and
Fig. 6. Pre-treated spectra 1st Savitzky-Golay derivative 17 points + MSC for
validation data sets. The fact that all NIR assay methods demonstrated
blank tablet (red), Pure API (blue), active tablets at 60% (green), 100% (or
acceptable accuracy and precision indicates that they are robust to tablet
ange) and 140% (purple) LC. (For interpretation of the references to colour in
this figure legend, the reader is referred to the web version of this article.) orientation.
calibration and validation data sets and the acceptable accuracy and 3.2. Tablet NIR identity method validation
precision of the three NIR models demonstrated that NIR assay methods
are robust to considerable variations in tablet physical properties. The The NIR identity method has been validated for specificity, precision,
robustness with respect to API sourced from two different vendors was and robustness in similar fashion as the assay method with results
also investigated and these NIR spectra were included in both the cali summarised in Table 9. Specificity was firstly assessed by comparing the
bration and validation data sets. Therefore, no difference in method NIR spectrum from a placebo 100 mg active core tablet (0% API) to pure
accuracy was noted between vendor API when predicting the indepen API along with tablets of nominally 60%, 100% and 140 %LC as already
dent validation set thus models are robust to API sources. shown in Fig. 6. Secondly, specificity was evaluated by challenging the
Robustness in regard to manufacturing mass line flow rate was method with positive and negative control samples. An independent
likewise assessed and NIR spectra included in the calibration and vali data set of active core tablets across the three strengths and covering
dation sets cover different mass flow conditions – thus each potency (25 60% – 140 %LC were used as positive controls. Furthermore, the positive
mg, 50 mg and 100 mg). Since the calibration models were used to control data set contains samples from two different API sources as well
predict API concentration at a range of mass line flow rates from 15 kg/h as samples with excipient concentration, weight, and hardness vari
to 90 kg/h, data from these predictions (Table 8) illustrated passing ability. In contrast, placebo tablets from all three strengths were used as
results when compared to performance criteria. Thus, methods are negative controls. As consequence, all positive control samples were
robust to manufacturing mass line flow rate. correctly identified as active tablets whereas all negative controls
The robustness with reference to tablet age was studied for all three correctly failed identity testing (Table 9). Based on these results, tablet
11
Table 9
Validation parameters for the tablet identification NIR method (Threshold = 0.512).
Validation criteria Parameter Strength
25 mg 50 mg 100 mg
Specificity Positive controls identified as active tablets 33 33 36

Positive controls not identified 0 0 0
Negative controls identified as active samples 0 0 0
Negative controls not identified 3 3 3
Results Pass Pass Pass
Measurement precision (repeatability) Measurements 10 10 10
Average (Euclidean distance) 0.039 0.021 0.043
Min – Max. 0.038 – 0.039 0.020 – 0.022 0.038 – 0.049
Method precision (reproducibility) Measurements 10 10 10
Average (Euclidean distance) 0.035 0.020 0.040
Min – Max. 0.033 – 0.037 0.019 – 0.021 0.031 – 0.054
Robustness (Mass line flow rate) Mass line flow rate (Kg/h) 21 15 25 42 15 50 75 90
Samples correctly identified / samples not identified 33 / 0 3/0 3/0 33 / 0 3/0 36 / 0 3/0 3/0
Robustness (tablet Age) Tablet age (h) 73 3 43
Samples correctly identified / samples not identified 15 / 0 18 / 0 16 / 0
Overall robustness Pass/Fail Pass Pass Pass
ID NIR method meets the performance criteria for specificity. assay values and the overall RMSEP across the RTD model validation
Measurement precision was assessed using a single active core tablet was calculated to be 1.4%.
from each strength. The tablets were scanned ten consecutive times
without moving the sample between scans. As shown in Table 9, all ten
3.4. Applying blend NIR model 1 to RTD development runs
measurement precision scans were correctly identified as active tablets,
consequently, tablet ID NIR method meets the performance criteria for
As aforementioned in Table 5, a fit-for-purpose blend NIR model 1
measurement precision.
was validated using an external validation set, therefore it was employed
Method precision was evaluated using a single active tablet from
to support RTD development experiments. As part of these experiments,
each strength. The tablets were scanned ten times running it through the
API spikes with different amounts of API (30 g or 100 g) were manually
complete NIR tablet analyser sampling system each time. As shown in
dispensed into the inlets of blender 1 and blender 2 during the CM
Table 9, all scans were correctly identified as active tablets meeting the
process. These spikes were performed once system has reached steady
performance criteria for method precision.
state and at predefined speeds for blender 1 (180, 315 and 450 rpm) and
Robustness was assessed by imposing deliberate or systematic vari
blender 2 (150, 210 and 300 rpm). As example, Fig. 7 shows the blend
ations to the library, as appropriate. Tablet attributes that varied in the
NIR predictions (API % LC) after API spikes for the three tablet strengths
prepared sample sets included drug substance concentration, drug
at multiple mass flow rates. The amount of API manually added was 100
substance source, excipient concentration, tablet weight and tablet
g and mainly into the inlet of blender 1, runs were performed using
hardness. The robustness of the library with respect to two sources of
speeds of 315 and 210 rpm for blender 1 and 2 respectively. Fig. 7 shows
API were included in the positive control data set. All samples within the
the NIR model successfully detected the API spikes and generated NIR
positive controls were correctly identified as active tablets thus illus
blend composition profiles which were used for RTD model fitting. The
trating robustness to API source. Robustness with respect to API con
spikes of API are clearly observed and correctly estimated by the NIR
centration, tablet weight and hardness was also investigated as the
model for the three potencies and at different mass line flow rates. Note
positive control sample set contained tablets with API concentration
that at lower mass flows, independently of the tablet strength, and in
varying from 60% − 140% of nominal concentration, weight values
crease of noise is notably observed generating disturbances in the API
ranging from 93 mg to 420 mg and hardness values from 5 to 20 kP.
predictions. This kind of disturbances are also observed at a mass flow of
These positive controls were correctly identified as active tablets thus
25Kg/h; however, it is considerably reduced after mass flow increased
demonstrating NIR ID method is robust to concentration, weight, and
from 50Kg/h. Additionally, blend NIR was able to differentiate and
hardness changes. Further robustness with respect to tablet age was also
detect the spikes of 30 g and 100 g API dispensed in the in inlet of
investigated by scanning a selection of tablets from each strength which
blender 2. As shown in Fig. 8, the magnitude of the peaks is proportional
were held and re-scanned after a period of several hours to a period of
at the amount of API dispensed. No impact of noise in the NIR pre
several days. As shown in Table 9, all samples were correctly identified
dictions was observed which suggest that under these conditions at
as active tablets thus demonstrating robustness to tablet age. Finally,
50Kg/h mass flow and fixed speed blenders (315 rpm and 210 rpm for
robustness in regard to mass flow rate was considered as mass flow rates
blender 1 and 2 respectively), the product blend flows smoothly through
of 21 kg/h to 50 kg/h are part of the positive control data set and thus
the NIR probe location, then by the PDV and finally into the press
the library is robust in this range. Samples at additional mass flow rates
without disruption in this flow which will require extra activation of the
were investigated to cover a range from 15 kg/h to 90 kg/h and results
material inlet hopper vibrator causing brief changes in the NIR diffuse
demonstrated all samples were correctly identified as active tablets thus
reflectance spectra during runs as observed for the runs at lower mass
demonstrating robustness of the ID NIR method to mass line flow rates.
flows (<50 Kg/h).
When the 30 g of API is added in the inlet of blender 2 at different
3.3. Validation of RTD model blender 2 speeds it was observed that the spike is clearly seen at higher
blender 2 speeds and in contrary poorly defined at lower blender 2
The validation activity demonstrated the RTD model is valid for the speeds (Fig. 9a-c). However, when spike of 100 g is added in the inlet of
three tablet strengths 25 mg, 50 mg, and 100 mg across the process mass blender 1 at different speeds, it was observed there was not a significant
flow ranges of 15 kg/h to 90 kg/h (Hurley et al., 2022). Overall, the RTD impact in the concentration profile, but just slight change in maximum
model predictions showed good alignment with the HPLC composite peak intensity (Fig. 9d & e). Note that a higher speed of blender 2 the
12
Fig. 7. NIR API concentration estimation (% of target) in the blend during RTD spikes experiments on a) 100 mg, b) 50 mg and 25 mg images.
Fig. 8. Blend NIR predictions for API spikes of 30 g and 100 g at same conditions (speeds blender 1 and blender 2 were 315 rpm and 210 rpm, respectively and spike
location inlet of blender 2).
13
Fig. 9. Blend NIR predictions for runs at different blender speeds and different API spike amount.
profile of spike presents less disturbances compared to other conditions. successfully validated using an independent validation set. Samples
included in this model encompasses different sources of variability.
3.5. PAT applied to NOC batches Indeed, Fig. 10 shows the scatter plots for the PLS scores from calibra
tion and validation sample sets used in the blend NIR model 2 to esti
As previously presented in Table 5, blend NIR model 2 was mate the BU as %LC by using the blend NIR. Note that the calibration
Fig. 10. Scatter plot of PLS scores for blend NIR model 2. a) PC1 vs PC2, b) PC1 vs. PC3, c) projected scores for NOC batches on PC1 vs PC2 and d) PC1 vs PC3.
14
samples are distributed over the first principal component (Var. Exp = process is highly consistent across all batches being within the accep
85%) according to their content on API %LC (Fig. 10a) and the valida tance limits for BU. Outlier diagnostic metrics show no outliers. RTD
tion set used to validate the model is inside the calibration space. Whilst data for batches 6 & 7 are not available, however RTD results for the rest
PC 2 (Var. Exp = 6%) seems to be linked to indeed process variability, of the batches show that processes are in control and values predicted
PC3 (Var. Exp = 3%) is strong related to mass flow and tablet strength are around 100 %LC target.
differences (Fig. 10b). It is noted that for blend NIR measurements shown in Fig. 11 API %
The NOC batches were manufactured mainly at 21 and 75 Kg/h for LC predictions and %RSD, multiple brief transients are observed
tablet strengths 25 and 100 mg, respectively. As can be observed in throughout the entire runs. There was no evidence that these brief
Fig. 10c-d, despite the two tablet strengths are covered by the samples in transients were linked to any product quality impact as all IPC data
the PLS model, the two mass flows were not initially included in the (weight, hardness and NIR assay) for these batches were in control
model, this represents a challenge for the model robustness. All spectral limits. It was remarked that the transients were at regular intervals. The
data shown correspond to a period when press was in running status. timing and duration of the excursions strongly pointed towards the
Fig. 10c-d show the distribution of blend spectra which are within the material inlet hopper vibrator being engaged for brief periods of time-
calibration space demonstrating that most of the spectral variability is thus causing brief changes in the NIR diffuse reflectance spectra during
covered by the model. All NIR spectra are distributed around the sam runs. The observed transients evidenced that the vibrator was turned on
ples at target API concentration. to mitigate bridging risks. By consulting the Engineering Parameters
The estimation of BU by NIR is not part of the control strategy for Report which contains the relevant settings for the vibrator it was
CM, consequently no rejection limits were in place. However, it was observed that for NOC batches the vibrator was enabled. The vibrator
explored the possibility to monitor the BU via NIR by using qualitative starts when hopper is 5% filled and stops when hopper is filled at 95%.
and quantitative approaches. Hence, Fig. 11 shows the qualitative (% The waiting time setup to start automatically the vibrator is 120 s, which
RSD) and quantitative (%LC) estimations by the blend NIR and RTD means every 2 min the vibrator is activated by the CDC-50. The duration
predictions for the NOC batches. As can be seen, the three methods of each vibration session is one second. According to that information, it
allowed to extract equivalent conclusions and the results shown that could be more than one vibration within 2 min if hopper is at 5% filled
Fig. 11. Blend uniformity assessment of NOC batches by a) PLS model 2, b) % RSD for NOC batches and e) RTD process model. PLS predictions also provided outlier
diagnostic metrics such as c) Hotelling’s T2 (T2crit(α=0.01) = 11.4) and d) DModX (DmodXLimit = 3.2).
15
for example, which could explain the presence of regular disturbances in required to identify and find a method to reduce/remove this impact on
the API %LC predictions and %RSD plots. Because baseline spectral the NIR spectra. One recommendation would be active outlier filtering
changes are heavily influenced by the density of the blend material at of blend NIR outputs with the adequate limits during operation which
the probe surface, it is apparent that there is a periodic change in blend will likely be required to avoid these artefacts during routine operation
density at the Fprob surface occurring throughout the run, and some as the use of vibrator material inlet hopper is expected to be activated for
brief changes in blend density at regular intervals, which seems to be future manufacture to facilitate the flow of blend.
strongly linked to vibrator. Previous investigations have been done on For instance, Fig. 11a, shows a drop on NIR predictions across the
the specific nature of the changes in the NIR spectra when the vibrator is batch 3, slightly similar scenario is observed for batch 4, but with lower
on although these are predominantly driven by changes in baseline effect. Note that for those batches an increase of disturbances is observed
offset and multiplicative scatter. Further investigations would be in %RSD (Fig. 11c), this could be linked to multiple stops and re-start
Fig. 12. Assessment of %RSD for Batch 8. a) Blend NIR spectra, b) Moving block %RSD plot for entire batch, c) the spectral data and d) RMS distance from zero
results near spectrum number 140 and e) the reference spectral data from the time of the baseline shift.
16
due to reported filling issues with excipient EX1, blockage of lubricant, 3.6. IPC tests on core tablets
real time optimization activities to improve the powder flow, issues with
sensor vacuum receiving hopper in the LIWF (down from IBC into the The in-process checks monitored throughout the process were core
feeder) or activation of material inlet hopper vibrator due to issues tablet weight, hardness and disintegration (latter not discussed in this
related to powder flow instabilities related to PDV parameter settings paper). Tablets were automatically collected by the tablet analyser at a
which interrupted steady powder flow through the press. Indeed, these frequency of 15 min. For most of the NOC batches, ten weight, five
were the first 25 mg tablet strength batches manufactured with the CM hardness, three for API concentration and identity by NIR and finally six
line and knowledge of correct parameters for the PDV was limited. Later disintegration data have been generated at each sampling point.
development allowed to understand correct functioning of PDV and Nevertheless, the exception was for batch 5 where multiple number of
allow to vent the powder flow more efficiently enable consistent powder samples were withdrawn going up to thirty-four weights, seventeen
flow. Thus, as depicted after batch 5, an improvement on powder flow hardness and 10 API concentration and identity generated per sample
was achieved in later batches (batches 8 and 9). Fig. 11e shows RTD set. The individual tablet assay prediction was monitored between 75.0
predictions, it is observed RTD results has less disturbances are results and 125.0 % of target concentration, this approach is aligned with UDU
are consistent throughout the batches and are around target 100% LC. criteria in USP 905 and Eur. Ph 2.9.40 for individual dosage units. Pri
mary control for maintaining target API concentration includes process
3.5.1. Interpretation and troubleshooting on %RSD plots loss-in-weight feeder control loops and the RTD models which could
To illustrate the interpretation of moving block %RSD results, batch evaluate fluctuations in the loss-in-weight feeder rate as RTD model
8 is used as example. The mean residence time (MRT) for this batch was predicts API concentration based on the LIWFs.
calculated for the blenders using the mass flow rate of 21.0 Kg/h, Fig. 13a & b show the average tablet weight data across the 100 mg
blender 1 & blender 2 speeds of 315 rpm and 210 rpm, respectively. The and 25 mg runs, respectively. As can be seen, there are no obvious
calculated block size was eleven spectra for these process conditions. adverse trends within the parts of the run or across the runs as a whole.
The average %RSD result for the entire batch 8 was 0.55% which is All samples collected meet the acceptance criteria for average tablet
within the RSD limit. For instance, Fig. 12a shows blend NIR spectra for weight of 380 – 420 mg for the 100 mg strength and 95 – 105 mg for the
entire batch 8, which look typical to the product. However, in the %RSD 25 mg strength.
plot (Fig. 12b) it could be observed one location with high %RSD of Moreover, all samples collected meet the acceptance criteria for
spectral variation around spectrum #140. The value is within the limit average tablet hardness; ≥ 4.5kP for the 25 mg strength and ≥ 8.9 kP for
and it is not a considered a disturbance, however it is enough to inves the 100 mg strength. For example, Fig. 13c and d show the average
tigate further. Firstly, by examining the spectra and RMS distance from tablet hardness data across the 100 mg and 25 mg runs, respectively. As
zero (Fig. 12c & d), it can be seen the event occurs at a pause in the seen, mean hardness values are well above their minimum specification
process based on press stop/start times. Indeed, at restart, the baseline limits. Despite mean hardness values for batch 6 shifted higher during
noticeably increased and the RMS distance from zero decreased. Finally, manufacturing, this is not atypical for a longer compression run as the
by reviewing the reference spectra for batch 8, it was identified a new equipment warms and tooling beds settle in. In addition, as mean and
reference was collected at this time. However, the difference (Fig. 12e) individual results are above specification limits, the slightly higher
does not appear to be significant enough to cause the baseline shift. hardness values encountered in batch 6 and 8 do not represent a process
Therefore, data suggested a change in powder density at result of concern.
restarting the process which would be responsible for the baseline shift. Individual tablet weights for each 100 mg batch can be observed in
the histogram shown in Fig. 14a. From the results, it is clear values for
all 100 mg batches were centred around the target (target = 400 mg),
with no individual weight values outside of internal control limits (i.e.,
Fig. 13. Trend plot of average weight and hardness for NOC batches for 100 mg image (a & c) and 25 mg image (b & d). Red dot lines are specification limits and
green lines are target values. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
17
Fig. 14. Histograms of individual weight, hardness and NIR assay for 100 mg batches (a & c respectively) and 25 mg batches (b & d respectively). USL, upper
specification limit; LSL, lower specification limit.
T2 limits = 380 – 420 mg). This data indicates good process control (Fig. 15f) shows batches exhibited a relatively wider distribution cen
during manufacture of the 100 mg strength batches. On the other hand, tred close to target (100%LC) with no individual %LC being outside of
histogram for individual tablet weights for each 25 mg batch (Fig. 14b) the 85 – 115 %LC range. In conclusion, data indicates good process
shows a relatively tight distribution centred close to target (target = 100 control of NIR assay during manufacture for the 100 mg and 25 mg
mg). However, it was observed two individual weight values outside of batches; in addition all samples analysed by NIR passed the ID test.
internal control limits (i.e., T2 limits = 92.5 – 107.5 mg); 1 tablet in As part of sampling testing, three tablets were collected across all
batch 5 with a weight of 91.9 mg and another in batch 6 with a weight of batches every 15 min to test for assay and identity. However, batch 6
91.8 mg. These samples were handled according to site procedures with was an exception as ten tablets were collected for assay and identity at
no risk to the process. Overall, the results demonstrated good process each sampling point. The tablet assay results (%LC) and corresponding
control for tablet weight for the six NOC 25 mg batches. outlier diagnostic for nine batches are illustrated on Fig. 16. As illus
Regarding individual tablet hardness, the histogram of 100 mg trated on Fig. 16a, all NIR predictions fell within the IPC specification
batches (Fig. 14c) shows values were centred around the target (target for individual %LC (75 – 125 %LC), which evidence the robust predic
= 16kP), with no individual hardness values below the minimum control tion of the NIR models. Nevertheless, it was noted that for most of the
limit (≥8.9 kP). Thus, this data indicates good process control during samples collected from process operated at 21 Kg/h mass line flow rate,
manufacture of the 100 mg strength batches. In contrast, histogram for such as batches 3, 4, 7 and 8, the outlier diagnostic MD values were
individual tablet hardness for each 25 mg batches (Fig. 14d) illustrates higher than initial model limits, which suggests a source of variation
two clusters of hardness values; one cluster for batches 3, 4, 5 and 7 from these batches not being full present into the 25 mg NIR model. In
centred around 7kP and a second cluster for batches 6 and 8 centred near this kind of situations an NIR model update is recommended to include
8.5kP. Overall, the individual hardness data indicate good batch control this new source of variation as other IPC values were inside specifica
as all hardness values across all 25 mg batches were above the minimum tions. However, it is important to highlight that 25 mg NIR model
control limit of ≥ 4.5kP. demonstrated to be robust as batch 5 was parallel tested by reference
In respect of core tablet NIR Assay and Identity results, Fig. 15 shows method HPLC. Results shown that NIR model predicts correctly with a
the comparison of pre-treated NIR spectra for 100 mg and 25 mg batches RMSEP of 1.17 %LC being acceptance criterion 2.5 %LC. Thus, the
and their NIR model calibration sample sets (Fig. 15a & b, respectively). source of variation is not related to API content, but mainly to physical
Note that spectral region is the calibration region as described in properties of the tablets.
Table 8. This region includes the main API spectral band located at As shown in Fig. 16, all NIR assay data was within the acceptance
approximately 8800 cm− 1 (note the up and down arrows indicating criteria for individual tablet assay of 85.0 –115.0 %LC. The results were
movement with increasing concentration of active). As illustrated in centred around the target assay (target = 100 %LC), with no obvious
Fig. 15, all core tablets collected throughout the nine NOC batches are adverse trends either within the parts of the run or across the run as a
within the spectral variation exhibited by the calibration samples. whole. The data indicates good process control for the run with respect
Fig. 15c & d show the scatter plots of PLS scores corresponding to NOC to tablet assay. Note that batch 6 shows a noticeable drop off in all re
batches and the NIR model calibration sample sets for 100 mg and 25 mg sults for the final IPC sample taken as well as a sudden increase of the
tablet strengths, respectively. As seen, samples collected from NOC MD outlier diagnostic metric. The issue was reviewed and small differ
batches are within the calibration space and are well covered by the ences in the spectra for the final three tablets compared to the spectra for
spectral variability built into the model. the other tablets tested across the batch were identified. The review
In the matter of individual tablet NIR assay values for each of the concluded that the variation is small, and API estimation is within the
100 mg batches, histogram of NIR values (Fig. 15e) shows batches were production limits which implies that this event does not have an impact
tightly controlled and centred around the target (100% LC) with no on the quality of the product. Moreover, this is a single occurrence
individual %LC being outside of the 85 – 115 %LC range. Contrastingly, observed across the manufactured batches and the analysis implies that
histogram of individual NIR values for each of the 25 mg batches it could be triggered by a complex combination of factors, with a low
18
Fig. 15. NIR spectra (a & b) and scatter plot of PLS scores for NOC batches samples and calibration sample set (c & d) for 100 mg and 25 mg images, respectively.
probability of recurrence. Other than for the final IPC set, there were no demonstrates the robust process in place for managing the NIR models.
other obvious adverse trends in IPC data across the batch, in conclusion Model updates may be executed based on the outcomes of routine, pe
the data indicating good process control during manufacture of the riodic or event driven model verifications. An existing NIR model is
batch. typically updated by the addition of samples to both the calibration and
validation sets that represent new variation to improve model robust
3.7. PAT model lifecycle management ness. In certain cases, removal of existing samples from the mid-point of
the calibration and validation sets is also considered when that vari
For regulatory purposes, MSD differentiates the NIR method from the ability is no longer present (e.g., change of raw material supplier). Some
supporting model. The NIR model is expected to change over the of the model updates may also include changes to model parameters that
product lifecycle and not the NIR method. The process for developing may include modifications to spectral quality check algorithm threshold
and maintaining the NIR model and method is similar irrespective of the limits, spectral pre-processing and PLS model parameters (calibration
manufacturing process and the 10-year experience from the batch pro range and number of latent variables) to account for the different
cess will be leveraged for the CM process. The comprehensive process spectral variability in the new calibration and validation sample popu
used to develop NIR models includes the risk assessment, design of ex lation. These changes are necessary to ensure model accuracy
periments, risk mitigations and preventative maintenance. As show in throughout the lifecycle.
Fig. 17, the NIR model lifecycle approach includes routine verification
on individual tablets using outlier diagnostics, periodic verification, 3.8. NIR assay model verification
event driven model verification and validation. The manufacturing site
procedures for each of the steps in the NIR model lifecycle management Three types of NIR model verifications are performed at the
19
Fig. 16. NIR assay outputs for NOC batches a) API predictions in %LC, b) Outlier diagnostic MD and c) Outlier diagnostic Fprob with corresponding limits.
Fig. 17. Diagram of CM NIR model lifecycle management.
manufacturing site for API assay; routine, periodic and event based. Germany). The use of these metrics is recommended by ASTM standard
D-6122 (ASTM, 2021), as they express different outlier behaviour. These
3.8.1. Routine verification: Model outlier diagnostics values are obtained for each, and every sample analysed. MD expresses
Routine NIR model verification is performed on each sample spec the distance of an unknown spectrum from the centre of the spectra used
trum by applying outlier diagnostics. A major benefit of multivariate to generate the multivariate calibration model. When compared to the
calibration modelling, such as partial least squares methodology, is the MD of the calibration spectra, this metric can provide an assessment of
inherent ability to provide an indicator of the quality of the model whether an unknown sample spectrum differs greatly from the popula
prediction through outlier diagnostics for each sample. The two outlier tion of calibration spectra. In contrast, the spectral residual outlier
diagnostics that will be implemented for NIR assay model verification diagnostic, Fprob ratio expresses the extent to which the unknown
are MD and Fprob, these values are calculated by OPUS (Bruker, spectrum cannot be explained by the multivariate calibration model.
20
The Fprob is the probability that the unknown’s spectral residual is physicochemical properties or instrument performance variations, a li
significantly greater than those of the calibration spectra. Thus, this brary update will be executed. A library update typically involves the
diagnostic provides an assessment of whether the unknown spectrum addition of samples into the library and positive control data sets but
contains absorptions do not present in the calibration spectra. Action may also include the removal of samples. The suitability of a library
limits are established in the PMS method to flag spectra as outliers based update will involve verification of validation elements.
upon internationally accepted standards for multivariate quantitative
analysis codified in ASTM standards E-1655 (ASTM, 2017) and D-6122
(ASTM, 2021) and aligned with ICH guideline Q2(R1) (ICH, 2005) for 3.10. RTD process model lifecycle management
the validation of analytical methods. These standards and guidelines
provide the framework from which a multivariate quantitative calibra Leveraging well established principles for PAT models, the RTD
tion can be developed, validated and maintained. Site procedures have process model is managed in a similar fashion. Lifecycle management is
been developed and implemented to manage the use and maintenance of designed to continuously assess the performance of the RTD process
outlier diagnostics. If a sample spectrum has been detected as an outlier model and adjust model parameters as needed to ensure predictive
by either of the diagnostics noted above, site procedures will be performance. The RTD model was developed, verified, and validated at
executed to evaluate and verify the NIR results. Such activities can the Cramlington manufacturing site. To maintain the RTD process model
include exploratory multivariate analysis of NIR spectral data, along as a robust element of the control strategy for the CM product, it is
with associated process data and event data on site, to assess root cause anticipated that periodic model verification and validation will be
of the outlier and to identify possible method improvement required on an ongoing basis to maintain the model relative to the CM
opportunities. process. Periodic verification of the RTD process model will be per
formed at a regular frequency and will be governed by site procedures
3.8.2. Periodic and event based verification within the pharmaceutical quality system (PQS). Model verification will
Periodic model verification is performed to assure predictive per include an intentional variation from target API concentration and a
formance on a periodic basis (Eur. Ph. 9.5, 2018a) and is performed a comparison of the concentration prediction by the RTD process model to
minimum of twice per year. Event based model verification will be measured tablet assay at approximately t10, t50, and t90. The root mean
triggered by changes to the NIR instrument (preventative maintenance), squared error of prediction (RMSEP) is calculated and assessed against
or changes to the process (new raw material supplier, change in the previously defined acceptance criteria. The performance of the RTD
manufacturing equipment). The procedure for both cases is similar; it process model, i.e., the RMSEP from the periodic verification, is trended
includes running a statistically significant sample population (20 sam and may prompt model update and validation when appropriate.
ples or greater) using a minimum of one commercial production batch Additionally, tablet weight variance, an input to RTD reject limits, will
with both the NIR and HPLC reference methods. A statistical comparison be monitored and adjusted, as appropriate, as part of ongoing model
of the parallel tested results will be completed and compared to the maintenance. Examples of events that may result in RTD model verifi
established model performance criteria. cation or validation are described in Table 10 and will be documented
via change control or as part of a deviation investigation.
3.8.3. Model updates Examples of a major equipment change include, but are not limited
If the outcome of an investigation, triggered by routine or periodic to, a change in blending equipment or orientation of mixing blades in
NIR model verification failure, significant changes in materials, equip the blender which may have the potential to affect the RTD of the system
ment and/or manufacturing process, erroneous NIR result, or trends/ outside of the previously validated range. An example of a minor
shifts in NIR model predictions, indicate an un-modelled tablet physi equipment modification includes, but is not limited to, a change in
cochemical properties or instrument performance variation, a model blender speed which may have the potential to affect the RTD of the
update will be executed. An existing model is typically updated by the system, but within the previously validated range. Equipment modifi
addition of samples to both the calibration and validation sets that cations such as changes in feeder set-up or configuration, for example,
represent new variation to improve model robustness. In certain cases, are not expected to impact the RTD of the system and would not require
removal of existing samples from the calibration and validation sets is verification or validation of the RTD process model.
also considered when that variability will no longer be present (e.g., To conclude, all work done after CM line installation and qualifica
change of raw material supplier). Some of the model updates may also tion allowed getting experience on the equipment functionality and PAT
include changes to model parameters that may include modifications to (RTD and NIR) model maintenance. All learnings and robust results
spectral range, model rank or spectral pre-treatment to account for the obtained with PAT and control strategy supported approval of the
different spectral variability in the new calibration and validation comparability protocol for use of CM as an alternative manufacturing
sample population. All such modified models must then be validated and
approved before they can be returned to service. The model performance Table 10
criteria to validate an updated NIR model are identical to those estab Events prompting RTD process model verification and (re)validation.
lished during the original model validation. Example event Anticipated Response to
frequency of event
occurrence
3.9. Identify NIR model verification Raw material vendor change Low Model
verification
Raw material specification change Low Model
3.9.1. Method and library verification verification
The performance of the identification method is periodically (mini Minor equipment or automation Medium Model
mum of once per year) verified using positive and/or negative controls. modification verification
Positive control samples (of known origin and corresponding to active Out of specification tablet NIR result 1 Low Model
verification
core tablets) must provide a passing result. Negative control samples
Major equipment changes Low Model (re)
(tablet placebo) must provide a non-passing test result. validation
Periodic RTD process model Low Model (re)
3.9.1.1. Library update. If method verification results or if the outcome verification RMSEP Above validation
acceptance criteria
of an out-of-specification (OOS) investigation indicates that a NIR li
brary update is necessary to incorporate un-modelled tablet 1
Not related to NIR model robustness.
21
process to the batch process. This achievement, result over several years, Gascoigne, Steven Appleby, David Todd, Steven Russell, Abbey Sum
has provided benefits not just to CM, but to processes in which control mers. From Merck US: Frank Witulski, Laura Wareham, Bob Meyer,
elements such as NIR or predictive RTD process models provide benefits Anita Lalloo.
to quality assurance and flexible, compliant manufacturing.
References
4. Conclusions
Almaya, A., De Belder, L., Meyer, R., Nagapudi, K., Lin, H.-R.-H., Leavesley, I.,
Jayanth, J., Bajwa, G., DiNunzio, J., Tantuccio, A., Blackwood, D., Abebe, A., 2017.
The continuous manufacturing equipment, fully automated with Control Strategies for Drug Product Continuous Direct Compression—State of
PAT for real time monitoring, real time rejection and real time release Control, Product Collection Strategies, and Startup/Shutdown Operations for the
testing, and successfully integrated with the existing site infrastructure Production of Clinical Trial Materials and Commercial Products. J. Pharm. Sci. 106,
930–943.
was demonstrated to be fully functional under commercial and Quality ASTM, 2017. ASTM E1655-17 Standard Practices for Infrared Multivariate Quantitative
systems and consistent with the principles of data integrity, system Analysis.
lifecycle and GMP qualification. As part of this journey, it was demon ASTM, 2021. ASTM D6122-21. Standard Practice for Validation of the Performance of
Multivariate Online, At-Line, Field and Laboratory Infrared Spectrophotometer, and
strated that CM Technology for OSD in both development campaigns Raman Spectrometer Based Analyzer Systems.
and prior PPQ commercial production provide manufacturing benefits Badman, C., Cooney, C.L., Florence, A., Konstantinov, K., Krumme, M., Mascia, S.,
versus batch manufacturing, and that fully flexible batch sizes can be Nasr, M., Trout, B.L., 2019. Why We Need Continuous Pharmaceutical
Manufacturing and How to Make It Happen. J. Pharm. Sci. 108, 3521–3523.
produced with consistent attributes. Notably, NIR models performed
Berntsson, O., Burger, T., Folestad, S., Danielsson, L., Kuhn, J., Fricke, J., 1999. Effective
well as predictions were robust and allowed to identify sources of Sample Size in Diffuse Reflectance Near-IR Spectrometry. Anal. Chem. 71, 617–623.
variability and act in a timely manner. Overall, RTD model validation Cárdenas V., Rosas J.G., Pinzón C., Romañach R.J., 2020. Statistical Methods in Quality
and integration into the control scheme for real time monitoring and by Design and Process Analytical Technologies for Continuous Processes to Enable
Real-Time Release, in: Nagy Z., E.H.A., Litster J. (Ed.), Continuous Pharmaceutical
rejection was successful. Hence, the RTD model performed well and Processing - AAPS Advances in the Pharmaceutical Sciences Series, pp. 361-393.
demonstrated consistency over all normal operational condition runs. De Leersnyder, F., Peeters, E., Djalabi, H., Vanhoorne, V., Van Snick, B., Hong, K.,
Indeed, RTD model results indicated that tablet predictions were around Hammond, S., Liu, A.Y., Ziemons, E., Vervaet, C., De Beer, T., 2018. Development
and validation of an in-line NIR spectroscopic method for continuous blend potency
the target value with no drift throughout the batches observed. As a determination in the feed frame of a tablet press. J. Pharm. Biomed. Anal. 151,
result of all this work, this novel RTD and PAT-based control strategies 274–283.
supported approval of the comparability protocol for CM. In addition, De Leersnyder, F., Vanhoorne, V., Kumar, A., Vervaet, C., De Beer, T., 2019. Evaluation
of an in-line NIR spectroscopic method for the determination of the residence time in
the batch-size flexibility and make-to-order approach was successfully a tablet press. Int. J. Pharm. 565, 358–366.
integrated with existing GMP systems. Finally, NOC runs allowed the EMA, 2014. Guideline on the Use of Near Infrared Spectroscopy by the Pharmaceutical
identification and remediation of several PAT issues: CDC-50 to PAT Industry and the Data Requirements for New Submissions and Variations EMEA/
CHMP/ CVMP/ QWP/ 17760/ 2009 Rev2.
instrument automation issues, data integrity issues with transfer of data Eur. Ph. 9.5, 2018a. Chapter 2.2.40. Near-infrared spectroscopy.
to the PMS and improved processes for the calibration of the tablet FDA, 2015b. Development and Submission of Near Infrared Analytical Procedures -
analyser. The resolution of these issues demonstrated that innovative Guidance for Industry.
FDA-EMA, 2013. EMA-FDA pilot program for parallel assessment of Quality-by-Design
methods and technologies were successfully implemented and validated
applications: lessons learnt and Q&A resulting from the first parallel assessment.
in a GMP facility. Fogler, H.S., 2016. Elements of Chemical Reaction Engineering, 5th ed. Prentice Hall,
Philadelphia, PA.
García-Muñoz, S., Butterbaugh, A., Leavesley, I., Manley, L.F., Slade, D., Bermingham, S.,
CRediT authorship contribution statement
2018. A flowsheet model for the development of a continuous process for
pharmaceutical tablets: An industrial perspective. AIChE J. 64, 511–525.
Juan G. Rosas: Conceptualization, Methodology, Software, Formal Gregory, A.P., Debasis, M., Michael, G.O., 2017. Process analytical tools for flow
analysis, Investigation, Writing – original draft, Validation. Peter analysis: A perspective. J. Flow Chem. 7, 82–86.
Hailey, P.A., Doherty, P., Tapsell, P., Oliver, T., Aldridge, P.K., 1996. Automated system
Brush: Conceptualization, Methodology, Software, Formal analysis, for the on-line monitoring of powder blending processes using near-infrared
Investigation, Writing – review & editing, Validation. Bruce Thomp spectroscopy part I. System development and control. J. Pharm. Biomed. Anal. 14,
son: Conceptualization, Methodology, Software, Formal analysis, 551–559.
Holman, J., Tantuccio, A., Palmer, J., van Doninck, T., Meyer, R., 2021. A very boring
Investigation, Writing – review & editing, Validation. Charles Miller: 120 h: 15 million tablets under a continuous state of control. Powder Technol. 382,
Conceptualization, Methodology, Formal analysis, Investigation, Vali 208–231.
dation. Paul Overton: Investigation, Writing – review & editing. Neil Hurley, S., Tantuccio, A., Escotet-Espinoza, M.S., Flamm, M., Metzger, M., 2022.
Development and Use of a Residence Time Distribution (RTD) Model Control
Tugby: Investigation, Writing – review & editing. Daria Stoliarskaia: Strategy for a Continuous Manufacturing Drug Product Pharmaceutical Process.
Visualization, Writing – review & editing. Samantha Hurley: Formal Pharmaceutics 14.
analysis, Investigation. Manoharan Ramasamy: Conceptualization, ICH, 2005. Tripartite Guideline: Validation of Analytical Procedures: Text and
Methodology: Q2(R1).
Methodology, Project administration, Writing – review & editing. Ste
ICH, 2019a. Technical and regulatory considerations for pharmaceutical product
phen L. Conway: Conceptualization, Supervision, Project administra lifecycle management Q12.
tion, Writing – review & editing. ICH, 2019b. Technical and regulatory considerations for pharmaceutical product
lifecycle management Q12 Annexes.
Järvinen, K., Hoehe, W., Järvinen, M., Poutiainen, S., Juuti, M., Borchert, S., 2013. In-
Declaration of Competing Interest line monitoring of the drug content of powder mixtures and tablets by near-infrared
spectroscopy during the continuous direct compression tableting process. Eur. J.
The authors declare that they have no known competing financial Pharm. Sci. 48, 680–688.
Johansson, J., Folestad, S., Josefson, M., Sparén, A., Abrahamsson, C., Andersson-
interests or personal relationships that could have appeared to influence Engels, S., Svanberg, S., 2002. Time-Resolved NIR/Vis Spectroscopy for Analysis of
the work reported in this paper. Solids: Pharmaceutical Tablets. Appl. Spectrosc. 56, 725–731.
Levenspiel, O., 1993. The Chemical Reactor Omnibook, Oregon St Univ Bookstores.
McKay, A.T., 1932. Distribution of the Coefficient of Variation and the Extended “t”
Data availability Distribution. J.R. Stat. Soc. 95, 695–698.
Nasr, M.M., Krumme, M., Matsuda, Y., Trout, B.L., Badman, C., Mascia, S., Cooney, C.L.,
The data that has been used is confidential. Jensen, K.D., Florence, A., Johnston, C., Konstantinov, K., Lee, S.L., 2017. Regulatory
Perspectives on Continuous Pharmaceutical Manufacturing: Moving from Theory to
Practice: September 26–27, 2016, International Symposium on the Continuous
Acknowledgements Manufacturing of Pharmaceuticals. J. Pharm. Sci. 106, 3199–3206.
Nelson, L.S., 1984. The Shewhart Control Chart—Tests for Special Causes. J. Qual.
Technol. 16, 237–239.
The authors would like to thank all people from MSD Cramlington
involved in the CM project including Tony Robinson, Andrew
22
Palmer, J., Reynolds, G.K., Tahir, F., Yadav, I.K., Meehan, E., Holman, J., Bajwa, G., Van Snick, B., Holman, J., Vanhoorne, V., Kumar, A., De Beer, T., Remon, J.P.,
2020. Mapping key process parameters to the performance of a continuous dry Vervaet, C., 2017b. Development of a continuous direct compression platform for
powder blender in a continuous direct compression system. Powder Technol. 362, low-dose drug products. Int. J. Pharm. 529, 329–346.
659–670. Vanarase, A.U., Alcalà, M., Jerez Rozo, J.I., Muzzio, F.J., Romañach, R.J., 2010. Real-
Pauli, V., Roggo, Y., Pellegatti, L., Nguyen Trung, N.Q., Elbaz, F., Ensslin, S., time monitoring of drug concentration in a continuous powder mixing process using
Kleinebudde, P., Krumme, M., 2019. Process analytical technology for continuous NIR spectroscopy. Chem. Eng. Sci. 65, 5728–5733.
manufacturing tableting processing: A case study. J. Pharm. Biomed. Anal. 162, Vanarase, A.U., Järvinen, M., Paaso, J., Muzzio, F.J., 2013. Development of a
101–111. methodology to estimate error in the on-line measurements of blend uniformity in a
Sekulic, S.S., Ward, H.W., Brannegan, D.R., Stanley, E.D., Evans, C.L., Sciavolino, S.T., continuous powder mixing process. Powder Technol. 241, 263–271.
Hailey, P.A., Aldridge, P.K., 1996. On-Line Monitoring of Powder Blend Vanhoorne, V., Vervaet, C., 2020. Recent progress in continuous manufacturing of oral
Homogeneity by Near-Infrared Spectroscopy. Anal. Chem. 68, 509–513. solid dosage forms. Int. J. Pharm. 579, 119194.
Shi, Z., McGhehey, K.C., Leavesley, I.M., Manley, L.F., 2016. On-line monitoring of blend Vargas, J.M., Roman-Ospino, A.D., Sanchez, E., Romañach, R.J., 2017. Evaluation of
uniformity in continuous drug product manufacturing process—The impact of analytical and sampling errors in the prediction of the active pharmaceutical
powder flow rate and the choice of spectrometer: Dispersive vs. FT. J. Pharm. ingredient concentration in blends from a continuous manufacturing process.
Biomed. Anal. 118, 259–266. J Pharm Innov. 12, 155–167.
Sierra-Vega, N.O., Román-Ospino, A., Scicolone, J., Muzzio, F.J., Romañach, R.J., Vargas, J.M., Nielsen, S., Cárdenas, V., Gonzalez, A., Aymat, E.Y., Almodovar, E.,
Méndez, R., 2019. Assessment of blend uniformity in a continuous tablet Classe, G., Colón, Y., Sanchez, E., Romañach, R.J., 2018. Process analytical
manufacturing process. Int. J. Pharm. 560, 322–333. technology in continuous manufacturing of a commercial pharmaceutical product.
USP38, 2015. General Chapter <1119> Near-Infrared Spectroscopy. Int. J. Pharm. 538, 167–178.
Van Snick, B., Holman, J., Cunningham, C., Kumar, A., Vercruysse, J., De Beer, T., Wahlich, J., 2021. Continuous Manufacturing of Small Molecule Solid Oral Dosage
Remon, J.P., Vervaet, C., 2017a. Continuous direct compression as manufacturing Forms. Pharmaceutics 13, 1308–1311.
platform for sustained release tablets. Int. J. Pharm. 519, 390–407.
23

1 s2.0 S037851732300234X Main

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1 s2.0 S037851732300234X Main

Uploaded by

Copyright:

Available Formats

International Journal of Pharmaceutics 636 (2023) 122814

Contents lists available at ScienceDirect

International Journal of Pharmaceutics

Implementation of a fully integrated CM direct compression and coating

1. Introduction and operational philosophy associated with a product, it can be used to

Fig. 3. High-level PAT/automation system architecture and communication.

1) the API concentration prediction by the RTD model and

2.4.2. Development batches

potencies. Tablet age is defined as the approximate elapsed time from

Specificity Positive controls identified as active tablets 33 33 36

Fig. 17. Diagram of CM NIR model lifecycle management.

You might also like