PrimSeq: A deep learning-based pipeline to quantitate rehabilitation training
Fig 4
(A) Sequence matching. Shown is a schematic depicting the different types of prediction outcomes, comparing the ground truth sequence (top row) against the predicted sequence (bottom row). Seq2Seq could produce a true positive (correct prediction; green), false negative (deletion or swap-out error; pink), or false positive (insertion or swap-in error; blue). In the example shown, transport was correctly predicted; stabilize was incorrectly deleted; reach was incorrectly swapped-out while idle was incorrectly swapped-in; and reach was incorrectly inserted. (B) Frequency of prediction errors per primitive class. Shown are the mean frequency ± standard deviation of prediction errors for patients in the test set; each dot represents a single subject. Activities were combined, and erroneous counts were normalized to ground truth counts in each primitive class. Deletion errors happened when primitives were incorrectly removed from the prediction, and occurred with modestly low frequency, except for stabilizations. Swap-out errors happened when primitives were incorrectly removed from the prediction and instead predicted as another class, and occurred with modestly low frequency. Insertion errors happened when primitives were incorrectly added to the prediction, and occurred with low frequency, except for stabilizations. Swap-in errors happened when primitives were incorrectly predicted instead of the actual primitive class, and occurred with modestly low frequency.
doi: https://doi.org/10.1371/journal.pdig.0000044.g004