FIGURE 3.

Expert (top) and automatic soring with one algorithm (CNN with 16-s window; bottom) of an MWT (40 min) in one patient of the validation set (patient y5We). A good match between the algorithm and an expert scoring for wakefulness (W) and microsleep episodes (MSE) are evident, but a poor match for episodes of drowsiness (ED) and microsleep episode candidates (MSEc). Scoring was performed with the resolution of one sample; for the illustration, we coarsened the result to a resolution of 0.5 s (100 samples), i.e., the most frequent class within an interval was plotted. Results for other patients of the validation set are illustrated in Supplementary Figure S1, those of the test set in Supplementary Figure S2.