Figure 1.

Overall study pipeline. (a) A training data-set arose in-the-clinic ($D{S}_2$) and used for training of the hybrid model along with a second training set ($D{S}_1$) arose from in-the-wild typing data remotely collected. (b) A hybrid model training using $D{S}_2$ and $D{S}_1$ produce best performing model for FMI estimation. (c) Separate data sets captured in-the-wild were used as test data for three test scenarios (T1, T2, T3); T1 correlates the estimated FMI severity with the clinical ground-truth, and the ability to classify early PD patients vs. HC in a test dataset ($T{S}_1$), T2 examines the classification performance to classify de novo PD patients vs. HC in a validation set ($T{S}_2$), a subset of $T{S}_1$, T3 examines the classification properties of the hybrid model in a large self-reported validation set, from the union of $T{S}_1$ and $T{S}_0$ ($T{S}_3$) against the subjects’ self-reported health status of being PD or not.