Skip to main content
. Author manuscript; available in PMC: 2020 Jul 22.
Published in final edited form as: Patterns (N Y). 2020 May 28;1(3):100042. doi: 10.1016/j.patter.2020.100042

Figure 4. Visualization of Deep Convolutional Neural Networks Unveils the Importance of Quiet Standing Behaviors during Walking when Detecting Parkinson’s Disease.

Figure 4.

(A) The comparisons of AUROCs performed by models on walking records during outbound walklng (outbound), qulet standing (rest), and return walking (return).

(B) Pairwise comparison between AUROCs achieved by outbound walking (Outbound) and quiet standing (Return) models.

(C) Pairwise comparison between AUROCs achieved by return walking (Return) and quiet standing (Rest) models. Quiet standing consistently performed better than both outbound and return walking (p < 1 × 10−6).

(D) Saliency maps that the trained deep neural network extracted from walking records (after padding to 40 s) of both PD patients and healthy controls during outbound walking (Outbound), quiet standing (Rest), and return walking (Return). Ground-truth labels and predictions by the machine-learning models are denoted for each rotation-rate signal and corresponding saliency map extracted from each signal. On the right, PD characteristics of perturbed steps and tremors in comparison with controls are zoomed in to show them in detail.