Segment-Level Gait Dynamics Reveal Early Parkinsonian Motor Signatures in Fall-Prone Older Adults
Segment-Level Gait Dynamics Reveal Early Parkinsonian Motor Signatures in Fall-Prone Older Adults
Where did the research take place?
The study site has not been established. Author addresses may differ from where the research occurred.
Publication status: preprint
A plain-language reading has not been prepared for this paper yet.
Original abstract
Abstract Background Early detection of Parkinson’s disease (PD) is limited by the late emergence of overt motor symptoms, while subtle gait abnormalities may appear years earlier. Older adults with recurrent unexplained falls have been suggested as a population with elevated prodromal PD risk, yet objective tools to quantify PD-like motor dynamics in this group remain scarce. We aimed to develop and validate a multi-channel time–frequency deep learning framework for segment-level identification of PD-like gait dynamics. Furthermore, we aimed to derive subject-level, model-defined transition metrics that characterize sustained PD-like motor states in fall-prone older adults. Methods Overall, 211 older adults (PD, n = 50; fall-prone older adults, n = 161) who completed inertial sensor–based gait assessment under slower (80%), preferred, and faster (120%) walking speeds were included. Data from 16 bilateral inertial measurement unit channels were transformed into continuous wavelet transform scalograms and used to train single- and multi-channel convolutional neural network models using subject-wise stratified five-fold cross-validation. Subject-level “transition” was defined as entry into a sustained PD-like state using smoothing and a hysteresis rule, with an operating threshold selected to control false positives (target false-positive rate [FPR] 5%). Results The multi-channel model achieved consistently high segment-level discrimination across speeds (area under the receiver operating characteristic curve [AUC]: faster 0.992 ± 0.007; preferred 0.974 ± 0.016; slower 0.975 ± 0.017), outperforming representative single-channel models (AUC: 0.65–0.75). Subject-level transition detection under the 5% FPR operating point yielded high sensitivity (0.96–1.00) with strong specificity (0.857–0.913) across speeds, while relaxing the FPR constraint to 10% substantially reduced specificity. Conclusions Multi-channel time–frequency deep learning enables robust identification of sustained PD-like gait states from short walking segments and provides interpretable model-derived transition metrics for motor risk stratification in fall-prone older adults.