Parkinson's Disease Diagnosis and Severity Assessment from Gait Signals via Bayesian-Optimized Deep Learning

Background/Objectives: Early diagnosis of Parkinson's Disease (PD) is essential for initiating interventions that may slow its progression and enhance patient quality of life. Gait analysis provides a non-invasive means of capturing subtle motor disturbances, enabling the prediction of both disease presence and severity. This study evaluates and contrasts Bayesian-optimized convolutional neural network (CNN) and long short-term memory (LSTM) models applied directly to Vertical Ground Reaction Force (VGRF) signals for Parkinson's disease detection and staging. Methods: VGRF recordings were segmented into fixed-length windows of 5, 10, 15, 20, and 25 s. Each segment was normalized and supplied as input to CNN and LSTM network. Hyperparameters for both architectures were optimized via Bayesian optimization using five-fold cross-validation. Results: The Bayesian-optimized LSTM achieved a peak binary classification accuracy of 99.42% with an AUC of 1.000 for PD versus control at the 10-s window, and 98.24% accuracy with an AUC of 0.999 for Hoehn-Yahr (HY) staging at the 5-s window. The CNN model reached up to 98.46% accuracy (AUC = 0.998) for binary classification and 96.62% accuracy (AUC = 0.998) for multi-class severity assessment. Conclusions: Bayesian-optimized CNN and LSTM models trained on VGRF data both achieved high accuracy in Parkinson's disease detection and staging, with the LSTM exhibiting a slight edge in capturing temporal patterns while the CNN delivered comparable performance with reduced computational demands. These results underscore the promise of end-to-end deep learning for non-invasive, gait-based assessment in Parkinson's disease.