IMU-Based Gait Analysis for Detection of Parkinson Disease Using Machine and Deep Learning Models

Parkinson's disease (PD) is the second most common neurodegenerative disorder worldwide, affecting especially elderly. Early diagnosis of PD remains challenging due to the subtle onset of symptoms. Traditional motor assessment tools, such as the MDS-UPDRS, often lack objectivity and sensitivity. This study evaluates Machine Learning (ML) and Deep Learning (DL) techniques combined with wearable inertial measurement unit IMU sensors to detect gait abnormalities associated with Parkinson. Several shallow ML models and DL architectures were trained on preprocessed IMU gait data partitioned at the subject level. Data augmentation techniques were applied exclusively to the training dataset to address class imbalance without information leakage. The LSTM model trained on augmented temporal signals achieved the highest performance, reaching an accuracy of 93.79%, a recall of 94.84%, a balanced accuracy of 93.78%, and a Matthews correlation coefficient (MCC) of 0.876. The MLP and Random Forest with DWT features achieved an accuracy of 92.48%, a recall of 93.55%, and an MCC of 0.850. Model robustness was further confirmed on an independent cross-site cohort achieving, LSTM: 93.10% accuracy, 94.16% recall; MLP: 92.13% accuracy, 93.07% recall. Furthermore, Explainable AI revealed that model decisions were predominantly driven by wrist rotational dynamics, physiologically corroborating asymmetric arm-swing reduction.