HYBRID DEEP LEARNING APPROACH FOR PARKINSON’S DISEASE CLASSIFICATION USING MEL-SPECTROGRAMS GENERATED FROM PADS TIME SERIES DATA

Abstract

The increasing integration of smart devices—such as smartphones and smartwatches—into the study of movement disorders has opened new avenues for continuous, real-world monitoring. However, a significant limitation remains due to the lack of robust, clinically annotated datasets that encompass Parkinson’s Disease (PD) alongside its differential diagnoses (DD). To address this gap, we employed the publicly accessible PADS time series dataset from PhysioNet, which features sensor-derived motion data from individuals diagnosed with PD, DD, and healthy controls (HC).

In this study The raw motion sensor data were transformed into Mel-spectrograms to facilitate the extraction of meaningful frequency-based motor features. A hybrid deep learning architecture—combining WaveNet and ResNet50—was developed to classify these spectrograms into diagnostic categories.

Our model demonstrated high performance in distinguishing PD from healthy individuals, achieving an accuracy of 95.24%, an F1 score of 0.9524, a recall of 0.9565, and a precision of 0.9565. In contrast, differentiating PD from other movement disorders proved more challenging, with the model attaining an accuracy of 76.19%, an F1 score of 0.7608, recall of 0.8163, and precision of 0.7843.

These findings underscore the promise of deep learning applied to spectrogram representations for identifying Parkinson’s Disease using smart device sensor data. Moreover, they highlight the diagnostic complexity inherent in distinguishing PD from other neurologically similar conditions. The PADS dataset—with its rich clinical annotations, including demographics, symptom profiles, and medical histories—provides a valuable foundation for advancing machine learning research in the domain of movement disorder analysis.