Adaptive frequency-time network with dual-path fusion for precise Parkinson’s disease staging from gait signals

Parkinson’s Disease (PD) is a common neurodegenerative disorder, with gait abnormalities serving as critical biomarkers for early diagnosis and severity assessment. Traditional gait analysis methods often rely on handcrafted features, which can be complex, expert-dependent, and lack generalizability. To address these limitations, this study introduces an Adaptive Frequency-Time Fusion Network (FTNet), which automatically extracts time-frequency features from gait signals for accurate PD staging and classification. The model integrates a multi-scale convolutional module with a cycle modeling component based on Fast Fourier Transform (FFT), facilitating a deep fusion of time-domain and frequency-domain features. This fusion enhances the model’s ability to capture dynamic gait changes. Experiments conducted on the PhysioNet gait dataset reveal that FTNet attains a classification accuracy of 90.06% and a macro F1 score of 0.90, outperforming the current deep learning models. Ablation studies confirm the complementary nature of time-domain and frequency-domain features. This research offers a novel approach for the early diagnosis and monitoring of PD.