Dominant Hand Invariant Parkinson's Disease Detection Using 1-D CNN Model and STFT-based IMU Data Fusion

Aleksei Shcherbak, Ekaterina Kovalenko, Ekaterina Bril, Anna Baldycheva, Andrey Somov

Abstract: In this paper, we report on a dominant hand invariant 1-D convolution model for distinguishing between Parkinson’s disease (PD) and healthy subjects. For realizing this approach, we propose an STFT-based method for IMU data axes fusion. We learn how both the different frequency ranges and the period of the STFT window affect the efficiency of Parkinson’s disease detection. We test this solution on the dataset collected from 58 subjects. Our results show superiority of the proposed axes fusion method in 70% of cases in comparison with the state-of-the-art. Results also prove the efficiency of the proposed 1-D convolution hand invariant model with the best scores 98% of AUC and 92% of F1 and accuracy metrics. In addition, we show the STFT window must be at least 2 seconds of length, while the frequency range must include the frequencies below 3 Hz.