Abstract
To develop applications for assisting Parkinson's disease (PD) patients, extracting Parkinsonian tremors from the raw signal is crucial; however, conventional methods such as filtering require a preset frequency range, and a poorly set frequency range may lead to the inclusion of undesired signals. AI algorithms can potentially overcome the heterogenous tremor characteristics in different patients, but they have been applied in the disease or tremor type classification rather than in tremulous-voluntary motion classification. Hence, this study presents an approach to automatically differentiate between voluntary and tremulous motions in PD patients, achieved by combining ensemble empirical mode decomposition (EEMD) and convolutional bi-directional long short-term memory (LSTM). Non-labelled raw hand-arm orientation data collected from PD patients was decomposed into sub-signals via EEMD to replace the conventional filtering techniques. A convolutional layer automatically extracted key features from these sub-signals to train the deep learning classifier, eliminating the need for manual feature engineering. The proposed method can be generalized to identify tremors from motion signals obtained during daily activities without pre-identified features. The proposed approach accurately extracted Parkinsonian tremors from raw signals across various activities, achieving low root-mean-square error and 94.2 ± 1.1% accuracy in differentiating voluntary and tremulous motions in PD patients.