Classification of Parkinson Tremor Using Accelerometer Signals and Long Short-Term Memory Network

Tremor is the most characteristic symptom of Parkinson's disease (PD), making its precise differentiation from other tremors crucial for successful diagnosis and subsequently effective treatment. This study presents a methodology for differentiation between patients with Parkinson's disease (PD) and healthy subjects (HS) based on accelerometer (ACC) signals. Information from both upper extremities, acquired simultaneously through a wireless system, is used in order to employ a less complex architecture for classification by means of a variation of a recurrent neural network (RNN), using as input ACC time series signals. An accuracy and precision of 88% were achieved using the “Finger-Nose” maneuver, which allows for better differentiation between groups. Clinical Relevance– The use of automated classification systems represents a potential supportive tool in diagnosis for clinical experts, especially during the early stages of PD, aiming to reduce misdiagnosis. On the other hand, employing deep learning techniques also enables the extraction of information not perceived by humans, which could enhance diagnostic sensitivity, specificity, and efficiency, as well as reduce diagnosis time.