Abstract
Tremors represent an early and primary symptom in diagnosing Parkinson’s disease (PD). Monitoring tremor’s frequency and intensity by wearable device sensors helps healthcare professionals understand patients’ conditions and enables personalized interventions. Due to privacy concerns, traditional centralized machine learning (ML) methods are unsuitable for analyzing sensitive data collected from wearable devices. Federated learning (FL) overcomes this issue, preserving data security and privacy by training ML models collaboratively without sharing raw data. This paper presents a comparative study of proposed CNN-LSTM and InceptionTime models trained with accelerometer data for PD tremor detection in centralized and federated environments. Results obtained from 27 PD patients’ wearable data show that the CNN-LSTM performs comparable to the InceptionTime model. Still, we highlight the lightweight of the InceptionTime model, a characteristic that makes it suitable to be executed in wearable devices.