Compressed and Lightweight CNN for Real-Time Parkinson's Tremor Detection from Wearable IMU Data

The real-time detection of Parkinsonian tremors on embedded devices has the potential to enable intelligent assistive technologies that can respond dynamically to tremor events. In this work, we present a lightweight 1D Convolutional Neural Network (CNN) architecture designed to classify tremor patterns from accelerometer data collected via wrist-worn inertial measurement units (IMUs). The proposed CNN is benchmarked against traditional signal processing and classical machine learning methods, demonstrating superior accuracy and robustness. To ensure suitability for deployment on resource-constrained embedded platforms, we apply post-training 8-bit quantization and pruning. These compression techniques reduce model size by over 70% with minimal loss in classification performance. While many existing studies emphasize continuous monitoring and clinical diagnostics, the goal of this work is to lay the groundwork for real-time, on-device tremor detection that can support the development of responsive assistive sys-tems-such as wearable or prosthetic devices capable of mitigating tremor episodes as they occur, enabling real-time intervention through intelligent assistive wearables.