Keywords
Parkinson's Disease, Machine Learning

Technologies
Python, scikit-learn, pandas, NumPy, Matplotlib

Parkinson's disease affects over 8.5 million people worldwide, and its severity is typically assessed through clinical rating scales that depend on trained specialists and subjective judgement. This project investigates whether wearable sensors can offer an objective, scalable alternative.

Using data from 20 participants collected by the Tyndall National Institute, the project extracts movement features from just two body-worn sensors, one on the wrist and one on the ankle, during the Timed Up and Go clinical mobility test. Rather than relying on a single stopwatch time, the TUG is segmented into five phases using video annotation, and 238 features are engineered to capture phase-specific movement characteristics such as gait variability, turning dynamics, and transition smoothness.

Machine learning regression models are trained to predict clinical severity scores from these features, with Principal Component Analysis applied to address the high-dimensional, small-sample challenge. Results demonstrate that a minimal sensor configuration combined with phase-aware feature engineering can capture movement patterns meaningfully associated with Parkinson's disease severity.