Prediction of Freezing of Gait in Parkinson’s Disease Patients Using Machine Learning

Parkinson’s disease is a complex neurological dis- order that severely impacts a person’s motor control, leading to various movement difficulties. Among the most challenging symptoms being experienced is the freezing of gait, where individuals suddenly and temporarily struggle to start or main- tain movement, causing them to freeze in place. Detecting and predicting these FOG episodes early is crucial for prompt intervention and improved patient results. This project introduces a machine learning method for predicting FOG episodes using 3D accelerometer data gathered from the lower backs of Parkinson’s patients in both lab and home environments. The research uses two datasets: tDCSFOG, with 833 unique participants, and DeFOG, with 137 unique participants. The goal is to find the most accurate model for predicting FOG and to estimate the probability of occurrence of four specific events : Start Hesitation, Turn, Walking, and Normal movement, and to identify which event is most and least likely to occur among the subjects. Five machine learning models were used : LightGBM, Random Forest, Decision Tree, Gradient Boosting Classifier, and CatBoost. These models were assessed according to their accuracy, recall, precision, and f1-score. LightGBM performed the best, with an accuracy rate of 98.11%, a f1-score of 98.10%, a precision rate of 98.12%, and a recall rate of 98.11%. Therefore, LightGBM was chosen for the final probability prediction.The results show that ‘Turn’ is the most frequently occurring event, while ‘Walking’ is the least frequently occurring event among the subjects. The aim of this research is to improve the prediction and understanding the instances of FOG in patients, thereby contributing to better management and care for those affected by this challenging symptom.