Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Sayma Akther

Second Advisor

Thomas Austin

Third Advisor

Katerina Potika

Keywords

Parkinson’s Disease, Freezing of Gait, Machine Learning, Detection, Random Forest Classifier, ANN, Bidirectional LSTM, Leave-One-Subject-Out, Leave-One-Episode-Out

Abstract

Among the most debilitating symptoms of Parkinson’s Disease is Freezing of Gait (FoG) which is a debilitation in motor abilities such as walking. Severe cases of FoG can lead to an increased risk of falling and thus injury. No cure has been created yet so focus remains on mitigating the severity of the disease. Current research uses accelerometer data and extracted frequency-domain features to detect FoG events, while sidelining patient-specific biographical data. This project seeks to fill this void by creating a machine learning model that incorporates these different features to identify FoG episodes. Multiple models are evaluated including a Random Forest Classifier, an Artificial Neural Network (ANN), and Bidirectional LSTMs (BiLSTMs). Variations of the BiLSTM model are tested with different feature sets to evaluate the performance impact of frequency-domain and biographical data features alongside temporal features. To test for generalization, Leave-One-Subject-Out (LOSO) and Leave-One-Episode-Out (LOEO) evaluation strategies are employed. Our results show that while the biographical data does not help, the frequency-domain features can be useful for both robust and patient-specific implementations making it practical for real-world utilization.

Download

Available for download on Saturday, May 22, 2027

Share

COinS