A Picture is Worth a Thousand Steps
Using Image Processing Techniques to Predict Freezing of Gait in Parkinson's Patients

Abstract

Parkinson's Disease is a neurodegenerative disease that affects the ability to perform activities of daily living. A debilitating syndrome of Parkinson's is Freezing of Gait (FoG), where patients are unable to move forward despite intention of walking, resulting in the forward momentum shifting to the torso and leading patients to fall, causing serious medical consequences. Prior work has explored that use of gait tests, medical questionnaires and inertial measurement units to predict FoG events. In this research, a comprehensive review of various input formats, signal processing algorithsm and machine learning algorithms to predict FoG has been performed. Raw signal data and the Moore-Bachlin algorithm have been studied, and we introduce a novel method for image representation via autoscaling and RGB pixelation. We find that a 2-dimensional convolutional neural network (CNN) performs the best on scaled images, attaining a state-of-the-art accuracy 99.50% and sensitivity of 99.65%, that has been verified via various K-Folds and overfitting tests. We also integrate this model into an Android application which can be used by patients and doctors to predict and track freeze events over a period of time. This can help doctors diagnose patients and gauge the severity of the disease to prescribe medication accordingly.