Developing a Hybrid Machine Learning System for Parkinson's Disease Detection Using a Combination of Freezing of Gait Video and Inertial Sensor Data

Detection of Parkinson's disease remains challenging due to the complexity and cost of diagnosis. Recently, different machine learning models have been proposed to detect Parkinson's. This paper proposes a novel hybrid machine learning detection system to detect Parkinson's disease using a combination of both videos of freezing of gait and numerical sensor data. Three machine learning models (sensor model, video model, and a hybrid model using a combination of the data) have been developed and evaluated. Four machine learning algorithms were used for model development, which were Logistic Regression, K-neighbors, Bayesian Regression, and Random Forest. All the models were evaluated using standard performance metrics of accuracy and precision based on the prediction of 12 commonly used Parkinson's rating scales (Mini-Mental, NFo-GQ, H&Y, UPDRS-II, UPDRS-III, PIGD, Dyskinesia, HADS, HADS-A, HADS-D, FES-I, mini-BESTest). This is the first time to use all 12 rating scales as output of a machine learning model for Parkinson's disease detection. The results indicate that the hybrid model has a significantly better performance than the video or sensor models. The hybrid model achieved an average accuracy of approximately 94% for all twelve rating scales, while the sensor model had an accuracy of approximately 91% and the video model had an accuracy of approximately 89%. This research indicates that the hybrid model is accurate and reliable because of its ability to fully use two sets of clinical data to make a final decision.