Machine Learning Discrimination of Parkinson’s Disease Stages from Walker-Mounted Sensors Data

Clinical methods that assess Parkinson’s Disease (PD) gait are mostly qualitative. Quantitative methods necessitate costly instrumentation or cumbersome wearables, limiting usability. This study applies machine learning to discriminate six stages of PD. The data was acquired by low cost walker-mounted sensors at a movement disorders clinic. A large set of features were extracted and three feature selection methods were compared using a Random Forest classifier. The feature subset selected by the ANOVA method provided performance similar to the full feature set: 93% accuracy, with a significantly shorter computation time. Compared to PCA, it enabled clinical interpretability of the features, an essential attribute in healthcare. All selected-feature sets were dominated by information theoretic and statistical features and offer insights into the characteristics of PD gait deterioration. The results indicate a feasibility of machine learning to accurately classify PD severity stages from kinematic signals acquired by low-cost, walker-mounted sensors and can aid medical practitioners in quantitative assessment of PD progression. The study presents a solution to the small and noisy data problem, typical to sensor-based healthcare assessments.