Machine learning model to estimate momentary status of activities of daily living in people with early-stage Parkinson’s disease: an experimental study in home setting

PURPOSE: The purpose of this study was to develop a model that estimates the momentary status of activities of daily living (ADLs) in people with Parkinson's disease (PwPD) in the early stages using data measured in real home settings of PwPD. MATERIALS AND METHODS: We conducted an experiment to measure kinematic data as PwPD performed ADLs in their home environment. We selected five instrumental ADLs (hanging out laundry, taking in laundry, going out, preparing a meal, and cleaning a room) and two basic ADLs (toileting and bathing). Six PwPD in early stages participated in the experiment, and we collected joint movement data and segment position data when they performed ADLs (number of instances = 3954). Then, we built random forest models with 10-fold cross validation and evaluated them in remained 20% data. RESULTS: The results demonstrated that two models developed in this study successfully estimates the momentary status of ADLs (an accuracy of 0.82 and a kappa coefficient of 0.83 with joint motion data, and an accuracy of 0.84 and a kappa coefficient of 0.87 with segment position data). CONCLUSION: This study demonstrated combination of kinematic data and machine learning algorithms enables estimation of the momentary status of ADLs and represents a proof of concept demonstrating that ADL monitoring system with kinematic data measurements and machine learning. Further research for practical implementation is warranted.