Interpretable AI for Decision Support for Mobility Changes of Older Adults

Monitoring the physical performance of older adults is important for their health and well-being. While conducting frequent standardised geriatrics assessments exceeds the capacities of physicians, sensor systems in domestic environments can continuously monitor the performance of older adults in their daily life. Our system uses motion sensors, a sensor graph, a groundplan, and an ensemble classifier to predict Timed Up & Go scores on daily scale. The ensemble classifier, comprised of three Random Forests, used 12 hand-engineered features and achieved an accuracy of 91.06%(+/- 4.91), 89.82%(+/- 1.93), and 90.53%(+/- 3.19) for the scores 1, 2, and 3, respectively; the complete ensemble had an accuracy of 79.17% on our real-world dataset containing data of 20 (pre-)frail older adults aged >= 75y. Since interpretability and reasoning are important in medicine, Shapley values were used for analysing the feature importance. Albeit the Timed & Go measures only a single feature, the Shapley values revealed that the classifiers were sensitive to a combination of features. Also, we show how the proposed system and the Shapley values support the decision making of medical experts.