Concurrent clustering and classification for assessing the risk of falling during ageing

During ageing, fall prevention represents one of the most sustainable plan of action to promote active ageing and reduce health care costs. To significantly prevent the falls, it is necessary to deploy predictive models that can accurately assess the risk of falling. However, current machine learning methodologies do not offer insights about the rules used for the assessment. Here, we proposed a method capable to concurrently cluster and classify data in the context of fall risk assessment. Such clustering provides support in analyzing the classification performed. We applied the method on a dataset composed by accelerometer signals collected using a wearable sensor from 90 subjects that underwent a Tinetti test (i.e., a clinical scale meant to assess the risk of falling). Thirty-three subjects had a Tinetti score <= 18 and considered has having high risk of falling. A training-validation-test procedure was designed to determine the classification accuracy of the proposed methodology. We evaluated the automatic clustering by observing how the subjects were splitted into three groups. The method achieved a test set accuracy of 0.85. The obtained clusters supported the presence of three macro groups, i.e., low risk, high risk and borderline.