Clinical profile prediction by multiple instance learning from multi-sensorial data

The last years there is a great interest in developing unobtrusive health monitoring systems with a predictive component, aiming to recognize signs of illness in an attempt to assist clinicians in delivering early interventions. The objective of this work is to investigate whether the physiological and kinetic functioning and human activity of daily living monitored by multiple sensors can be used as surrogate of the standard clinical assessment. We focus on the older population and propose to utilize Multiple Instance Learning (MIL) to predict their clinical profile from the multi-sensorial data. ReliefF-MI is applied to achieve dimensionality reduction and to discover the most important features that are associated with each clinical metric, while the BagSMOTE algorithm is utilized to mitigate the class imbalance problem. The proposed methodology was evaluated on a multi-parametric dataset of 86 older adults containing clinical parameters from various domains (cognitive, physical, medical, psychological, social and showed high prognostic capacity for the person’s functionality (Katz index) and social interaction (phone calls).