APPLICATIONS OF MACHINE LEARNING FOR PERSONALIZED HEALTHCARE OF OLDER ADULTS LEVERAGING IN-HOME SENSORS

A growing population of baby boomers and their demands for independent aging in place settings necessitates in-home sensing frameworks equipped to make the aging experience more affordable and comfortable. Aging-in-place facilities like TigerPlace are an example of such a community. With in-home sensing frameworks, a daily deluge of health data is expected to be gathered.  Therefore, automated health monitoring and early illness predictions are critical needs for this bulk of data. An apt solution is integrating applications of artificial intelligence and machine learning (AI/ML) for automated monitoring and predictions related to health, which can serve as indications for clinicians to act upon early. The Center to Stream Healthcare in Place (C2SHIP) (previously known as the Center for Eldercare and Rehabilitation Technology (CERT)) is a research organization at the University of Missouri that works towards integrating AI/ML for automated health monitoring purposes and health decline predictions.

This current research is an exemplar of applying AI/ML for personalized healthcare of older adults with minimal human intervention, detecting the early stages of health decline, and enabling interventions before a health condition becomes medically evident.  Linguistic health summary data of older adults from aging-in-place facilities is used to find frequent repeating patterns in an individual's health and associate it with a health condition. Association rules are integrated with transactional encodings to discover patterns.  Further, electronic health records (EHR) data are scraped and re-structured to associate conditions with these repeating patterns.  The increased frequency of similar patterns can be a sign of a health condition becoming chronic, and this applicability and utility are demonstrated through various case studies. To make explorations at the deeper level, a transition is made here to an in-depth study of sensor data from the aforementioned aging-in-place facilities. Specifically, the focus is on the hydraulic bed sensor (HBS) which helps in capturing the respiratory and heart function in the form of composite signals.  Past HBS data gathered from Chronic Obstructive Pulmonary Disorder (COPD) diagnosed individuals, pre and post-COPD over 36 months is used.  Robust and explainable feature engineering, coupled with data mining, is used to find changing patterns in this data over time and effectively learn to code differentiable types of breathing patterns (normal/abnormal), in essence, a code book. This code book is then used to build a predictive model that makes predictions for other unseen older adults in the future, in an automated way.  The predictive model demonstrates ~99% accuracy.  Further, the clinical relevance of the results is testified by building retrospective case studies for six older adults and comparing prediction results with the health checkpoints mentioned in the EHR of the respective individuals.