Personal health monitoring refers to the continuous, long-term monitoring of an individual’s health by integrating and analysing diverse sources of health data. This can facilitate the early detection, prediction, and prevention of diseases. The aim of this research is to design and evaluate an AI system architecture for personal health monitoring. The architecture will be a guide for the development of personal health monitoring systems that combine expert knowledge with data-driven analyses. It integrates three AI techniques: knowledge graphs (KGs), machine learning (ML), and Bayesian networks (BNs).
I demonstrate the architecture through a use case of atrial fibrillation (AF), the most commonly occurring heart rhythm disorder worldwide. The specific objectives for this research are: 1.) to predict the risk of new-onset AF in the general population; 2.) to predict the onset of AF episodes in AF patients with lead times ranging from 5 minutes to 1 hour; 3.) to predict the risk of stroke and major bleeding in AF patients; and 4.) to provide explainable and human-centered decision support to mitigate the predicted situations.
I will use the following data from the UK Biobank: healthcare records, self-reported/questionnaire lifestyle data, activity monitoring, and electrocardiogram (ECG) data, to develop and validate the models. Specifically, I will: 1.) train ML models to identify patterns preceding AF episode onset using ECG data; 2.) build KGs to represent an individual’s profile, including health conditions, body attributes, and activity; 3.) develop BNs to calculate personalised risk scores for AF, stroke, and major bleeding based on the individual’s profile; and 4.) validate the predictive performance of the models. This builds on my prior work where I built models using MIMIC-IV data, a dataset consisting primarily of intensive care unit patients. UK Biobank data will demonstrate the efficacy of the models for more general populations.