Mobile health for cardiovascular disease risk prediction and management in resource-constrained environments

It is well established that the leading global cause of mortality and morbidity, cardiovascular disease (CVD), is more severe in resource-constrained environments such as rural India (RI). This thesis explores how best to manage CVD risk in RI by using a mobile-based, point-of- care (POC) Clinical Decision Support System (CDSS), SMARThealth, that is designed to assist Accredited Social Health Activists (ASHAs) or minimally trained health workers. The four major focus areas are: (a) Design, development, and large-scale data collection using SMARThealth - an agile development process and user-centred design approach were followed to pilot test the CDSS with 292 participants. Evaluation metrics included system efficiency, end-user variability, usability, and sub-group analysis to identify better or poorly performing ASHAs. An improved version of SMARThealth was used for baseline data collection across 54 villages (62,194 participants) in Andhra Pradesh, India. 9864 (15.8%) of the participants were at high CVD risk. (b) Improvement of the sole CVD risk prediction algorithm for RI, the WHO/ISH CVD risk prediction charts (WHO-ISHc) - the choice of the low information (LI) model or high information (HI) model of WHO-ISHc was statistically significant for CVD risk prediction in RI (p=0.008;χ2=7.03) with 155 subjects (or 14.5% of 1066 patients) having different CVD risk scores according to the LI and HI WHO-ISHc. A parsimonious POC test was developed to identify patients for whom risk prediction by the HI and LI WHO/ISHc differ (that is, for whom the assessment of total cholesterol would be beneficial). The POC test showed good discrimination (out-of-sample AUC 0.85 with Random Forests). (c) Assessment of best prediction algorithm for RI - eight highly predictive features of CVD risk were identified based on labelled data, and the resulting model (Model 1) had higher or equal AUCs and log-likelihood scores, and lower Brier scores when compared to a benchmark algorithm. The contribution of age and gender alone offered good discrimination and recalibration of Model 1 for RI was introduced. The lack of recorded end outcomes in RI prompted the use of an unsupervised approach to identify high-risk patients. Clusters of low and high CVD risks were found when ˆK =2, but also clusters with intermediate risk when ˆK =4 offering an alternative approach to identifying groups of high-risk patients. (d) Analysis from a randomised controlled trial evaluation of SMARThealth - preliminary data analysis of 131 high-risk patients during the first year of the randomised controlled trial showed a statistically significant reduction in median blood pressure between the 1st and 5th assessment (p=0.0097). The proportion of patients under treatment for high blood pressure continued to increase throughout.