Machine Learning for Integrating Social Determinants in Cardiovascular Disease Prediction Models: A Systematic Review

Background: Cardiovascular disease (CVD) is the number one cause of death worldwide, and CVD burden is increasing in low-resource settings and for lower socioeconomic groups worldwide. Machine learning (ML) algorithms are rapidly being developed and incorporated into clinical practice for CVD prediction and treatment decisions. Significant opportunities for reducing death and disability from cardiovascular disease worldwide lie with addressing the social determinants of cardiovascular outcomes. We sought to review how social determinants of health (SDoH) and variables along their causal pathway are being included in ML algorithms in order to develop best practices for development of future machine learning algorithms that include social determinants. Methods: We conducted a systematic review using five databases (PubMed, Embase, Web of Science, IEEE Xplore and ACM Digital Library). We identified English language articles published from inception to April 10, 2020, which reported on the use of machine learning for cardiovascular disease prediction, that incorporated SDoH and related variables. We included studies that used data from any source or study type. Studies were excluded if they did not include the use of any machine learning algorithm, were developed for non-humans, the outcomes were bio-markers, mediators, surgery or medication of CVD, rehabilitation or mental health outcomes after CVD or cost-effective analysis of CVD, the manuscript was non-English, or was a review or meta-analysis. We also excluded articles presented at conferences as abstracts and the full texts were not obtainable. The study was registered with PROSPERO (CRD42020175466). Findings: Of 2870 articles identified, 96 were eligible for inclusion. Most studies that compared ML and regression showed increased performance of ML, and most studies that compared performance with or without SDoH/related variables showed increased performance with them. The most frequently included SDoH variables were race/ethnicity, income, education and marital status. Studies were largely from North America, Europe and China, limiting the diversity of included populations and variance in social determinants. Interpretation: Findings show that machine learning models, as well as SDoH and related variables, improve CVD prediction model performance. The limited variety of sources and data in studies emphasize that there is opportunity to include more SDoH variables, especially environmental ones, that are known CVD risk factors in machine learning CVD prediction models. Given their flexibility, ML may provide opportunity to incorporate and model the complex nature of social determinants. Such data should be recorded in electronic databases to enable their use.