Modeling Social Determinants of Health through Human-Centered Artificial Intelligence

Many Americans have social and cultural hurdles that prevent them from getting timely cancer diagnoses and effective treatments. For example, lack of transportation can prevent women from being regularly screened for breast cancer, increasing the risk of a late cancer diagnosis that can be difficult to treat. Such social factors, also called social determinants of health, are well-known but have been difficult to analyze and interpret despite the use of powerful machine learning methods based on three hurdles: (1) patients are similar and different in complex ways based on their social determinants of health. For example, some patients lack transportation, health insurance, and steady employment, whereas others partially overlap with this group by also lacking steady employment, but additionally having language and communication problems with their providers. We need methods that can not only identify such complex overlapping patterns to help design targeted solutions, but also measure their statistical significance and replicability, which are critical in the biomedical sciences; (2) many Al and machine learning methods use complex mathematical formulas that transform the data in ways that are difficult to interpret by clinicians - the so called "black box" Al problem. This problem prevents researchers from inspecting whether the data and algorithms are biased against some groups. For example, if the data only includes individuals with health records, the results are biased because they exclude individuals who have never seen a doctor due to low income; and (3) Black and Hispanic researchers are currently underrepresented in Al and machine learning research, which further increases the risk of biased data, analyses, and their interpretations.

 

This project addresses these three hurdles by using an approach called Human-Centered Al. This approach will use graphical networks to automatically identify complex patterns in very large datasets, while also visualizing the results at each step so ethicists, biostatisticians, and clinicians can inspect the data and interpret the results. In addition, the collaboration with Dr. Hunter from Texas Southern University, a historically black college or university (HBCU), will leverage his rich experience in treating cancer patients from disadvantaged communities to inspect the analysis and interpret the

results.