T2D is a public health crisis that must be managed by addressing patients’ unmet social needs. We aimed to develop an electronic health record (EHR)-based individualized polysocial risk score (iPsRS) of T2D, which can be used for identifying patients of high social risk and their unmet social needs. We identified T2D patients using 2015-2020 EHR data from University of Florida Health system and followed a year to record hospitalizations. We (1) used natural language processing to extract person-level social determinants of health (SDoH, e.g., employment, housing and food) from clinical notes; and (2) spatiotemporally linked 117 contextual-level SDoH (e.g., neighborhood walkability). We incorporated both contextual and person-level SDoH and used machine learning(ML) to develop an iPsRS that predicted 1-year hospitalization. We assessed iPsRS’ fairness across racial-ethnic groups (Black or Hispanic vs White) and explored different approaches to mitigate identified bias. Of 10192 T2D patients, mean age was 59±21years, and 16% were hospitalized in a year. Our iPsRS showed a good utility for predicting individuals’ risk of hospitalization (C statistic=0.70). After adjusting for clinical characteristics, iPsRS explained 41% of the increased 1-yr hospitalization risk. The naïve iPsRS was biased on Hispanic vs. White patients, that is, Hispanics (false negative rates [FNR]=0.87) were more likely to be misclassified as having low hospitalization risk compared to their White counterparts (FNR=0.77). We identified that a disparate impact remover approach reduced the bias without compromising much prediction utility, producing a final iPsRS with an optimized prediction (C-statistic=0.69) - fairness (FNR-Hispanic=0.62, FNR-White=0.64) balance. We developed a fair iPsRS that can identify T2D patients at a high risk of hospitalization due to SDoH, supporting the integration of social care with clinical care of T2D.


Y.Huang: None. T.Zhang: None. W.T.Donahoo: None. E.Shenkman: None. J.Guo: Consultant; Pfizer Inc., Research Support; PhRMA Foundation, NIH - National Institutes of Health. J.Bian: None.


National Institute of Diabetes and Digestive and Kidney Diseases (R01DK133465)

Readers may use this article as long as the work is properly cited, the use is educational and not for profit, and the work is not altered. More information is available at http://www.diabetesjournals.org/content/license.