2987-LB: Fairness of Machine Learning–Based Mortality Prediction in Acute Hyperglycemia Patients across Community-Level Food Access

Résumé fourni par la source

Introduction and Objective: Acute hyperglycemia is common and linked to higher mortality. Community healthy food access may shape this risk and how discharge risk-prediction tools classify patients. The objective of this study was to evaluate the fairness of a machine learning (ML) model for predicting long-term mortality among hospitalized patients with acute hyperglycemia, across strata of community-level food access. Methods: A retrospective cohort of 4,627,091 U.S. hospital patients without baseline diabetes from the National Clinical Cohort Collaborative was used to evaluate whether a Gradient Boosting Machine was fair by community-level food access in predicting all-cause mortality >30 days after discharge (median follow-up: 3.3 years). Acute hyperglycemia was defined as two or more glucose readings ≥180 mg/dL during hospitalization. The primary outcome was all-cause mortality >30 days post-discharge. Community-level food access was categorized using the modified Retail Food Environment Index: (1) No Healthy Access (0-5), (2) Limited Healthy Access (5.1-10), (3) Enhanced Healthy Access (10.1-37.5), and (4) Balanced Healthy Access (37.6-100). Discrimination (AUC), calibration, and group fairness metrics were compared across food access groups. Results: The ML model’s sensitivity and specificity were similar across food access strata, with true positive rates ranging of 68%-72% and false positive rates of 7.5%-10.4%, suggesting modest accuracy loss in disadvantaged communities. Patients from “No Healthy Access” areas were more often flagged as high risk (19.1%, 95% CI: 18.5%-19.6%) than those from “Enhanced Healthy Access” areas (15.0%, 95% CI: 14.4%-15.6%). The resulting disparate impact ratio was 0.78 (95% CI: 0.76-0.80), falling below the 0.80 threshold and signaling a significant disparity. Conclusion: Sensitivity and specificity were comparable across strata, but high-risk flagging was higher in the stratum with lowest healthy food access. Disclosure M.P. Santos: None. A. Culotta: None. K. Theall: None. S.H. Ley: None. Funding National Institutes of Health (T32292732)

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.