Disaggregating health differences and disparities with machine learning and observed-to-expected ratios: Application to major lower limb amputation.

Background: Major lower limb amputation is a devastating but preventable complication of peripheral artery disease. It is unclear whether racial and ethnic and rural differences in amputation rates are due to clinical, hospital, or structural factors.

Methods: We included all peripheral artery disease hospitalizations of patients ≥40 years old between 2017-2019 in Florida, Georgia, Maryland, Mississippi, or New York (HCUP State Inpatient Databases). We estimated the expected number of amputations using three models: 1) unadjusted, 2) adjusted for clinical factors, and 3) adjusted for clinical factors, hospital factors, and social determinants of health) using LASSO. We calculated and compared observed-to-expected ratios and quantified the role of these factors in amputation rates.

Results: Overall, 1,577,061 hospitalizations (990,152 unique patients) and 21,233 major lower limb amputations (1.4%) were included. After accounting for clinical differences, we observed amputation disparities among rural Black, Hispanic, Native American, and White patients and non-rural Black and Native American patients. After accounting for hospital factors and social determinants of health, disparities were no longer present among rural White adults (0.93, 95% CI 0.77-1.09); however, disparities persisted among rural Black (1.26, 95% CI 1.01-1.51), Hispanic (1.50, 95% CI 0.89-2.12), and Native American patients (1.13, 95% CI 0.68-1.58) and non-rural Black (1.12, 95% CI 1.09-1.15) and Native American (1.15, 95% CI 0.86-1.44) patients.

Conclusions: Clinical factors did not fully explain differences in amputation rates, and hospital factors and social determinants did not fully explain disparities. These findings provide additional evidence that implicit bias is associated with amputation disparities.