Machine Learning for Causal Inference in Hospital Diabetes Care: TMLE Analysis of Selection Bias in Diabetic Foot Infection Treatment—A Cautionary Tale

Background/Objectives: Diabetic foot infections (DFIs) are a leading cause of hospitalization, amputation, and costs among patients with diabetes. Although early treatment is assumed to reduce complications, its real-world effects remain uncertain. We applied a causal machine-learning (ML) approach to investigate whether early DFI treatment improves hospitalization and clinical outcomes. Methods: We conducted an observational study using de-identified UCSF electronic health record (EHR) data from 1434 adults with DFI (2015–2024). Early treatment (<3 days after diagnosis) was compared to delayed/no treatment (≥3 days or none). Outcomes included DFI-related hospitalization and lower-extremity amputation (LEA). Confounders included demographics, comorbidities, antidiabetic medication use, and laboratory values. We applied Targeted Maximum Likelihood Estimation (TMLE) with SuperLearner, a machine-learning ensemble. Results: Early treatment was associated with higher hospitalization risk (TMLE risk difference [RD]: 0.293; 95% CI: 0.220–0.367), reflecting the triage of clinically sicker patients. In contrast, early treatment showed a protective trend against amputation (TMLE RD: −0.040; 95% CI: −0.098 to 0.066). Results were consistent across estimation methods and robust to bootstrap validation. A major limitation is that many patients likely received treatment outside UCSF, introducing uncertainty around exposure classification. Conclusions: Early treatment of DFIs increased hospitalization but reduced amputation risk, a paradox reflecting appropriate clinical triage and systematic exposure misclassification from fragmented healthcare records. Providers prioritize the sickest patients for early intervention, leading to greater short-term utilization but potentially preventing irreversible complications. These findings highlight a cautionary tale; even with causal ML, single-institution analyses may misrepresent treatment effects, underscoring the need for causally informed decision support and unified EHR data.