Leveraging automated machine learning to benchmark, deconstruct, and compare frailty indices for predicting adverse spinal surgery outcomes.

Frailty indices, composite summary scores of diagnostic codes, labs, or vitals, are widely used to assess patient vulnerability. However, their value in predictive modeling remains uncertain. In a cohort of 2,912 elective lumbar spinal fusions, we retrospectively evaluated two established indices, the Hospital Frailty Risk Score (HFRS) and the electronic Frailty Index (eFI), against a curated set of 63 clinical features from electronic medical records. Using automated machine learning optimized logistic regression models, we compared predictive performances and feature importances of frailty indices, and their individual components, against curated clinical features. The curated clinical feature set consistently outperformed (measured by AUC) both frailty indices and their respective components across three outcomes: discharge disposition, length of stay, and 90-day readmission. While individual frailty components improved performance over their respective summary scores, they underperformed compared to clinical variables. However, integrating select frailty-based components with curated clinical features modestly improved predictive performance. Interaction analysis uncovered nonlinear relationships, such as between hemoglobin and smoking, that frailty indices alone did not capture. These findings highlight that while frailty indices aid broad risk stratification, their composite nature limits utility for precise, patient-specific predictive modeling, supporting hybrid strategies centered on curated variables enhanced with selected frailty elements.