Multimodal machine learning for risk-stratified bundled payments in spinal surgery.

Accurate prediction of financial metrics in spine surgery is crucial as healthcare transitions to value-based care. While bundled payment models have succeeded in other orthopedic procedures, the heterogeneity of spinal surgery complicates their adoption. We develop the first preoperative risk-stratified multimodal machine learning model that integrates structured clinical data and unstructured surgeon notes using natural language processing to predict financial parameters. The model achieved ROC-AUC values of 0.845 and 0.883 for outlier total and variable costs, respectively, reflecting good-to-excellent performance. Among 1898 spinal surgery patients, 209 (11.0%) were identified as financial outliers, contributing to $12.8 million in losses, while the remaining cases yielded $1.8 million in profits. Financial outliers exhibited higher ICU admissions, 90-day reoperations, and longer LOS (all P < 0.001). We propose a patient-specific payment plan by quantifying predicted risk, enabling fair payment adjustments for high-risk spinal surgery patients. Institutions treating higher-risk patients face a greater financial burden in a flat bundled payment structure, emphasizing the need for individualized, risk-based models to improve payment equity and align resource allocation with patient complexity.