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

Area of Science:

• Health Economics
• Medical Informatics
• Spine Surgery

Background:

• Value-based care models are transforming healthcare reimbursement.
• Spine surgery's complexity poses challenges for standardized bundled payment models.
• Accurate financial prediction is essential for sustainable healthcare delivery.

Purpose of the Study:

• To develop and validate a machine learning model for predicting financial outliers in spine surgery.
• To integrate structured clinical data and unstructured surgeon notes for enhanced prediction accuracy.
• To inform the development of risk-stratified, patient-specific payment plans.

Main Methods:

• Developed a multimodal machine learning model using structured clinical data and natural language processing (NLP) of surgeon notes.
• Utilized a dataset of 1898 spine surgery patients.
• Validated model performance using ROC-AUC for predicting total and variable cost outliers.

Main Results:

• The model achieved high predictive performance (ROC-AUC 0.845 for total cost, 0.883 for variable cost).
• Identified 11.0% of patients as financial outliers, contributing to significant financial losses.
• Financial outliers demonstrated increased ICU admissions, 90-day reoperations, and longer length of stay (LOS).

Conclusions:

• A risk-stratified machine learning model can accurately predict financial outliers in spine surgery.
• Individualized, risk-based payment models are necessary for equitable reimbursement in spine surgery.
• Implementing such models can improve payment equity and align resource allocation with patient complexity.