Predicting Postoperative Discharge Status and Readmissions in Spinal Metastatic Disease Using Machine Learning Models

Area of Science:

• Neurosurgery and Orthopedic Oncology
• Health Services Research
• Machine Learning in Medicine

Background:

• Spinal metastatic disease surgery is primarily for palliation, not cure, with outcomes depending on postoperative recovery.
• Existing survival prediction models lack integration of diverse predictors for short-term recovery outcomes.
• Accurate prediction of non-home discharge (NHD) and 90-day readmission is crucial for managing patients undergoing surgery for spinal metastases.

Purpose of the Study:

• To develop and validate integrated predictive models for short-term postoperative outcomes after surgery for extradural spinal metastatic disease.
• To identify key predictors influencing non-home discharge (NHD) and unplanned 90-day readmission.
• To assess the performance of machine learning algorithms in predicting these outcomes.

Main Methods:

• Retrospective analysis of adult patients undergoing surgery for extradural spinal metastatic disease (2006-2023) using Merative™ MarketScan® data.
• Primary outcomes: non-home discharge (NHD) and unplanned 90-day readmission; secondary outcome: inpatient length of stay (LOS).
• Five machine learning models (Extreme Gradient Boosting, Support Vector Machine, Neural Network, Random Forest, Penalized Logistic Regression) were trained and validated.

Main Results:

• Thoracic spine localization increased NHD odds; postresection arthrodesis and intraoperative neuromonitoring decreased NHD odds.
• Combined anterior-posterior approach and arthrodesis were associated with lower 90-day readmission odds.
• Random Forest model showed best predictive performance for NHD (AUC=0.68) and 90-day readmission (AUC=0.67).

Conclusions:

• Parsimonious predictive models for NHD and 90-day readmission after spinal metastatic surgery were successfully developed and validated.
• These models, particularly the Random Forest algorithm, offer valuable prognostic insights.
• Integration into clinical workflows can enhance decision-making for patient management and prognostication.