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

Operative management of spinal metastatic disease is largely for symptom palliation rather than curative and revolves around the expectation that postoperative survival will exceed recovery time. While several scoring systems and models to predict survival exist, few studies have unified diverse predictors into integrated models to predict short-term postoperative outcomes as indicators of recovery.The Merative™ MarketScan® Commercial Database and the accompanying Medicare Supplement were queried for adult patients receiving surgery for extradural spinal metastatic disease between 2006 and 2023. Primary outcomes of interest were non-home discharge (NHD) and unplanned 90-day postdischarge readmission. Inpatient length of stay (LOS) was assessed as a secondary outcome. Five models (Extreme Gradient Boosting, Support Vector Machine, Neural Network, Random Forest, and Penalized Logistic Regression) were trained on a 70% training sample and validated on the withheld 30%.A total of 1,926 patients were included. Thoracic spine localization (vs. cervical, odds ratio [OR]: 2.83, 95% confidence interval [CI]: [1.74-4.58]) was associated with higher odds, while postresection arthrodesis (vs. no arthrodesis, OR = 1.24, [0.59-0.97]) and intraoperative neuromonitoring (vs. not, OR = 0.45, [0.31-0.66]) were associated with lower odds, of NHD. Utilizing a combined anterior and posterior approach (vs. anterior, OR = 0.50, [0.33-0.75]) and arthrodesis (OR = 0.96, [0.75-1.23]) were associated with lower odds of 90-day readmission. Similarly, using intraoperative neuromonitoring (B = - 1.84, [-2.72, -0.97]) or operating microscope (vs. not, B = - 1.71, [-2.66, -0.76]), postresection arthrodesis (B = - 0.17 [-2.66, -0.76]) were associated with shorter LOS, while thoracic localization (B = 1.67, [0.57, 2.76]) was associated with extended LOS. The random forest algorithm demonstrated the best overall predictive performance in the withheld validation cohort when assessing NHD (area under the curve [AUC] = 0.68, calibration slope = 0.82) and unplanned 90-day readmission (AUC = 0.67, calibration slope = 0.87).We developed and validated parsimonious predictive models to estimate the risk of NHD and 90-day readmission after surgery for extradural spinal metastatic disease. After integration into physician- and patient-facing interfaces, these models may serve as clinically useful decision tools to enhance prognostication and management.