Tree-based and sparse logistic models for predicting one-month postoperative performance status after surgery for spinal metastases.

Abstract

We aimed to develop and internally validate prediction models for one-month postoperative performance status (PS) after surgery for spinal metastases and to identify patients likely to achieve PS 0-2 at one month. We performed a retrospective analysis of a prospectively collected spine surgery registry. We compared three tree-based models (Random Forest, XGBoost, and CatBoost) with two regularized logistic regression models (ridge-regularized logistic regression and a sparse elastic-net logistic regression model constrained to ≤ 15 predictors). Model development and hyperparameter tuning were performed using nested cross-validation. Missing data were handled using model-specific strategies within the cross-validation pipeline, and a sensitivity analysis excluded the predictor with the highest missingness. Performance was assessed using discrimination and calibration metrics, including the area under the receiver operating characteristic curve (AUC-ROC), accuracy, precision, recall, F1 score, Brier score, calibration intercept, and calibration slope. The primary analysis included 375 patients with available one-month PS out of 413 enrolled patients. Random Forest achieved the highest discrimination (AUC-ROC 0.811 ± 0.079) and showed calibration measures closest to the ideal among the evaluated models (Brier score 0.168; calibration intercept - 0.024; slope 1.121). The sparse elastic-net model showed good discrimination (AUC-ROC 0.796 ± 0.081) with a limited set of predictors, although its calibration metrics suggested less reliable absolute probability estimates (Brier score 0.217; intercept 0.612; slope 3.228). Excluding the predictor with the highest missingness yielded similar performance for the main models. Tree-based models, particularly Random Forest, provided the most favorable overall predictive performance for one-month postoperative PS after surgery for spinal metastases, whereas a sparse elastic-net logistic regression model preserved reasonable discrimination with a small predictor set and coefficient-based interpretability. These findings support clinically oriented prediction of early postoperative functional status while highlighting the need to assess calibration before clinical implementation.