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

Area of Science:

• Oncology
• Neurosurgery
• Data Science

Background:

• Spinal metastases significantly impact patient quality of life and functional outcomes.
• Accurate prediction of postoperative performance status (PS) is crucial for surgical planning and patient counseling.

Purpose of the Study:

• To develop and validate prediction models for one-month postoperative performance status (PS) in patients undergoing surgery for spinal metastases.
• To identify patients likely to achieve a favorable PS (0-2) one month after surgery.

Main Methods:

• Retrospective analysis of a prospectively collected spine surgery registry.
• Comparison of three tree-based models (Random Forest, XGBoost, CatBoost) and two regularized logistic regression models.
• Nested cross-validation for model development and hyperparameter tuning, with specific strategies for handling missing data.

Main Results:

• Random Forest demonstrated the highest discrimination (AUC-ROC 0.811 ± 0.079) and superior calibration.
• Sparse elastic-net logistic regression offered good discrimination with a limited predictor set.
• Model performance remained consistent in sensitivity analyses excluding predictors with high missingness.

Conclusions:

• Tree-based models, especially Random Forest, show the most promising predictive performance for early postoperative functional status after spinal metastasis surgery.
• Sparse elastic-net models provide interpretability with a reduced set of predictors.
• Clinical implementation requires careful assessment of model calibration alongside discrimination.