Enhancing preoperative risk stratification: an interpretable machine learning model for ICU admission in spinal...

Spinal metastases frequently cause debilitating symptoms and require complex surgical management, with postoperative intensive care unit (ICU) admissions representing a major concern. This multicenter study aimed to develop and validate machine learning (ML) models to predict 30-day unplanned ICU admission following metastatic spinal tumor surgery.

A total of 642 patients with metastatic spinal disease were enrolled, and 525 from 2 major institutions were randomly split into derivation (80%) and internal validation (20%) cohorts. External validation was performed using an independent cohort ( n = 117) from a third medical institution. Six ML algorithms were trained on 11 clinically significant features selected after multicollinearity analysis.

In the model development cohort, significant differences were observed between ICU ( n = 101, 19.2%) and non-ICU groups, with ICU patients demonstrating higher comorbidity burdens, elevated inflammatory markers, and impaired renal function. Among six machine learning models evaluated, the KNN algorithm demonstrated superior predictive performance with the highest discriminative power [Area Under the Curve (AUC): 0.884], accuracy (82.1%), recall (96.4%), and favorable calibration (Brier score: 0.149). The ANN also performed well, achieving the second-highest AUC (0.847), precision (0.808), and F1 score (0.778), as well as a competitive log loss (0.491). The composite scoring system confirmed KNN and ANN as top performers (total scores: 42 each), but, in the external validation cohort, the KNN model demonstrated significantly superior discriminative ability compared to the ANN model, with an AUC of 0.834 (95% CI: 0.773-0.894) versus 0.741 (95% CI: 0.665-0.816) respectively (Delong test, P < 0.001).

This study presents validated ML models specifically designed for ICU admission prediction following spinal metastasis surgery, with the KNN model performing satisfactory performance and demonstrating strong potential for clinical implementation.