Development and temporal validation of a machine learning–based cancer-specific early warning score for predicting clinical deterioration in oncology ICU patients.

24

Background: Cancer patients admitted to the intensive care unit (ICU) are at high risk of clinical deterioration, yet conventional early warning scores (EWS) such as MEWS and NEWS were developed for general populations and lack cancer-specific parameters. These scores show limited discrimination in oncology patients, who often present with altered physiological baselines due to disease burden, immunosuppression, and treatment toxicity. We developed and temporally validated a machine learning-based Oncology Early Warning Score (Onc-EWS) incorporating cancer-specific biomarkers to predict clinical deterioration in critically ill cancer patients. Methods: Using MIMIC-IV (v3.1), we identified 5,729 adult ICU admissions with active cancer diagnoses across nine malignancy types (2008–2022). Clinical deterioration was defined as a composite of ICU mortality, new mechanical ventilation, new vasopressor initiation, or cardiac arrest. We extracted 90 predictor features from the first 24 hours of ICU stay spanning vital signs, standard and cancer-specific laboratory values (albumin, LDH, calcium, lactate), and derived indices (shock index, NLR). An XGBoost classifier was trained on 2008–2016 data (n=4,499) and temporally validated on 2017–2022 data (n=1,230). Onc-EWS was compared against MEWS, NEWS, and SOFA using AUROC. SHAP analysis provided model interpretability. Results: The cohort had a mean age of 66.3 years (44.3% female), mean SOFA of 9.8, and 40.7% deterioration rate. On temporal validation, Onc-EWS achieved an AUROC of 0.901, significantly outperforming SOFA (0.809), NEWS (0.746), and MEWS (0.664) (Table 1). Performance was consistent across cancer subtypes (AUROC 0.847–0.975). SHAP analysis identified FiO₂, GCS, systolic blood pressure, SpO₂, and cancer-specific markers (calcium, albumin, lactate, LDH) as top predictors. Decision curve analysis confirmed net clinical benefit across a wide range of threshold probabilities. Conclusions: Onc-EWS, a cancer-specific ML early warning score incorporating tumor-related biomarkers alongside physiological parameters, substantially outperformed conventional EWS in predicting clinical deterioration among oncology ICU patients. These findings support the development of disease-specific risk stratification tools for critically ill cancer patients.

Discrimination performance on temporal validation set (2017–2022).