Acute illness severity index for severe cardiovascular disease patients using machine learning approaches

Common intensive care severity scores may be less accurate for critically ill patients with cardiovascular disease. We develop and validate a data-driven severity index for 30 day mortality prediction using routinely collected hospital data. We performed a retrospective multicenter cohort study. Model development used the Medical Information Mart for Intensive Care databases (39,458 admissions), split into training (10,457) and internal validation (29,001). External validation used a large United States multicenter intensive care database (9,222) and an independent Guangdong Second Provincial General Hospital cohort (235). We compared machine learning models with three established severity scores. We evaluated discrimination, calibration, clinical utility, and model explanations. Here we show that a gradient boosted decision tree model achieves the highest discrimination, with areas under the receiver operating characteristic curve of 0.853, 0.802, and 0.853 in the internal and two external cohorts, versus 0.724 to 0.765 for established scores. Calibration is good (standardized mortality ratio 1.00; 95% confidence interval 0.98 to 1.02) and remains consistent across age, sex, and race subgroups. Decision curve analysis shows a higher net benefit than conventional scores across thresholds of 0.20 to 0.60. Key contributors include comorbidity burden, urine output, activity status, respiratory rate, and minimum oxygen saturation. A gradient boosted decision tree-based severity index improves discrimination, calibration, and clinical utility compared with conventional intensive care scores and generalizes across external cohorts. The model explanations support interpretable risk stratification for critically ill cardiovascular disease. Many people with serious heart and blood vessel disease are treated in intensive care units. Doctors often use standard scoring tools to estimate a patient’s risk of dying, but these tools may be less accurate for this group. In this study, we aimed to build a more reliable risk score using information that is routinely recorded in electronic health records. We trained and tested several computer models on a large intensive care database and then checked the best model in two independent patient groups from different hospitals. We find that the best model predicts 30 day death risk more accurately than the standard scores and remains reliable across different patient subgroups. This could help clinicians identify high risk patients earlier and support better planning of care. Future work should test how it performs in real time clinical use. Ren et al. develop and validate an XGBoost-based acute illness severity index for critically ill cardiovascular patients using multicenter ICU data and compare it with APSIII, OASIS, SOFA and machine learning baselines. The index shows better discrimination, calibration and net benefit across validations, with SHAP highlighting key drivers of risk.