Machine learning-based ICU mortality prediction across hematologic malignancy subtypes: A comparative analysis using MIMIC-IV.

262

Background: Patients with hematologic malignancies face high ICU mortality, yet prediction models largely treat these diseases as a homogeneous group. Whether risk profiles differ across subtypes is poorly understood, particularly for multiple myeloma (MM) where ICU-specific prediction tools are absent. We developed machine learning (ML) models to predict in-hospital mortality and compared disease-specific predictive features across non-Hodgkin lymphoma (NHL), MM, and acute myeloid leukemia (AML). Methods: Using MIMIC-IV v3.1 (2008-2022), we identified 1,372 adult ICU patients (NHL=742, MM=330, AML=298). A CatBoost model was trained on 134 features extracted from the first 24 hours (demographics, labs, vitals, SOFA score, treatments) with temporal validation (train: 2008-2016, n=1,019; test: 2017-2022, n=353). Combined and subtype-specific models were developed. SHAP analysis identified differential predictors. Patients were stratified into low, intermediate, and high risk groups. Results: Overall in-hospital mortality was 24.0%. AML had the highest mortality across all endpoints (in-hospital 40.3% vs NHL 19.7% vs MM 19.1%; 90-day 52.0% vs 33.2% vs 32.4%). The combined model achieved AUROC 0.83 on temporal validation. The MM-specific model outperformed the combined model (AUROC 0.92 vs 0.89), while the AML-specific model underperformed (0.72 vs 0.77), suggesting myeloma harbors distinct predictive signals diluted in mixed cohorts. Risk groups showed observed mortality of 7.9% (low, n=189), 31.6% (intermediate, n=79), and 60.0% (high, n=85). BUN/creatinine ratio and respiratory rate were top predictors across all subtypes; platelet parameters were disproportionately important in AML, while SpO2 carried greater weight in MM. Model performance remained stable in sensitivity analyses excluding transplant patients (0.81), early deaths (0.80), and across age subgroups (Table 1). Conclusions: An ML model using first-24-hour ICU data predicts in-hospital mortality across hematologic malignancy subtypes (AUROC 0.83). Disease-specific models reveal that MM benefits most from tailored prediction, while AML benefits from shared cross-disease features. These findings support subtype-aware risk stratification in critically ill hematologic malignancy patients.