Abstract 8864: TRANSFORM-HF Score: A Machine Learning Survival Model for Mortality Risk Prediction After Heart Failure Hospitalization
Abstract
Background: Mortality after heart failure (HF) hospitalization remains high, yet commonly used risk scores may have limited discrimination, calibration, and generalizability in contemporary practice. Contemporary models should be derived from broadly representative HF populations. Hypothesis: A machine learning survival model using routinely available discharge variables would improve mortality prediction versus established HF risk scores, maintain performance in an external real-world cohort, and identify interpretable predictors of post-discharge mortality. Methods: We developed the TRANSFORM-HF Score using TRANSFORM-HF, a contemporary, pragmatic, multicenter randomized trial of loop diuretic strategy among 2,859 adults hospitalized with HF at 60 U.S. hospitals from 2018-2022, including 36.9% women and 33.9% Black. An XGBoost Cox survival model (Figure 1) was trained on 54 discharge-available clinical variables spanning demographics, vital signs, laboratories, biomarkers, cardiac phenotype, comorbidities, medications, and hospitalization characteristics available in routine practice. Model development used 5-fold stratified cross-validation. External validation was performed in 3,427 consecutive HF hospitalizations from the Yale New Haven Hospital (YNHH) from 2018-2024. Performance was assessed using time-dependent area under the receiver operating characteristic curve (TD-AUC), C-index, Brier score, calibration, and head-to-head comparison with MAGGIC, BCN-Bio-HF, and GWTG-HF scores. Results: The TRANSFORM-HF Score achieved internal TD-AUCs of 0.73 at 12 months and 0.85 at 30 months, with Harrell C-index 0.71 and integrated Brier score 0.13. Compared with existing models (Figure 2), it outperformed MAGGIC (0.677/0.657 at 12/30 months), BCN-Bio-HF Clinical (0.65/0.66), and GWTG-HF (0.68/0.67). External validation tested transportability to an older, more comorbid, HFpEF-predominant YNHH cohort (see Figure 1). Despite these differences, the model maintained discrimination with TD-AUCs of 0.74 at 12 months and 0.73 at 30 months and a C-index of 0.69, again exceeding comparator scores. Feature importance analysis (Figure 3) identified systolic blood pressure (BP), blood urea nitrogen, diastolic BP, serum sodium, and NT-proBNP among the leading contributors, supporting clinical interpretability. Calibration at 12 months was precise without recalibration (observed/expected ratio 0.97); 30-month calibration corrected overestimation (observed/expected ratio from 0.80 to 1.00). Conclusions: The TRANSFORM-HF Score improved mortality risk prediction after HF hospitalization compared with established risk scores and demonstrated generalizability in a distinct real-world cohort. This approach may support individualized post-discharge prognostication, surveillance intensity, and risk-guided clinical decision-making.