Feature-token clinical transformer for 24-h landmark mortality risk stratification in critically ill patients with heart failure: development and multicenter external validation

The first 24 h after ICU admission in critically ill patients with heart failure yield heterogeneous EHR evidence across vital signs, laboratory tests, treatments, organ support, and nurse-charted monitoring. Conventional scores and flat-vector ML models may not preserve individual variable identity or context.

We developed FTCT, encoding each predictor recorded during the first 24 h after ICU admission as a clinically tagged token (value, missingness, identity, group). MIMIC-IV (2008-2019) served for development, validation, and internal testing; eICU-CRD (2014-2015) served for external validation. Included were adults with heart failure, a complete 24-h ICU observation window, and known discharge status. The endpoint was in-hospital all-cause death.

MIMIC-IV included 11,156 stays and 1,484 deaths, partitioned into training (n = 7,809), validation (n = 1,673), and internal test (n = 1,674) sets. eICU-CRD included 9,837 stays and 1,319 deaths. FTCT achieved an AUROC of 0.849 (95% CI, 0.835-0.863) internally and 0.804 (0.788-0.820) externally. Patient-paired AUROC differences vs. XGBoost were 0.027 (0.011-0.043) internally and 0.032 (0.014-0.050) externally. A matched FT-Transformer-style baseline achieved an external AUROC of 0.781 (0.760-0.801). In eICU-CRD, FTCT achieved an AUPRC of 0.296 (0.278-0.314), Brier score of 0.091 (0.085-0.097), and expected calibration error of 0.036 (0.030-0.042). In a post hoc cohort-specific descriptive operating analysis targeting approximately 85% specificity, external sensitivity was 0.662 (0.636-0.687), positive predictive value 0.406 (0.385-0.427), negative predictive value 0.942 (0.937-0.947), and F1 score 0.503. Removing feature-token encoding produced the largest ablation loss in external AUROC (-0.031).

FTCT combines established feature-token Transformer modeling with clinical-group embeddings, explicit missingness representation, and group-aware pooling for 24-h landmark mortality risk stratification from structured EHR data recorded during the first 24 h after ICU admission. Its external validation and ablation results support prospective investigation, but do not yet demonstrate clinical utility, workflow integration, admission-time applicability, or autonomous use.