Early radiology reports and ECG machine measurements do not measurably improve 30-day ICU mortality prediction from the first 6 h of the stay: a multimodal ablation with external validation of the structured EHR backbone

Whether electrocardiographic measurements and unstructured clinical notes add prognostic information to structured electronic health records for early ICU mortality prediction remains unresolved. We trained a modality-attention model on tabular EHR features, machine-derived ECG measurements, printed diagnostic statements, and biomedical language-model representations of notes, all drawn from the first 6 h after ICU admission. Within that window, the only usable text in MIMIC-IV is the radiology report, available for 19.6% of stays. The discharge summary is the only other free-text table, but in 98.6% of stays in which it was charted during the window, it was filed after the window had closed. Mortality was assessed at 48 h, 7 days, and 30 days after ICU admission, with 30 days as the primary endpoint. The 6-h window is the period from which the predictors are drawn, not the interval over which death is predicted. Post-window features, including total ICU length of stay, were excluded. The cohort comprised each adult's first ICU stay in MIMIC-IV v3.1 (65,366 stays), holding out 9,806 (30-day mortality 10.6%), without a minimum stay to avoid day-one survival conditioning. Five variants used five seeds per horizon. EHR-only AUROCs were 0.9371, 0.9106, and 0.8969. Adding ECG, text, or both produced no reliable improvement: at 30 days, every paired AUPRC difference included zero, and the largest gain compatible with the data, 0.0068, was of the order of the study's own training noise. These are confidence limits, not equivalence bounds. No defect was identified in the preprocessing checks performed on the ECG matrix, and the matrix’s highest-ranked input still ranked 106th of 1,827. L2-penalized logistic regression reached 0.8844, and the matched architecture added 0.0126 AUROC (0.0083–0.0168). All learned models outperformed SOFA at 0.731 and SAPS-II at 0.793, with DeLong p below 10 ⁻⁵⁹ . Transferred without retraining, the models reached AUROC 0.833 on HiRID, 0.774 on MC-MED, and 0.7268 on eICU-CRD; none of these cohorts supplied ECG or notes, so they tested the structured backbone with both extra modalities masked, and the HiRID and MC-MED 30-day outcomes are derived from length of stay. Mean calibration gaps concealed over-prediction up to 0.167 in the highest-risk decile. AUROC was 0.853 for individuals aged 80 or above; race-stratified AUROC remained 0.878–0.914 compared with 0.897 overall. These findings pertain to early radiology reports and ECG summaries within a 6-h window.