ECG data and early radiology reports add little to ICU mortality prediction from structured records

Adding electrocardiographic measurements and early clinical text to structured electronic health records did not measurably improve predictions of death within 30 days of ICU admission, according to a multimodal modeling study published in Frontiers in Medical Technology.

The research team, led by Aigerim Ismukhamedova of D. Serikbayev East Kazakhstan Technical University with collaborators at Kazakh-American Free University, Turan University and Politechnika Lubelska, trained a modality-attention model on four kinds of input drawn from the first six hours after ICU admission: tabular EHR features, machine-derived ECG measurements, printed diagnostic statements, and biomedical language-model representations of clinical notes. Mortality was then assessed at 48 hours, seven days and 30 days after admission, with the 30-day endpoint treated as primary. The six-hour window is the period from which predictors were drawn, not the interval over which death was predicted, and features recorded after that window, including total ICU length of stay, were excluded.

A notable constraint shaped the text analysis. Within the first six hours in MIMIC-IV, the only usable free text is the radiology report, which was available for just 19.6 percent of stays. The discharge summary is the only other free-text table in the database, but in 98.6 percent of stays in which it was charted during the window, it was filed only after the window had closed. The study’s conclusions therefore apply specifically to early radiology reports and ECG summaries available in that narrow opening period.

The cohort comprised each adult’s first ICU stay in MIMIC-IV version 3.1, totaling 65,366 stays, of which 9,806 were held out for testing, where the 30-day mortality rate was 10.6 percent. The authors deliberately imposed no minimum stay length, avoiding a form of survival conditioning that can inflate apparent performance. Five model variants were trained with five random seeds per horizon.

Using structured EHR features alone, the model achieved AUROCs of 0.9371, 0.9106 and 0.8969 for the 48-hour, seven-day and 30-day endpoints respectively. Adding ECG data, text, or both produced no reliable improvement. At the 30-day endpoint, every paired AUPRC difference included zero in its confidence interval, and the largest gain compatible with the data, 0.0068, was of the same order as the study’s own training noise. The authors note these are confidence limits, not formal equivalence bounds. No defect was identified in the preprocessing checks performed on the ECG matrix, yet the matrix’s highest-ranked input still ranked only 106th of 1,827 features overall.

The deep architecture did outperform simpler alternatives. L2-penalized logistic regression reached an AUROC of 0.8844, and the matched modality-attention architecture added 0.0126 AUROC, with a confidence interval of 0.0083 to 0.0168. All learned models outperformed conventional severity scores, exceeding SOFA at 0.731 and SAPS-II at 0.793, with DeLong test p-values below 10 to the power of minus 59.

External validation tested the structured backbone rather than the full multimodal system. Transferred without retraining, the models reached an AUROC of 0.833 on HiRID, 0.774 on MC-MED and 0.7268 on eICU-CRD. None of these cohorts supplied ECG data or notes, so both extra modalities were masked, and the 30-day outcomes in HiRID and MC-MED were derived from length of stay rather than recorded directly.

The study also surfaced weaknesses that aggregate metrics concealed. Mean calibration gaps hid over-prediction of risk reaching 0.167 in the highest-risk decile, meaning the model told some of the sickest patients they were more likely to die than the data supported. Performance also varied across subgroups: AUROC was 0.853 for individuals aged 80 or above, while race-stratified AUROC ranged from 0.878 to 0.914 compared with 0.897 overall.

For clinicians and data scientists weighing investments in multimodal ICU pipelines, the message is measured. Structured electronic health records captured in the first six hours already carry most of the available prognostic signal for early mortality prediction, and the marginal value of ECG summaries and early radiology text within that window appears small, at least in the data examined here.

Subject of Research: Multimodal machine learning for early ICU mortality prediction using electronic health records, ECG measurements and clinical text

Article Title: 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

Article References: Ismukhamedova, A., Bakanova, A., Belginova, S., Gola, A., & Khalimov, A. (2026). 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. Frontiers in Medical Technology, 8, 1902843. https://doi.org/10.3389/fmedt.2026.1902843

Image Credits: AI Generated

DOI: 10.3389/fmedt.2026.1902843

Keywords: ICU mortality prediction, electronic health records, electrocardiogram, clinical notes, MIMIC-IV, machine learning, AUROC, external validation, modality fusion, SOFA, SAPS-II, model calibration