Development and external validation of a multimodal artificial intelligence mortality prediction model of critically ill patients using multicenter data
Background Early prediction of in-hospital mortality in critically ill patients can aid clinicians in optimizing treatment. The objective was to develop a multimodal deep learning model, using structured and unstructured clinical data, to predict in-hospital mortality risk among critically ill patients after their initial 24 hour intensive care unit (ICU) admission. Methods We used data from MIMIC-III, MIMIC-IV, eICU, and HiRID. A multimodal model was developed on the MIMIC datasets, featuring time series compon…

"Intensive Care" by emilio labrador is licensed under CC BY 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/2.0/.
By August 2026, researchers had developed and externally validated a multimodal deep learning model that uses structured data, clinical notes, and chest X-ray images from the first 24 hours of ICU admission to predict subsequent in-hospital mortality. Trained on MIMIC datasets and tested on more than 200 hospitals including eICU and HiRID, the model achieved high discrimination and calibration.
The work matters because early mortality risk stratification can help clinicians optimize treatment for critically ill patients, and the study demonstrates that combining multiple data sources improves performance over structured data alone. What remains uncertain is prospective clinical impact, workflow integration, and generalizability beyond the retrospective multicenter datasets reported.
- Model developed on MIMIC-III and MIMIC-IV using time-invariant variables, time-variant variables, clinical notes, and chest X-ray images from first 24 hours of ICU admission.
- Total of 203,434 ICU admissions from more than 200 hospitals between 2001 to 2022 included, with mortality rate ranged from 5.2% to 7.9% across four datasets.
- External validation performed in temporally separated MIMIC population, HiRID, and eICU datasets, including eight different institutions within eICU with AUROCs ranging from 0.84-0.92.
- In patients with notes and imaging available, multimodal integration improved Brier score from 0.37 to 0.17 and AUROC improvement was statistically significant based on DeLong test (p=0.0167).
A multimodal deep learning model using first-24-hour ICU data predicted subsequent in-hospital mortality with high discrimination, and adding clinical notes and chest X-ray images further improved performance.
The rundown
Researchers built a multimodal deep learning model on MIMIC-III and MIMIC-IV that used time series components occurring within the first 24 hours of ICU admission, including time-invariant and time-variant variables, clinical notes, and chest X-ray images, to predict subsequent inpatient mortality.
The study included 203,434 ICU admissions from more than 200 hospitals between 2001 to 2022 across MIMIC-III, MIMIC-IV, eICU, and HiRID, and reported external validation in a temporally separated MIMIC population and in HiRID and eICU, with eICU results broken out across eight institutions.
Sources
- Peer-reviewedAnesthesiology2026-08-04
How should this claim be treated?
ace
The debate