K-MIMIC: a nationwide Korean multi-institutional Multimodal intensive care dataset

Recent advancements in critical care have highlighted the need for comprehensive, multimodal datasets to support clinical decision-making and advancing artificial intelligence (AI) research. However, such datasets are scarce in Asia. We developed the Korean Multi-Institutional Multimodal Intensive Care (K-MIMIC) dataset by integrating structured electronic medical records (EMRs), high-resolution bio-signals, and medical imaging from multiple hospitals in Korea.

This retrospective multicenter study collected intensive care unit (ICU) data from 278,274 patients admitted to 71 ICUs across 10 hospitals between 2001 and 2023. The data modalities included structured EMRs, physiological waveforms, and imaging studies. Data extraction followed standardized protocols and de-identification procedures in compliance with the Korean Health Data Utilization Guidelines. Multimodal linkage was achieved at the patient level to enable temporal trajectory analysis.

The K-MIMIC dataset contains 287,274 ICU admissions from 241,805 unique patients, including 22,588 bio-signal files and 496,999 imaging studies, primarily chest X-rays aligned with EMRs. Nearly 47% of ICU admissions originated in the emergency department (ED). Elderly patients (65-90 years old) constituted the largest age group. Fifteen thousand, five hundred forty-eight patients had EMR data linked with both bio-signals and imaging, enabling full multimodal analyses.

The K-MIMIC is the first large-scale, multicenter, multimodal ICU dataset in Asia to provide a robust resource for critical care research, including AI-based prediction, monitoring, and longitudinal outcome studies. The dataset demonstrates the feasibility of secure and standardized ICU data integration across diverse institutions.