MIMIC-ED-Assist

To support the machine learning (ML) community in developing a time-cost-effective diagnostic assistant, we collaborate with ED clinicians to curate a benchmark, called MIMIC-ED-Assist, that is derived from MIMIC-IV and MIMIC-ED. MIMIC-ED-Assist is designed to test the ability of AI systems to provide both accurate and time-cost saving laboratory recommendations. Our benchmark consists of two prediction targets identified by our clinical collaborators to reflect patient risk: critical outcomes, which include patient death and ICU transfer, and lengthened ED stay, defined as ED LOS exceeding 24 hours. Accurately identifying patients at high risks of these outcomes reduces time-cost by allowing clinicians to perform timely interventions and efficiently allocate resources. MIMIC-ED-Assist mirrors real-world ED practices by grouping individual laboratory tests into commonly performed groups, e.g., complete blood count (CBC). MIMIC-ED-Assist then tests AI systems on their ability to recommend the most informative groups to make accurate diagnostic suggestions while minimizing the total time required to perform these tests, thereby reducing LOS.