Researchers have developed a medical artificial intelligence system that can reliably predict a patient's risk of death even when some test results or medical records are missing. Because the technology addresses the problem of "missing data" — a common occurrence in clinical settings that degrades the performance of medical AI — it is seen as more likely to work in actual hospital practice.
Professor Park Yu-rang and researcher Hong Jae-sung of the Department of Biomedical Systems Informatics at Yonsei University College of Medicine said on the 16th that their team had developed a multimodal AI framework called ETF-UML that can operate on incomplete medical data. The findings were published in the international journal Information Fusion.
Multimodal medical AI predicts a patient's condition by combining different types of information, including basic patient details, test results, vital signs and clinicians' notes. In emergency rooms and intensive care units, however, it is difficult to obtain all of that information at the same time. Blood test results and vital signs change depending on when they are measured, and some clinical records, such as discharge summaries, do not exist at all in the early stages of hospitalization.
Existing AI models have sometimes filled in missing information with estimated values. The research team instead chose not to supplement the data arbitrarily, training the system so that each type of data shared a common basis for judgment. Tabular data such as basic patient information, diagnoses and prescriptions, time-series data such as blood tests and vital signs, and text data such as clinicians' notes were each trained independently.
The team used MIMIC-IV and eICU, two publicly available international intensive care databases, to evaluate performance in predicting the risk of death within 30 days, within 90 days and within 10 days in hospital. The analyses drew on 190,027 records, 164,723 records and 119,852 records, respectively.
ETF-UML recorded a higher AUROC than existing comparison models in all three evaluations. The performance gap widened as the amount of available data was restricted. When only two types of data were used, the system outperformed existing models by 4.4 to 10.7 points in many settings.
Computing efficiency also improved. The peak graphics memory used when combining the data was 0.44 gigabytes, about one-tenth of the 4.70 gigabytes required by MUSE, a comparison model, and the computational load was less than half.
