AI models reach 98.33% accuracy in ICU mortality study

An ICU team may need to identify a high-risk patient while that patient's condition is still changing. In a study led by Australian Catholic University (ACU) and Charles Darwin University (CDU), an extra trees machine-learning model predicted ICU mortality with 98.33% accuracy; gradient boosting reached 98.23%.

Clinicians commonly use APACHE, the Acute Physiology and Chronic Health Evaluation, and SAPS, the Simplified Acute Physiology Score. The study says these tools can struggle to capture evolving conditions, take time to validate and need frequent recalibration. Machine-learning systems have outperformed such traditional scores, but their lack of explainability has slowed adoption.

The researchers used the extra trees results with explanation models to identify the factors behind the predictions and compare them with medical knowledge. Hypertension, tumors, endocrine disease, digestive disease and cardiovascular disease emerged as key factors. Lead author Niusha Shafiabady, an ACU professor and director of its Women in AI for Social Good lab, led work with colleagues from Charles Darwin University, Amirkabir University of Technology in Tehran and other Australian universities.

So what changes in practice? If embedded in clinical decision-support systems, explained predictions could help clinicians focus urgent attention on high-risk patients, monitor them continuously and intervene earlier. Shafiabady said the tools should complement rather than complicate frontline care.

The study used the MIMIC-III dataset. The researchers identify larger datasets and different health care settings as the next tests, so the reported accuracy remains a research result based on that dataset.