This large-scale registry study leveraged interpretable machine learning to identify distinct, high-risk subgroups of intensive care unit (ICU) patients with a six-month post-admission mortality rate of 80% or higher. Utilizing data from the Dutch National Intensive Care Evaluation (NICE) registry, the researchers analyzed an expansive dataset of 807,727 ICU admissions across 84 Dutch hospitals between 2013 and 2023.

The machine learning model was trained and validated using a robust split-hospital approach (70% training, 30% validation) across different stages of ICU admission to account for varying levels of available data, followed by a final temporal validation on 2023 data.

Ultimately, the model successfully isolated ten specific high-mortality subgroups, with reduced urine output and low Glasgow Coma Scale (GCS) eye and motor scores emerging as the most dominant risk factors.

External and temporal validation confirmed the consistency of these findings, with only minor deviations in mortality rates and only one subgroup dropping slightly below the 80% threshold.

The study concludes that routinely collected clinical data can effectively flag these ultra-high-risk populations, offering vital insights that could help shape future ethical frameworks and policy-making for more patient-centered, appropriate intensive care.

Read More