Identifying subgroups of ICU patients with high mortality rates using machine learning: A nationwide, population-based study

PURPOSE: Identifying subgroups of intensive care unit (ICU) patients with high mortality rates can provide directions in policy making about appropriate intensive care medicine. The objective of the study was to identify demographical and clinical characteristics of subgroups of ICU patients with high (≥80%) mortality rates at six months post-ICU admission. METHODS: Data from all Dutch ICU patients admitted between 2013 and 2023 are used. Data are obtained from the Dutch National Intensive Care Evaluation (NICE) registry, including 807,727 ICU admissions from 84 hospitals. A machine learning model was trained on four samples, defined at different stages of the ICU admission with varying amounts of available data. Training was performed on 70% of the hospitals and validation on the other 30% for the years 2013-2022. A temporal validation was performed on data from 2023. RESULTS: Ten high-mortality subgroups were identified. Reduced urine output and a low combined score on the eye & motor components of the Glasgow Coma Scale (GCS) were the most common risk factors defining high-mortality subgroups. External validation showed small deviations in mortality (median absolute deviations -1% and -2%), with one subgroup falling below the 80% mortality threshold. CONCLUSION: Interpretable machine learning can identify ICU patient subgroups with ≥80% 6-month mortality using routinely collected data. These groups are predominantly marked by impaired consciousness and reduced urine output. Future work should integrate these insights into ethical, patient-centered frameworks that support appropriate care.