A comprehensive PROBAST-based assessment of mortality prediction models in adult intensive care units; an updated scoping review

Background Numerous mortality prediction models have been developed for general adult intensive care unit (ICU) populations, but concerns remain regarding their methodological quality, risk of bias, and validation. Previous reviews primarily identified available models and their predictors, without a comprehensive model-by-model risk-of-bias assessment. We therefore updated the existing evidence base and systematically assessed identified ICU mortality prediction models using the Prediction model Risk Of Bias ASsessment Tool (PROBAST), aiming to identify methodologically robust models warranting further validation. Methods and results We built on a systematic scoping review and its first update covering studies published up to February 8, 2023, and updated the search in PubMed (MEDLINE), Embase (Ovid), Web of Science, and CENTRAL through May 11, 2026. Eligible studies developed mortality prediction models for general adult critically ill populations. Data extraction and PROBAST assessment were performed independently by reviewer pairs using a standardized CHARMS- and PROBAST-based form. In total, 56 studies describing 84 mortality prediction models were included. Overall, 82/84 (98%) models were judged at high risk of bias, predominantly because of shortcomings in the analysis domain. Internal validation was absent in 21/84 (25%) models, while only 2/84 (2%) underwent external validation in the original publication. Only one model, ICNARC II, was judged at overall low risk of bias, while PREDICT was judged at unclear risk of bias. A prospective citation search identified no external validation studies for either model. Conclusion Nearly all identified ICU mortality prediction models were at high risk of bias, limiting confidence in their reported performance and practical implementation of their intended use. Rather than continuing to develop additional models, future research should prioritize rigorous external validation and, where necessary, updating of promising existing models while adhering to established methodological standards.