An updated systematic scoping review evaluating mortality prediction models for general adult intensive care unit patients has revealed that nearly all identified tools exhibit substantial methodological flaws. Among 84 distinct mortality prediction models described across 56 published studies, 82 models—representing 98 percent—were judged to be at a high risk of bias. The assessments, conducted using the Prediction model Risk Of Bias ASsessment Tool, indicate that severe methodological limitations fundamentally compromise confidence in the reported performance metrics of these clinical prediction systems.

Researchers initiated the investigation because numerous prognostic models have been introduced into critical care medicine without sufficient scrutiny of their underlying validity. While earlier literature reviews cataloged available ICU mortality models and their predictor variables, those prior assessments did not execute a comprehensive, model-by-model evaluation of bias. To address this persistent gap, the study team sought to systematically evaluate the existing evidence base to determine whether any available models demonstrate sufficient methodological rigor to justify further clinical validation.

The review expanded upon an initial systematic scoping review and a subsequent update that had captured literature published through February 8, 2023. The authors updated comprehensive database searches across PubMed, Embase, Web of Science, and the Cochrane Central Register of Controlled Trials through May 11, 2026. The search focused exclusively on studies describing the development of mortality prediction models intended for general adult critically ill populations admitted to intensive care units.

Data extraction and methodological evaluations were carried out independently by reviewer pairs using standardized forms structured around the CHARMS checklist and the PROBAST framework. Across the 56 eligible studies and 84 individual models analyzed, the high risk of bias was predominantly attributable to analytical shortcomings.

Validation practices across the surveyed literature were found to be remarkably limited. Internal validation was completely absent in 21 of the 84 models, meaning that a quarter of the tools were never statistically tested against the development dataset. Furthermore, only two models—representing just two percent of the total sample—underwent external validation within their original developmental publications, leaving the generalizability of nearly all algorithms unverified at publication.

Across the entire review, only the ICNARC II model achieved an overall judgment of low risk of bias, while the PREDICT model received an unclear risk of bias rating. A subsequent prospective citation search identified zero external validation studies for either of these two leading candidates. Given these pervasive weaknesses, the authors concluded that current ICU mortality prediction models should not be implemented for their intended clinical use in their present form, advising researchers to prioritize the rigorous external validation and updating of promising existing tools rather than generating additional unvalidated models.

Subject of Research: Risk of bias in adult intensive care unit mortality prediction models

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

Article References: Bos, V. B. A., Wang, K. Z., Schellens, J., Cox, E. G. M., van Rosmalen, F., van Herpt, T. T. W., Driessen, R. G. H., van der Horst, I. C. C., Sels, J.-W. E. M., Wynants, L., & van Bussel, B. C. T. (2026). A comprehensive PROBAST-based assessment of mortality prediction models in adult intensive care units; an updated scoping review. Critical Care. https://doi.org/10.1186/s13054-026-06452-5

Image Credits: AI Generated

DOI: 10.1186/s13054-026-06452-5

Keywords: intensive care units, mortality prediction models, PROBAST assessment, risk of bias, scoping review, critically ill adults, ICNARC II, PREDICT model, external validation, internal validation, CHARMS, critical care medicine, prognostic models, methodological quality, analysis domain shortcomings, citation search