New in-hospital mortality prediction model for patients treated in the intensive care units
20.12.2021 - „Premorbid functional status is a highly relevant predictor for in-hospital mortality of patients treated in intensive care units [ICUs]. Almost 30% of our analysed patients had a premorbid functional status, which was strongly associated with mortality. This highlights the need of the inclusion of this predictor in future ICU scoring systems“ says Prof. Jukka Takala, the principal investigator of this multinational, multicenter observational study which included 61,224 patients treated in ICUs to create a new in-hospital mortality prediction model. „Our aim was to develop a prediction model which included established predictors like age, severity of acute illness, admission type, but also relevant predictors like premorbid functional status and diagnoses. This newly developed prediction model is key for the evaluation and benchmarking of ICUs in the Finnish Intensive Care Consortium [FICC] benchmarking program, which includes ICUs from Finland, Estonia and Switzerland.“
The study results summarize that premorbid functional status was strongly associated with mortality and increased the predictive performance of the model, while the inclusion of diagnosis accounted for the large heterogeneity between diagnosis groups. Overall, the prediction model showed excellent internal discrimination and calibration performance.
„The multicenter study design of the FICC cohort study with a large number of admissions allowed us to investigate validation and transportability properties of the prediction model using state-of-the-art statistical methodology“, says Dr. André Moser from CTU Bern. „Internal validation is only one key aspect in the development of a prediction model. Transportability - another important aspect - indicates performance outside the study population used for the development of the prediction model, for example, in other hospitals or time periods. This is an important piece of information for the future application and implementation of the predicition model. Often, ICU scoring systems do not report such transportability properties“.
The study authors used a proposed framework for interpreting validation results which investigated case-mix differences and transportability properties. „We found a large heterogeneity of performance measures between ICUs which can be explained by case-mix differences. This assessment allowed us to make conclusions about where model performance measures of new ICUs likely will lie“, explains Moser. They concluded that the geographical (ie, across hospital units) discriminative ability of the prediction model was very good, while other performance measures had large prediction uncertainty. „Moving from an internal validation approach towards an internal-external validation approach is key for the translational process of a new prediction model and should be regularly used by researchers - if possible“, adds Moser.
The developed prediction model and the gained knowledge of key validation properties will be implemented in the current FICC benchmarking program and will be used for monitoring and health economic evaluations. For example, Takala et al. (2021) used the developed prediction model for the investigation of variation in severity-adjusted resource use in intensive care units (https://doi.org/10.1007/s00134-021-06546-4).
„The whole development process of the prediction model was a very successful collaboration among researchers from different health care systems and settings and was based on fruitful discussions between clinicians, data scientists and statisticians“, conclude Takala and Moser.
The full article can be accessed here.