Fig. 1: Development of a machine learning algorithm to predict 48-hour mortality across ICU care.

a Data preparation: ICU patient data recorded in a patient data management system (PDMS) were exported from the Data Integration Center of the University Medical Center Mannheim, Germany. Data was excluded based on exclusion criteria. Variables included vital parameters, medications, laboratory results, treatments, and outcome metrics and were aggregated based on 24 h time intervals starting at admission. All available static and dynamic categorical and continuous aggregate variables were selected as features and underwent feature processing. Missing values were imputed using previous 24 h time intervals where available or median values from corresponding ICD-10 based disease groups. A separate dataset was created that contained change features indicating day-over day change. b Machine learning: the final dataset was split into a training and test sets. Applying five different linear and tree-based machine learning algorithms, models were trained to predict 48 h ICU mortality at the end of each 24 h time interval. c Model evaluation: for the selected LGBM-48 algorithm, performance was assessed using area under the receiver operating characteristics curve. Selection of cutoffs was performed to identify and map low-, intermediate and high-risk categories based on predicted 48 h mortality risk across the ICU stay. Lastly, sensitivity analyses were performed for different stay days across the ICU stay, different disease groups and endpoints.