Abstract
Abstract
Background
Critically ill patients with liver cancer have high in-hospital mortality, but transparent early ICU prediction tools with external validation across independent databases remain limited.
Objective
To develop and externally validate a regression-based early ICU mortality prediction model for adults with liver cancer using routinely available first-24-hour variables and diagnosis-derived baseline clinical/oncologic phenotype features.
Methods
We performed a retrospective cohort study using MIMIC-IV for model development and eICU-CRD for external validation. Eligible patients were adults with current primary liver cancer, ICU length of stay of at least 24 hours, and valid in-hospital outcome data. Candidate predictors included demographic variables, diagnosis-derived phenotype variables, and first-24-hour physiologic, laboratory, and organ-support measures. Missing predictors were handled with multiple imputation. Predictors were selected using bootstrap-stabilized LASSO followed by multivariable logistic regression. Performance was evaluated by discrimination, calibration, Brier score, decision curve analysis, and calibration plots, with emphasis on interpretable predictors available early in the ICU stay.
Results
The derivation cohort included 770 patients from MIMIC-IV, and the validation cohort included 90 patients from eICU-CRD. In-hospital mortality was 16.5% and 13.3%, respectively. The final 12-predictor model included bilirubin, creatinine, mean arterial pressure, heart rate, respiratory rate, SpO2, Glasgow Coma Scale, potassium, temperature, bicarbonate, age, and extrahepatic metastatic disease. The model achieved an AUROC of 0.874 (95% CI 0.842–0.907) in derivation, with an optimism-corrected C-statistic of 0.860, and an external AUROC of 0.707 (95% CI 0.517–0.898). In external validation, the calibration intercept was − 0.252 (95% CI − 1.000 to 0.496), calibration slope 0.474 (95% CI 0.124 to 0.824), and Brier score 0.104.
Conclusions
This transparent early ICU model showed good apparent performance and was externally evaluated in eICU-CRD, supporting cross-database mortality risk stratification. The model is intended to complement, not replace, clinical judgment. Further multicenter validation and local calibration are needed before clinical implementation.