CanICU

CanICU predicts 28-day mortality among critically ill cancer patients in intensive care units using a random forest classifier trained on multi-cohort clinical and laboratory data.

Key Features:

• Algorithm: Employs a random forest machine learning algorithm to construct the predictive classifier.
• Input variables: Utilizes nine clinical and laboratory factors as model features.
• Training cohorts: Developed using data from MIMIC (USA, 766 patients), Yonsei Cancer Center (YCC, Korea, 3,571 patients), and Samsung Medical Center (SMC, Korea, 2,563 patients), totaling 6,900 patients collected from January 2, 2008 to December 31, 2017.
• Evaluation metrics: Performance assessed using sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC).
• Primary cohort performance: Achieved 96% sensitivity, 73% specificity, and AUROC of 0.94 in the primary analysis.
• External validation performance: Maintained sensitivity of 79–89%, specificity of 58–59%, and AUROC of 0.75–0.78 across YCC, SMC, and MIMIC-III cohorts.
• Comparative performance: Outperforms APACHE and SOFA prognostic scores for short-term mortality prediction in critically ill cancer patients.
• Long-term prediction: Demonstrates predictive capability for one-year mortality with reported specificity of 88–93%.

Scientific Applications:

• Short-term mortality prediction: Predicts 28-day mortality for critically ill cancer patients admitted to ICUs.
• External validation and generalizability: Validated across MIMIC-III, YCC, and SMC cohorts to assess model generalizability.
• Prognostic benchmarking: Provides comparative prognostic assessments relative to APACHE and SOFA scoring systems.
• Long-term outcome prediction: Correlates model scores with one-year mortality for longer-term prognostication.
• ICU risk stratification: Supports risk stratification for ICU care allocation among cancer patients based on predicted mortality risk.

Methodology:

Trained a random forest classifier using nine clinical and laboratory variables on combined datasets from MIMIC (766 patients), YCC (3,571 patients), and SMC (2,563 patients) collected between January 2, 2008 and December 31, 2017, and evaluated model performance using sensitivity, specificity, and AUROC with external validations on YCC, SMC, and MIMIC-III cohorts.