Development and validation of an interpretable machine learning model for early prediction of mortality in ICU patients with chronic obstructive pulmonary disease and heart failure.
An interpretable CatBoost model built on MIMIC-IV data predicted 28-day all-cause mortality in ICU patients with COPD and heart failure, with an AUC of 0.788 internally and 0.739 in external validation on eICU, and SOFA, SAPS II, anion gap, vasopressor use, and respiratory rate were the top contributing features.
Key Findings
Methods
The study built and compared five machine learning models to predict 28-day all-cause mortality in ICU patients with comorbid COPD and heart failure using data from the first 24 hours of ICU admission.
Development cohort came from the MIMIC-IV database (3167 patients).
External validation cohort came from the eICU Collaborative Research Database (1781 patients).
Models compared: eXtreme Gradient Boosting, Categorical Boosting, Light Gradient Boosting Machine, Random Forest, and Support Vector Machine.
Features were screened using LASSO regression combined with the Boruta algorithm.
Results
Categorical Boosting (CatBoost) showed the best overall performance in internal validation.
The abstract reports an AUC of 0.788 for the CatBoost model.
Sensitivity was 0.482 and specificity was 0.900.
The low sensitivity relative to specificity means the model at its operating threshold identified fewer than half of the patients who died.
Results
The CatBoost model retained moderate discrimination in independent external validation, with a lower AUC than in internal validation.
External AUC was 0.739 in the eICU cohort (1781 patients).
This compares with 0.788 internally, a drop of about 0.05.
The authors describe this as good discriminative ability and reasonable generalizability.
Results
Feature selection with LASSO plus Boruta identified 14 key predictive variables from first-24-hour ICU data.
The two methods were combined to screen clinical features.
The 14 variables were used to build all five machine learning models.
Results
SHAP analysis identified SOFA score, SAPS II score, anion gap, vasopressor use, and respiratory rate as the top five features contributing to model predictions.
The Shapley Additive Explanations framework was applied to the optimal (CatBoost) model.
The authors state the interpretability analysis clarifies core predictive factors to aid clinical risk stratification and intervention.
What This Means
This study asked whether a computer model could use information gathered in the first day of intensive care to predict which patients with both COPD (a chronic lung disease) and heart failure are likely to die within 28 days. The researchers trained several machine learning models on records of 3167 patients from one large US database and then tested the best one on 1781 patients from a separate multi-hospital database.
The best-performing model, called CatBoost, distinguished patients who died from those who survived reasonably well (AUC 0.788 internally, 0.739 on the outside data). It was good at correctly identifying patients who would survive (specificity 0.90) but found less than half of those who died (sensitivity 0.48). The factors that mattered most were two standard illness-severity scores (SOFA and SAPS II), anion gap (a blood chemistry measure), use of blood-pressure-supporting drugs (vasopressors), and breathing rate.
This research suggests such a model could help clinicians sort patients by risk in the ICU, and the explanation tools show which factors drive each prediction. However, performance dropped on external data and the model missed many patients who died, so it would need further testing before being relied on in practice.
Shen J, Liu Y, Chen Y. (2026). Development and validation of an interpretable machine learning model for early prediction of mortality in ICU patients with chronic obstructive pulmonary disease and heart failure.. Medicine. https://doi.org/10.1097/MD.0000000000051224