Interpretable machine learning based on the Charlson comorbidity index predicts 28-day mortality in acute hypercapnic respiratory failure.

Acute hypercapnic respiratory failure (AHRF) is a major cause of mortality in intensive care units (ICUs). Early identification of high-risk patients with poor prognosis is crucial, as timely and effective interventions can significantly improve survival rates. However, effective models for predicting short-term mortality in AHRF patients remain limited. This study extracted clinical data from 4,302 AHRF patients in the Medical Information Mart for Intensive Care IV (MIMIC-IV) database (version 3.1), with 28-day all-cause mortality as the primary outcome. Univariate and multivariate Cox regression analyses, along with the Boruta algorithm and least absolute shrinkage and selection operator (LASSO) regression, consistently revealed a significant association between the Charlson Comorbidity Index (CCI) and 28-day mortality. Even after propensity score matching (PSM) was applied to balance the baseline characteristics between the high and low CCI groups, the difference in 28-day mortality remained statistically significant. Restricted cubic spline (RCS) analysis demonstrated an U-shaped relationship between CCI and the 28-day survival probability, indicating that higher CCI values were associated with an increased risk of adverse outcomes. Compared to the low-CCI group, patients in the high-CCI group exhibited a significantly elevated risk of 28-day mortality (P = 0.004). Subgroup analyses further suggested that the predictive value of CCI was more pronounced among patients with chronic pulmonary disease and those with a partial pressure of arterial oxygen/fraction of inspired oxygen (PaO₂/FiO₂) ratio between 100 and 149, compared to patients with a ratio > 150 or < 100. Feature selection using the Boruta algorithm identified CCI as a key variable with a high Z-score. Among the developed machine learning models, the Light Gradient Boosting Machine (LightGBM) algorithm achieved the best overall performance (internal validation results accuracy = 0.7793 [0.7516, 0.8070], area under the curve [AUC] = 0.8158 [0.7595, 0.8721]; external validation results on the eICU Collaborative Research Database (eICU-CRD): accuracy = 0.8003 [0.7860–0.8147], AUC = 0.7629 [0.7213–0.8045]), underscoring its robustness in predicting 28-day mortality in AHRF patients. Moreover, the Acute Physiology Score III (APSIII) also demonstrated potential predictive value for adverse clinical outcomes in this population.