Machine learning models to predict ICU admission of patients with COPD and depression: A retrospective cohort study

Clinical Snapshot

Key Findings

• P Value: Not reported in abstract

• Effect Size: Random Forest model: sensitivity 0.853, specificity 0.948, PPV 0.924, NPV 0.898, F1 score 0.887, balanced accuracy 0.901. Top SHAP predictors: AKI (mean |SHAP| 0.212), sepsis (0.092), invasive ventilation (0.073), hemoglobin (0.023), haematocrit (0.023), MCHC (0.022)

• Primary Outcome: ICU admission among patients with comorbid COPD and depression (n=1121 first-time hospitalisations from MIMIC-IV v3.1)

• Nnt Or Sensitivity: Sensitivity (recall) 0.853 for ICU admission prediction; specificity 0.948; AUC not explicitly stated in abstract but implied to be high given balanced accuracy of 0.901

• Confidence Interval: Not reported for any performance metric

Clinical Application

The 16 predictor variables are largely derived from routine admission blood tests and clinical documentation, making data collection feasible in most acute care settings. However, the model is not operationalised as a bedside tool, no risk threshold is defined, and implementation would require prospective validation and integration into clinical workflows. The model is not currently available as a validated clinical decision support tool. COPD affects approximately 1 in 13 Australians aged 45 and over (AIHW), and depression is a recognised comorbidity associated with worse outcomes. Australian COPD management is guided by RACGP and Lung Foundation Australia guidelines, which do not currently incorporate ML-based ICU triage tools. The MIMIC-IV database reflects US tertiary-care practice, which differs from Australian public hospital systems in ICU admission thresholds, staffing ratios, and care escalation norms. PBS-listed therapies for COPD (e.g., long-acting bronchodilators, inhaled corticosteroids) and antidepressants are not captured in this model. TGA approval and RACGP endorsement would be prerequisites before any such tool could be considered for Australian clinical use. The findings regarding AKI, sepsis, and invasive ventilation as dominant ICU admission predictors are clinically consistent with Australian critical care practice and may inform risk stratification discussions, but the model itself cannot be directly applied without local validation. Hospitalised adults with confirmed comorbid COPD and depression, particularly those presenting with AKI, sepsis, or requiring invasive ventilation, in tertiary-care settings with access to routine biochemical and haematological data

Abstract

This study aimed to develop and validate a machine learning (ML)-based predictive model to identify risk factors associated with intensive care unit (ICU) admission among patients with comorbid Chronic Obstructive Pulmonary Disease (COPD) and depression. Data of patients diagnosed with both COPD and depression were extracted from the MIMIC-IV version 3.1 database. A total of 1121 patients with first-time hospitalization and comorbid COPD and depression were included. Feature selection was performed using the Boruta algorithm. Five ML algorithms were employed to construct prediction models. Model performance was evaluated using the receiver operating characteristic curve, area under the curve, calibration curves, and decision curve analysis. Shapley additive explanations (SHAP) plots were used to interpret the contribution of each feature to the model's predictions. Sixteen variables identified by the Boruta algorithm were used to build the predictive models, including chloride, prothrombin time, phosphorus, bicarbonate, international normalized ratio, mean corpuscular hemoglobin (MCH) concentration (MCHC), mean corpuscular hemoglobin, hemoglobin, hematocrit (HCT), total calcium, paraplegia, cerebrovascular disease, smoker status, invasivevent, sepsis, and acute kidney injury (AKI). The Random Forest model performed the best: specificity (0.948), positive predictive value (0.924), precision (0.924), F1 score (0.887), balanced accuracy (0.901), sensitivity (0.853), negative predictive value (0.898), and recall (0.853). SHAP analysis indicated that AKI (0.212), sepsis (0.092), and invasivevent (0.073) were the most influential predictors, followed by anemia-related features such as hemoglobin (0.023), HCT (0.023), and MCHC (0.022). ML analysis revealed that AKI, sepsis, invasivevent, and anemia-related indicators are key risk factors for ICU admission among COPD patients with depression. However, it is important to note that the findings lack external validation, which warrants further investigation in diverse patient populations.

References

1. 1.Chen, H., Guo, Y., You, T., & Shi, K. (2026). Machine learning models to predict ICU admission of patients with COPD and depression: A retrospective cohort study. Medicine. https://doi.org/10.1097/MD.0000000000049519