A Machine Learning Approach for Predicting 28-Day ICU Mortality in ARDS Patients Based on the New Global Definition

Technical Summary: A Machine Learning Approach for Predicting 28-Day ICU Mortality in ARDS Patients Based on the New Global Definition

Problem Statement
Acute Respiratory Distress Syndrome (ARDS) is a critical condition with a high in-hospital mortality rate (35–45%). While the 2012 Berlin definition has guided clinical practice, it presents limitations, particularly in diagnosing patients not receiving mechanical ventilation or lacking arterial blood gas analysis. The 2023 New Global Definition of ARDS addresses these gaps by incorporating noninvasive ventilation and pulse oximetry-derived oxygenation (SpO₂/FiO₂). However, there is a lack of research utilizing machine learning (ML) to predict clinical outcomes specifically for patient cohorts defined by this updated framework. This study aims to bridge that gap by developing and validating an interpretable prognostic model for 28-day ICU mortality in ARDS patients diagnosed under the New Global Definition.

Methodology
The study employed a two-cohort design utilizing data from the MIMIC-IV database (version 2.2) for training and internal validation, and clinical data from the Affiliated Hospital of Xuzhou Medical University for external validation.

• Data Selection: Patients were included if they met the New Global Definition criteria (acute onset, bilateral infiltrates, specific ventilation/oxygenation thresholds) within 24 hours of ICU admission. The training cohort consisted of 3,382 patients, while the external validation cohort included 100 patients.
• Feature Engineering: From an initial pool of clinical variables (demographics, vital signs, laboratory results, imaging text, and treatments), missing values were handled via multiple imputation (mice forest). Variables with >30% missingness were excluded.
• Feature Selection: LASSO regression with cross-validation (LASSO-CV) was used to mitigate collinearity and select prognostic variables. This process reduced 64 preprocessed features to 15 key variables, including Logistic Organ Dysfunction Score (LODS), metastatic solid tumors, respiratory rate, age, anion gap, and various electrolytes and hematological markers.
• Model Development: Eight machine learning algorithms were constructed and compared: Logistic Regression (LR), Random Forest (RF), Decision Tree (DT), Support Vector Classifier (SVC), LightGBM, XGBoost, AdaBoost, and Multilayer Perceptron (MLP).
• Validation Strategy: A five-fold nested cross-validation approach (repeated five times) was used for hyperparameter optimization and performance assessment. The Synthetic Minority Oversampling Technique (SMOTE) and Tomek links addressed class imbalance.
• Interpretability: SHapley Additive exPlanations (SHAP) values were calculated to elucidate feature importance and directional effects on the model's predictions.

Key Results

• Feature Importance: The LASSO-CV selection identified 15 significant predictors. The top features influencing prognosis were LODS, presence of metastatic solid tumors, mean respiratory rate, age, and the anion gap.
• Model Performance: Among the eight algorithms, the Support Vector Classifier (SVC) and Logistic Regression (LR) demonstrated the strongest performance.
  • Internal Validation: Both SVC and LR achieved an Area Under the Curve (AUC) of 0.819 (95% CI: 0.786–0.855 for LR; 0.789–0.859 for SVC).
  • External Validation: The SVC model outperformed the LR model, achieving an AUC of 0.871, a lower Brier score, and a higher net benefit in Decision Curve Analysis (DCA).
• Interpretability: SHAP analysis revealed that higher values of LODS, respiratory rate, age, and anion gap were positively correlated with mortality risk. Conversely, higher body temperature, creatinine, and body weight were negatively correlated with risk. The analysis also identified non-linear relationships, such as a threshold age of 71.28 years and a respiratory rate threshold of 21 breaths/min, beyond which mortality risk significantly increased.

Significance and Claims
The authors claim that their study successfully develops a reliable, interpretable machine learning model for predicting 28-day mortality in ARDS patients based on the 2023 New Global Definition. The primary significance lies in:

1. Adaptation to New Criteria: The model is specifically tailored to the updated diagnostic criteria, which expand applicability to patients on noninvasive ventilation and those in resource-limited settings where arterial blood gases may be unavailable.
2. Clinical Utility: The SVC model demonstrated superior generalizability in external validation compared to traditional logistic regression, suggesting robustness in real-world clinical settings.
3. Interpretability: By utilizing SHAP values, the study moves beyond "black box" predictions, offering clinicians insights into specific risk factors (e.g., the impact of metastatic tumors or specific physiological thresholds) to support personalized treatment planning.

The paper concludes that while limitations exist regarding the demographic differences between training and validation cohorts and the retrospective nature of the study, the model provides a valuable tool for assisting clinicians in formulating personalized treatment strategies to improve survival rates in ARDS patients.