Decoding the Prognosis of Sepsis-Induced Acute Lung Injury: Machine Learning Model Development and SHAP Interpretability Analysis Based on Critical Care Databases

Background: Sepsis-induced acute lung injury (SI-ALI) has high morbidity and mortality. Early prognosis prediction is critical, but traditional scoring systems lack dynamic, interpretable, and personalized capabilities. This study aimed to develop and validate an interpretable machine learning model for predicting mortality in SI-ALI. Methods: Data from the MIMIC-IV database (training/internal validation) and from the eICU-CRD database (external validation) were analyzed. The Lasso regression model with cross-validation was utilized for the identification of the key features. Subsequently, a series of predicted models were formulated utilizing six distinct machine learning algorithms, and the performance of the models was evaluated using various analytical tools. SHAP values were employed to interpret the machine learning models. Results: Among 7250 patients, eight key indicators were identified. The GBDT model achieved an internal AUC of 0.801 (95% CI: 0.775–0.827), accuracy 0.730, sensitivity 0.692, and specificity 0.739. External validation showed an AUC of 0.681 (95% CI: 0.650–0.708), accuracy 0.601, sensitivity 0.686, and specificity 0.574. However, precision, F1, and MCC were modest, indicating limited predictive accuracy. SHAP identified urine output, respiratory rate, BUN, PLR, lactate, INR, age, and anion gap as the most influential variables. Conclusions: The GBDT model showed modest discriminative ability and limited generalizability. It should be considered an exploratory tool requiring further prospective multicenter validation before clinical application.