Interpretable Machine Learning Model Using Oxygenation and Respiratory Variability to Predict Hemorrhagic Stroke Mortality: A Multicenter Validation Study.

BACKGROUND: Hemorrhagic stroke has high in-hospital mortality, but early risk prediction remains suboptimal. While prior studies have focused on ventilation strategies and mechanical support, limited research has explored the prognostic significance of oxygenation status or hyperoxia. This may be because single-point measurements often fail to capture the dynamic instability of oxygenation and respiratory status during early ICU care. To address this gap, we aimed to evaluate the prognostic value of oxygenation and respiratory variability and incorporate it into an interpretable machine learning framework for mortality prediction in hemorrhagic stroke.

METHODS: We performed a retrospective cohort study using two publicly available ICU databases. Patients from MIMIC-IV were used for model development and internal validation (n = 2,262), and patients from the eICU Collaborative Research Database were used for external validation (n = 2,076). Candidate predictors were constructed from demographic characteristics, vital signs, laboratory tests, treatments, and oxygenation and respiratory measurements recorded during the first 72 hours after ICU admission. Variability features were derived from SpO₂, PaO₂, and respiratory rate. After feature selection using LASSO and recursive feature elimination, multiple machine learning models were trained and compared. The final gradient boosting machine (GBM) model was interpreted using SHapley Additive Explanations (SHAP).

RESULTS: The GBM model showed strong discrimination in the internal test set and acceptable discrimination in the external validation cohort, with AUROCs of 0.911 and 0.792, respectively. In the external cohort, the model maintained a high negative predictive value but had a low positive predictive value, indicating limited precision for identifying high-risk patients at the applied threshold. SHAP analysis identified HR-min, Glucose-mean, Sodium-mean, and several oxygenation and respiratory variability features, including TPE-RR, Slope-PaO₂, Mean-SpO₂, and SD-SpO₂, as important contributors to mortality prediction.

CONCLUSIONS: We developed and externally validated a machine learning model incorporating oxygenation and respiratory variability for mortality risk estimation in patients with hemorrhagic stroke. The model showed strong internal performance and acceptable external discrimination after a 72-hour ICU observation window. However, the low external positive predictive value and cross-cohort differences in event prevalence and treatment patterns limit immediate clinical application. The model should therefore be considered a risk-estimation framework that requires recalibration and prospective validation before integration into clinical decision-support systems.