Embedded Files
In a retrospective MIMIC-IV cohort of 5,280 adult ICU patients with ischemic stroke (cerebrovascular accident) plus intracranial artery stenosis/occlusion, machine learning models predicted all-cause mortality with AUC ~0.82–0.83 and accuracy >73% on an independent test set (n=1,584). LightGBM showed balanced performance (recall 0.70; precision 0.72) with good calibration, and SHAP interpretability highlighted acute physiology score III, suspected infection, Charlson comorbidity index, age, admission weight, and RDW as key drivers.
Why It Matters To Your Practice
Conventional ICU risk scores often underperform for individualized prognostication in complex stroke patients; this study suggests ML can maintain discrimination (AUC ~0.83) while adding patient-level interpretability.
Calibration and explainability (via SHAP) are central if model outputs are to be trusted for real-time triage, goals-of-care discussions, and resource planning.
Clinical Implications
These models could augment—not replace—clinical judgment when estimating mortality risk in ICU ischemic stroke with intracranial arterial disease, especially when severity and comorbidity profiles are extreme.
Top predictors (APS III, suspected infection, CCI, age, RDW) align with bedside reasoning and may help clinicians sanity-check model outputs before acting on them.
Good calibration matters: a well-calibrated probability estimate is more actionable than rank-order risk alone for discussions about escalation, monitoring intensity, and disposition.
Insights
Methodologically, feature selection and imputation were performed within the training set then applied to the test set—an approach that helps reduce leakage risk in retrospective modeling.
Logistic regression performed comparably to more complex learners (LightGBM/Bagging), a reminder that “simpler” models can be competitive when the signal is strong and features are well curated.
Interpretability was delivered at both global and individual levels using SHAP, which can support clinician adoption but does not guarantee transferability across hospitals or EHR workflows.
The Bottom Line
In MIMIC-IV ICU ischemic stroke with intracranial stenosis/occlusion, ML models achieved AUC ~0.82–0.83 with >73% accuracy and interpretable drivers of risk.
Next steps before practice impact: external validation, monitoring for dataset shift, and prospective workflow testing to confirm real-world performance and usability.
Page updated
Report abuse