Construction and Validation of a 90-Day Mortality Risk Prediction Model Based on Interpretable Machine Learning for...

Area of Science:

• Neurology
• Medical Informatics
• Biostatistics

Background:

• Accurate prediction of postprocedural mortality is crucial for clinical decisions in mechanical thrombectomy.
• Limited research exists on mortality risk models for patients undergoing mechanical thrombectomy.

Purpose of the Study:

• To develop and validate machine learning models for predicting 90-day mortality in patients undergoing mechanical thrombectomy.
• To identify key predictors of mortality in this patient population.

Main Methods:

• Retrospective-prospective cohort study with 973 patients.
• Development and evaluation of eight machine learning models using 10-fold cross-validation.
• Performance assessment via discrimination, calibration, decision curve analysis, and interpretability.

Main Results:

• Key predictors included age, hyperlipidemia, atrial fibrillation, pre-stroke statin use, and dysphagia.
• Logistic regression (AUC=0.87) and multilayer perceptron (AUC=0.77) models showed significant predictive power.
• 90-day mortality rates were 25.6% (modeling) and 32.0% (validation).

Conclusions:

• Machine learning algorithms accurately predict 90-day mortality post-mechanical thrombectomy.
• The multilayer perceptron model offers a validated tool for personalized risk assessment and clinical decision-making in stroke patients.