Development of an explainable machine learning model for predicting poststroke anxiety: A multicenter study using...

Area of Science:

• Neurology
• Artificial Intelligence
• Clinical Medicine

Background:

• Poststroke anxiety (PSA) negatively impacts recovery and quality of life.
• Existing predictive methods for PSA lack feature selection and interpretability.
• Early detection of high-risk PSA patients is crucial for timely interventions.

Purpose of the Study:

• To develop an interpretable machine learning model for early detection of high-risk PSA patients.
• To utilize a comprehensive dataset including demographic, clinical, biochemical, and psychosocial factors.
• To enable personalized interventions for improved poststroke outcomes.

Main Methods:

• Retrospective multicenter study of 238 stroke patients.
• Feature selection using univariate analysis and LASSO regression.
• Development and evaluation of seven machine learning models, including logistic regression and XGBoost, with cross-validation.
• Application of SHAP (Shapley Additive Explanations) for feature importance and nomogram development.

Main Results:

• Logistic regression model achieved an AUC of 0.981, accuracy of 0.917, sensitivity of 0.867, specificity of 0.952, and F1 score of 0.897.
• Key predictors identified include recurrent stroke, socioeconomic factors, lifestyle, comorbidities (hypertension, diabetes), biochemical markers (WBC, TC, LDL, FIB, APTT), and clinical scores (NIHSS, Barthel index).
• A nomogram incorporating top 10 SHAP-ranked features was created for clinical decision-making.

Conclusions:

• The developed machine learning model demonstrates high accuracy and interpretability in predicting PSA risk.
• SHAP analysis and nomogram visualization provide a practical tool for clinicians.
• This approach facilitates early identification of high-risk PSA patients and personalized management strategies.