A novel machine learning framework for stroke type identification in resource constrained settings with robustness to...

Area of Science:

• Neurology
• Medical Informatics
• Artificial Intelligence in Healthcare

Background:

• Stroke is a leading global cause of death and disability.
• Accurate stroke-type identification (ischemic vs. hemorrhagic) is crucial for effective treatment.
• Neuroimaging is often inaccessible in resource-limited settings, hindering timely diagnosis.

Purpose of the Study:

• To develop a cost-effective machine learning (ML) framework for stroke-type identification using only clinical data.
• To provide a diagnostic tool for resource-limited settings lacking neuroimaging facilities.
• To improve prompt primary care and timely referral for stroke patients.

Main Methods:

• Utilized a dataset of 2,190 stroke patients with 79 clinical attributes.
• Employed Multiple Imputation by Chained Equation (MICE) to handle missing data.
• Applied SHAP-analysis to identify key predictive clinical attributes for classification.

Main Results:

• Achieved 82.42% weighted accuracy, 82.33% accuracy, 82.19% sensitivity, 82.65% specificity, and 86.68% F1-score.
• A reduced set of 19 significant attributes maintained 82.20% weighted accuracy.
• Prospective validation showed a 16.42% improvement over the Siriraj clinical score.

Conclusions:

• The proposed ML framework offers a robust and cost-effective method for stroke-type identification.
• This approach can significantly aid clinical decision-making in resource-limited environments.
• The framework has the potential to reduce treatment delays and improve patient outcomes.