Automating Ischemic Stroke Subtype Classification Using Machine Learning and Natural Language Processing.

Area of Science:

• Neurology
• Medical Informatics
• Computational Biology

Background:

• Manual classification of ischemic stroke (IS) subtypes is essential for patient management and outcome prediction but is inefficient and prone to errors.
• Scaling manual classification to large datasets is challenging, hindering epidemiological research.

Purpose of the Study:

• To develop and validate an automated system for ischemic stroke subtyping using natural language processing (NLP) and machine learning (ML) on electronic health records (EHR).
• To assess the agreement between ML-based IS subtyping and manual adjudication by expert neurologists.

Main Methods:

• Unstructured EHR data, including progress notes and neuroradiology reports from IS patients with TOAST subtyping, were analyzed using NLP.
• Feature selection and 5-fold cross-validation were employed to optimize ML models.
• Multiple ML algorithms were tested, and their agreement with manual classification was measured using kappa values.

Main Results:

• The best ML model, using combined EHR data, achieved a kappa of 0.57 compared to manual classification.
• Agreement varied by subtype, with cardioembolic stroke showing the highest kappa (0.64) and cryptogenic stroke the lowest (0.47).
• In a blinded test set of 50 cases, the ML system achieved a kappa of 0.72, agreeing with expert raters in 40 cases.

Conclusions:

• Automated ML approaches utilizing textual EHR data demonstrate significant agreement with manual TOAST classification for ischemic stroke.
• This automated pipeline holds potential for enabling large-scale stroke epidemiology research, pending external validation.