Clinical decision support for vestibular diagnosis: large-scale machine learning with lived experience coaching

Area of Science:

• Neurology
• Medical Informatics

Background:

• Diagnosing vestibular disorders is challenging due to complex symptoms and lengthy patient history requirements.
• Machine learning (ML) applications in medical diagnostics show promise, but their use in classifying vestibular disorders is limited.

Purpose of the Study:

• To develop and evaluate a CatBoost ML model for classifying six common vestibular disorders.
• To integrate algorithmic feature selection with clinical expertise for improved diagnostic accuracy.

Main Methods:

• A retrospective patient dataset was used to train a CatBoost ML model.
• Fifty clinical features were selected using Recursive Feature Elimination-Support Vector Machine (RFE-SVM), SKB score, and expert clinical knowledge.
• The model was designed for high sensitivity in common disorders (e.g., Benign Paroxysmal Positional Vertigo [BPPV], Vestibular Migraine [VM]) and high specificity for complex conditions (e.g., Meniere's Disease [MD], Persistent Postural-Perceptual Dizziness [PPPD]).

Main Results:

• The CatBoost model achieved 88.4% overall accuracy on test data.
• Classification breakdown included 60.9% correct, 27.5% partially correct, and 11.6% incorrect classifications.
• The model demonstrated potential for supporting clinical decision-making in vestibular disorder diagnosis.

Conclusions:

• Machine learning, when combined with clinical expertise, can effectively support the diagnosis of vestibular disorders.
• The developed model shows promise in differentiating between various vestibular conditions, potentially reducing unnecessary invasive treatments.