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

Abstract

Diagnosing vestibular disorders remains challenging due to complex symptoms and extensive history-taking required. While machine learning approaches have shown promise in medical diagnostics, their application to vestibular disorder classification has been limited. We developed a CatBoost machine learning model to classify six common vestibular disorders using a retrospective dataset of patients. The model incorporates 50 clinical features, selected through a hybrid approach combining algorithmic methods (RFE-SVM and SKB score) and expert clinical knowledge. We designed the system to achieve high sensitivity for common vestibular disorders (BPPV and VM) and high specificity for conditions requiring intensive interventions (MD and HOD) or careful differential diagnosis (PPPD and VEST) to minimize unnecessary invasive treatments. When applied to test data, reaches 88.4% accuracy, with 60.9% correct classifications, 27.5% partially correct, and 11.6% incorrect classifications. Results suggest that machine learning can support clinical decision-making in vestibular disorder diagnosis when combining algorithmic capabilities with clinical expertise.