Using Base-ml to Learn Classification of Common Vestibular Disorders on DizzyReg Registry Data

Area of Science:

• Neuroscience
• Medical Informatics
• Ophthalmology

Background:

• Multivariable analyses (MVA) and machine learning (ML) show potential for clinical decision support in neuro-otology.
• Large patient datasets can advance vestibular research.
• The DizzyReg patient registry provides data for developing and testing MVA/ML tools.

Purpose of the Study:

• To develop and evaluate base-ml, a comprehensive MVA/ML software tool for neuro-otology.
• To apply base-ml to classify common vestibular disorders using clinical data.
• To benchmark the performance of various MVA/ML algorithms on distinct clinical tasks.

Main Methods:

• Base-ml integrates data pre-processing, nested cross-validation, hyper-parameter optimization, and 11 diverse classifiers.
• Classifiers include logistic regression, random forests, artificial neural networks, and a graph-based deep learning model.
• Explainable AI tools were used to analyze feature importance and input distributions.

Main Results:

• High accuracy (up to 92.5%) was achieved in classifying bilateral vestibular failure versus functional dizziness.
• Moderate accuracy (56.5-64.2%) was obtained for differentiating primary versus secondary functional dizziness.
• Lower accuracy (25.9-50.4%) was observed when classifying four episodic vestibular disorders.

Conclusions:

• Classification accuracy correlates with the clinical diagnostic difficulty of vestibular disorders.
• Employing multiple MVA/ML algorithms is recommended to avoid underestimating classification performance.
• Base-ml offers a standardized, open-source platform for benchmarking MVA/ML classifiers in clinical neuro-otology.