Machine learning method intervention: Determine proper screening tests for vestibular disorders

Abstract

Objective: To evaluate the performance of different vestibular indicators in disease classification based on machine learning method.

Methods: This study use retrospective analysis of the vertigo outpatient database from a tertiary care general hospital. 1491 patients with definite clinical diagnoses were enrolled in this study. Spontaneous nystagmus, head-shaking nystagmus, positional nystagmus, unilateral weakness in caloric test, and gain and saccade in video head impulse test (vHIT) were recorded as variables. Diagnoses were mainly reorganized as acute vestibular syndrome, episodic vestibular syndrome, and chronic vestibular syndrome. The trained random forest model was applied based on exploratory data analysis results.

Results: Random forest accuracies on acute, episodic, and chronic vestibular syndrome are 90%, 81.74%, and 91.3%, respectively. The most important features in acute vestibular syndrome are spontaneous nystagmus, and vHIT variables. In episodic vestibular syndrome, unilateral weakness in caloric test, gain and saccades on lateral semicircular canal are the top three parameters. Lateral vHIT gain, head-shaking nystagmus, and unilateral weakness in caloric test are the main parameters on chronic vestibular syndrome. In acute vestibular syndrome, spontaneous nystagmus and vHIT make major contributions in vestibular disorders distinction. When the disease course prolongation, unilateral weakness and head-shaking nystagmus become increasingly important.

Conclusion: Fast clinical test sets including spontaneous nystagmus, head shaking nystagmus, and vHIT should be the first consideration in screening vestibular disorders.

Keywords: Machine learning; Random forest; Vestibular function screening; Video head impulse test.