Abstract
Using machine learning, we attempted to differentiate between peripheral vestibular disorders (n = 466) and non-peripheral vestibular disorders (n = 254) based on the results of stabilometry. Six algorithms were used for machine learning: random forest, gradient boosting, support vector machine, logistic regression, k-nearest neighbor, and multilayer perceptron. Due to the large difference in the amount of data between the two groups, SMOTE (Synthetic Minority Over-sampling Technique) was used during learning to correct for the amount of data between the two groups.
The results were as follows. (1) The average value and standard deviation of accuracy for the six models were 0.64 and 0.05. Precision and recall were relatively good in the peripheral vestibular disorders group, but poor in the non-peripheral vestibular disorders group. (2) The accuracy rate of prediction of peripheral vestibular disorders by the three algorithms, RF, LR, and KNN, was as high as 90%, whereas their accuracy rate for predicting non-peripheral vestibular disorders was poor (53%).
The insufficient number of cases in the non-peripheral vestibular disease group appeared to have a large influence on the results. Therefore, we would like to collect more cases and repeat the analysis.
