Application of machine learning in the diagnosis of vestibular disease

Abstract

Machine learning is considered a potential aid to support human decision making in disease prediction. In this study, we determined the utility of various machine learning algorithms in classifying peripheral vestibular (PV) and non-PV diseases based on the results of equilibrium function tests. A total of 1009 patients who had undergone our standardized neuro-otological examinations were recruited. We applied five supervised machine learning algorithms (random forest, adaboost, gradient boosting, support vector machine, and logistic regression). After preprocessing the data, optimizing the hyperparameters using GridSearchCV, and performing a final evaluation on the test set using scikit-learn, we evaluated the predictive capability using various performance metrics, namely, accuracy, F1-score, area under the receiver operating characteristic curve, precision, recall, and Matthews correlation coefficient (MCC). All five machine learning algorithms yielded satisfactory results; the accuracy of the algorithms ranged from 76 to 79%, with the support vector machine classifier having the highest accuracy. In cases where the predictions of the five models were consistent, the accuracy of the PV diagnostic results was improved to 83%, whereas it increased to 85% for the non-PV diagnostic results. Future research should increase the number of patients and optimize the classification methods to obtain the highest diagnostic accuracy.

Figure 1

Overview of our machine learning process.

Figure 2

Comparison of ROC curves among the five machine learning models. ROC receiver operating characteristic, AUC area under the curve, RF Random Forest, AB AdaBoost, GB Gradient Boosting, SVM Support Vector Machine, LR Logistic Regression.

Figure 3

Top 10 most important features, ranked from high to low by classifier models. (a) Random forest. (b) Adaboost. (c) Gradient boosting. RF Random Forest, AB AdaBoost, GB Gradient Boosting, CP canal paresis in caloric test, DP directional preponderance in caloric test, PSRT pendular sinusoidal rotation test, R right, L left, Op open, Cl close, OKN optokinetic, CW clockwise, CCW counterclockwise, FFS failure of fixation suppression test, ETT eye tracking test.

Figure 4

Evaluation of the accuracy of the five machine learning models in the testing dataset (n = 253). PV peripheral vestibular disease, Non-PV non-peripheral vestibular disease.