Abstract: BACKGROUND: Electroencephalogram (EEG) is a common means to detect and analyze vertigo in clinic. Currently, unipolar or multistage lead tracing is mostly used to record and analyze whether the EEG frequency is abnormal. However, the EEG process of vertigo is extremely complex. It is difficult to accurately classify and detect the vertigo state only using a frequency speed analysis. 
OBJECTIVE: To classify the types of vertigo based on the combination of machine learning and EEG signal analysis, which has certain research significance and clinical application value for the diagnosis of vertigo.
METHODS: The non-invasive vestibular function regulation technology for vestibular electrical stimulation was used to create a reversible vertigo state. The stimulation current intensity was 1, 2, and 4 times that of the skin perception threshold. All subjects were required to fill in a dizziness handicap inventory after different intensity current stimulations. The vertigo symptoms were divided into different grades according to the evaluation results of the dizziness handicap inventory, which was used as the data label for supervised learning of EEG classification. The stimulated EEG signals were collected, the wavelet energy and wavelet entropy sample features of EEG signals were extracted by wavelet transform, and a variety of machine learning classification models were used to classify the features of samples with or without vertigo and with different levels of vertigo.
RESULTS AND CONCLUSION: By comparing the classification results of various classification models, we found that the supervised learning classification based on the wavelet transform characteristics of EEG signals could realize the binary classification and multi-classification of vertigo and vertigo level. Compared with logistic regression model, support vector machine model, back propagation neural network model, and random forest classification model showed higher accuracy in the binary classification and multi-classification of vertigo detection. The accuracy of the binary classification was up to 82.5% and the operating characteristic curve area was 0.913; the accuracy of the three-way classification was up to 75.8% and the operating characteristic curve area was 0.927. All these findings indicate that the random forest model has a relatively higher accuracy in the classification of EEG features with or without vertigo and vertigo level. This method provides a feasible supplementary scheme for the classification and detection of vertigo symptoms and offers a new perspective on the diagnosis of vertigo.

Key words: vestibular electrical stimulation, vertigo, electroencephalogram, wavelet transform, classification algorithm