Machine learning detection of severity level of maladaptive plasticity in tinnitus and neuropathic pain

Abstract

Tinnitus and NP datasets were analyzed separately and in conjunction, under the hypothesis that they share an underlying mechanism of maladaptive plasticity. Linear and non-linear features were extracted from the EEG signal data, including power spectral density, Shannon entropy and imaginary coherence between channel data. Feature selection with BorutaSHAP wrapper method was applied as part of the model construction process. Classification was performed using both traditional ML algorithms, Random Forest, Support Vector Machine, and k-Nearest Neighbors, and then implementing a prediction voting mechanism, as well as the deep learning neural network EEGNet. In general, SVM had the best performance across experiments. For the tinnitus dataset, the ensemble classifier had the highest accuracy of 50.08%. For the NP dataset, SVM had the best performance measured through an accuracy of 58.43%. For the BC dataset, the highest accuracy score of 42.46% was obtained by SVM. For the EEGNet implementation, the average accuracy obtained in the tinnitus dataset was 51%, the NP dataset was 97.92%, and the BC dataset was 74.31%. EEGNet was the best performing model, particularly for the NP dataset. The top selected features of the feature selection algorithm suggests gamma as a potential biomarker for the detection of maladaptive plasticity.