Design and development of a musical tone detection and identification system in brain wave signals

• Main Author: Navea, Roy Francis R.
• Format: text
• Language: English
• Published: Animo Repository 2017
• Subjects: Electroencephalography; Machine learning; Algorithms; Theory and Algorithms
• Online Access: https://animorepository.dlsu.edu.ph/etd_doctoral/576
• Institution: De La Salle University

Electroencephalogram (EEG) is one of the electrophysiological signals that possesses a high level of randomness and complexity. EEG signals are the electrical activities of the brain which reflect the different activities going on inside a human body. These signals can be stimulated.In this study, EEG signals were stimulated by acoustic energy through musical tones. Stimulated and non-stimulated EEG signals were detected and the tone of stimulation were identified. A musical piece was composed in the key of C with which the C, F and G tones were used. A total of 27 subjects were asked to listen to this piece and while they were listening, their EEG responses were captured using a 14-channel neuro headset. The captured signals were used to train different learning algorithms (Naïve Bayes, Multilayer Perceptron, Support Vector Machine, kNN and Tree) to perform detection and identification.Fourier-based filters were used to extract the EEG frequency range and wavelet denoising was performed to smoothen the signals. The statistical characteristics of the windowed power spectrum vectors of the EEG signals were obtained and were used as features. Several feature selection algorithms (ANOVA-based, Sequential, Ranking, Reliefbased) were used to determine which statistical characteristic/s of the signal could best match the target detection / identification.Classification tasks include detection whether the EEG signals are tone stimulated or not, and identification whether the EEG signals are stimulated by the C, F or G tone. Results show an accuracy of 87% to 94% in detection and 76% to 83% in identification with F-score of 82% to 92% and 66% to 80%, respectively. These were obtained using the alpha and beta power spectrum vectors which were denoised using the reverse biorthogonal (rbio) 3.1 and rbio 3.3 wavelets. The soft thresholding method used was rigorous Stein’s unbiased risk (rigrsure). The feature selection method and classifier pairs that produce high accuracies are the ranking and relief-based feature selection methods paired with a tree classifier for detection while the ANOVA-based paired with a tree classifier, for identification.An attempt to increase the number of instances was made to further investigate the performance of the classifiers. It is noticeable that there are instance imbalances for both tasks. Since machine learning algorithms work best when the number of instances of each class are roughly equal, the SMOTE algorithm was used to balance the instances. The classification was performed in WEKA considering the five aforementioned classification algorithms. The leave-one-out cross validation (LOOCV) and the ten-fold cross validation (TFCV) methods were used as test options. For the identification task, the C segment has more intances as compared to the F and G segments. To roughly balance the instances, the SMOTE algorithm was applied once with 100% increase for both the F and G segments. The percentage of CCI obtained ranges from 45% to around 56% only for both LOOCV and TFCV. The kappa values obtained roughly range from 0.2 to 0.3. The range of data reliability is from 5% to 15% which indicates fair or minimal representation of the features to the labels considered.The five initial machine learning algorithms used might not be the best considering the data sets used in this study. To further explore other learning algorithms, this study explored the utilization of the autoWEKA package for both detection and identification tasks to possibly search for an optimized learning algorithm that could match the nature of its data sets. for both feature vector sets, the RandomForest classifier was recommended by autoWEKA to yield the highest possible CCI percentage, kappa values, highest precision and recall, and lowest optimized RMSE as compared to the other available learning algorithms found in WEKA. The percentage of CCI obtained ranges from 91% to almost 99% and the kappa values obtained range from 0.82 to roughly 0.99. The range of data reliability is from 64% to 100% which indicates strong to almost perfect representation of the features to the labels considered. Result relevancy is more than 0.9 (or 90%) as indicated by precision and truly relevant results are returned at more than 0.9 (or 90%) as indicated by the recall. High scores for both precision and recall indicates that the classifier is returning accurate results, as well as returning a majority of all positive results.This study extended the practice of digital signal processing in analyzing EEG signals in relation to musical tones. Some researches focus on the given information of a complete musical piece or song such as genre and artists. The manner on how songs were created and composed is not yet thoroughly explored. This initial study on understanding how the brain responds to musical tones can be further explored by incorporating the other elements of sounds and rudiments of a musical piece.