Classification for EEG report generation and epilepsy detection

Author(s)

Oliva, Jefferson Tales ^[1] ; Garcia Rosa, Joao Luis ^[1]

Affiliation

^[1] Univ Sao Paulo, Bioinspired Comp Lab, Grad Program Comp Sci & Computat Math, Ave Trabalhador Sao Carlense 400, BR-13560970 Sao Carlos, SP - Brazil

• Document type: Journal article
• Source: Neurocomputing; v. 335, p. 81-95, MAR 28 2019.
• Web of Science Citations: 2

Abstract

An automatic generation of medical report method (AGMedRep) was proposed in order to process electroencephalogram (EEG) segments using machine learning (ML) to generate textual reports for epilepsy detection. This method is applied in two phases: (1) predictive model building, and (2) automatic generation of medical reports. In Phase 1, classifiers are built and evaluated, in which the best predictive model is chosen to perform Phase 2 for report generation. The AGMedRep was applied in a set of 500 EEG segments, 100 for each class, in a total of five classes. In the first phase, 90 signal segments for each class were selected for feature extraction and classifier building. Posteriorly, five different report expression were defined, one for each class. In each EEG segment, all possible combination of cross-correlogram, power spectrum, spectrogram (SG), and bispectrogram (BS) were computed for feature extraction, generating a total of 15 different datasets. Afterwards, ML methods, such as decision tree, 1-nearest-neighbor (1NN), naive Bayes, backpropagation based on multilayer perceptron (MLP), and support vector machine (SVM) were applied in each dataset considering a multiclass decomposition approach, generating a total of 75 (15 datasets x 5 ML techniques) classifiers. In the evaluation step, it was found that the ``SG{''} 1NN, ``SG and BS{''} MLP, and ``SG and BS{''} SVM reached, statistically, the best performances for signal classification, but a statistical difference was not found among them. In this sense, the ``SG{''} 1NN model was chosen for report generation due to its lower computation cost, and it proved to be one of the most accurate models. In the second phase, the 50 remaining EEG segments were equally distributed, randomly, into ten folders in order to simulate individual EEG exams. Finally, the chosen classifier was computed in each examination to construct textual reports. As a result, the 1NN model reached an average of 84% accuracy in the report generation. Our results suggest that the AGMedRep can be able to assist medical experts in the identification of patterns related to epileptic events in EEG. (C) 2019 Elsevier B.V. All rights reserved. (AU)

• FAPESP's process: 16/02555-8 - Development of algorithms and computational techniques for application in brain-computer interfaces
• Grantee: João Luís Garcia Rosa
• Support Opportunities: Regular Research Grants