<p>Electroencephalography (EEG) seizure detection automation remains a challenge due to the nonlinear and nonstationary characteristics of brain signals. This paper proposes a data augmentation method that utilizes the synchronization properties of the single-transistor chaotic oscillator (STCO) and EEG signals. Unlike previous works, which were based on simulated data, this paper is the first to apply the framework to patient EEG data. Based on preprocessed EEG samples from the CHB-MIT dataset, which was recorded at a sampling rate of 256 Hz and filtered between 0.5-25 Hz, a controlled coupling between the oscillator and EEG enables the creation of additional data with nonlinear perturbations. This data augmentation improves the dimensionality of the EEG dataset, thus improving feature extraction and classification accuracy for various sub-bands of EEG signals, including delta, theta, alpha, beta, and raw signals. The patterns of synchronization are investigated by means of phase-locking values (PLV), time-frequency analysis, first return maps, and information-theoretic analysis, showing dominant frequency-selective entrainment especially in the alpha band, as well as enhanced dynamical complexity due to the chaotic circuit. A multi-layer perceptron neural network with a single hidden layer, implemented by means of patternnet with 10 neurons, trained for a maximum of 50 epochs using leave-one-out cross-validation, is used. Normalized feature vectors including wavelet-based energy, entropy, mean, variance, and standard deviation features extracted from band-filtered EEG signals and their corresponding chaotic circuit responses are used as input to the classifier to distinguish between seizure-related brain states. The results show that the use of the augmented EEG dataset increases the accuracy and F1-scores for a variety of binary and multi-class classification problems. For instance, numerical computations show that the accuracy for pre-ictal vs. ictal classification is approximately <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$79.06\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>79.06</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> without augmentation, while it is <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$85.11\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>85.11</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> with augmentation. Similarly, accuracy for post-ictal vs. ictal classification is increased from <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$75.00\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>75.00</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> to <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$86.51\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>86.51</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> with augmentation. Experiments on analog hardware also further validate that the proposed framework leads to a substantial improvement in performance over the baseline classification approach, with even more improvements than those seen in the numerical simulations. For example, the experimental results show that the accuracy for pre-ictal vs. ictal classification without augmentation is <InlineEquation ID="IEq5"> <EquationSource Format="TEX">$97.67\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>97.67</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>, which increases to <InlineEquation ID="IEq6"> <EquationSource Format="TEX">$100\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>100</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> with augmentation. Similarly, the accuracy for post-ictal vs. ictal classification is improved from <InlineEquation ID="IEq7"> <EquationSource Format="TEX">$83.72\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>83.72</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> to <InlineEquation ID="IEq8"> <EquationSource Format="TEX">$100\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>100</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> with augmentation. Such improvements are also seen in multiclass classification and raw EEG classification problems. In the case of the 8 Hz circuit used for improving feature extraction, the alpha and raw frequency bands were most affected by the augmentation method in both numerical and experimental simulations. These results provide proof of concept and reiterate the usefulness of chaotic oscillators as a valuable, analog, and efficient tool for physiological signal augmentation and classification, and mark this work as the first application of the technique to epilepsy patient data, providing a physically grounded, hardware-efficient analog data augmentation approach conceptually related to reservoir computing.</p>

Enhancing epileptic seizure detection via augmentation of electroencephalography (EEG) signals using a single-transistor chaotic oscillator

期刊论文
发表日期：

2026年7月30日

查看全文

Zeric Njitacke Tabekoueng,
Longxiang Fu,
Yi Zhang,
Yuri Antonacci,
Jules Fossi Tagne,
Paul Didier Kamdem Kuate,
Manyu Zhao,
Pedro A. Valdes-Sosa,
Natsue Yoshimura,
Ferruccio Panzica,
Mattia Frasca,
Ludovico Minati

摘要

Electroencephalography (EEG) seizure detection automation remains a challenge due to the nonlinear and nonstationary characteristics of brain signals. This paper proposes a data augmentation method that utilizes the synchronization properties of the single-transistor chaotic oscillator (STCO) and EEG signals. Unlike previous works, which were based on simulated data, this paper is the first to apply the framework to patient EEG data. Based on preprocessed EEG samples from the CHB-MIT dataset, which was recorded at a sampling rate of 256 Hz and filtered between 0.5-25 Hz, a controlled coupling between the oscillator and EEG enables the creation of additional data with nonlinear perturbations. This data augmentation improves the dimensionality of the EEG dataset, thus improving feature extraction and classification accuracy for various sub-bands of EEG signals, including delta, theta, alpha, beta, and raw signals. The patterns of synchronization are investigated by means of phase-locking values (PLV), time-frequency analysis, first return maps, and information-theoretic analysis, showing dominant frequency-selective entrainment especially in the alpha band, as well as enhanced dynamical complexity due to the chaotic circuit. A multi-layer perceptron neural network with a single hidden layer, implemented by means of patternnet with 10 neurons, trained for a maximum of 50 epochs using leave-one-out cross-validation, is used. Normalized feature vectors including wavelet-based energy, entropy, mean, variance, and standard deviation features extracted from band-filtered EEG signals and their corresponding chaotic circuit responses are used as input to the classifier to distinguish between seizure-related brain states. The results show that the use of the augmented EEG dataset increases the accuracy and F1-scores for a variety of binary and multi-class classification problems. For instance, numerical computations show that the accuracy for pre-ictal vs. ictal classification is approximately $79.06\%$ $79.06 %$ without augmentation, while it is $85.11\%$ $85.11 %$ with augmentation. Similarly, accuracy for post-ictal vs. ictal classification is increased from $75.00\%$ $75.00 %$ to $86.51\%$ $86.51 %$ with augmentation. Experiments on analog hardware also further validate that the proposed framework leads to a substantial improvement in performance over the baseline classification approach, with even more improvements than those seen in the numerical simulations. For example, the experimental results show that the accuracy for pre-ictal vs. ictal classification without augmentation is $97.67\%$ $97.67 %$ , which increases to $100\%$ $100 %$ with augmentation. Similarly, the accuracy for post-ictal vs. ictal classification is improved from $83.72\%$ $83.72 %$ to $100\%$ $100 %$ with augmentation. Such improvements are also seen in multiclass classification and raw EEG classification problems. In the case of the 8 Hz circuit used for improving feature extraction, the alpha and raw frequency bands were most affected by the augmentation method in both numerical and experimental simulations. These results provide proof of concept and reiterate the usefulness of chaotic oscillators as a valuable, analog, and efficient tool for physiological signal augmentation and classification, and mark this work as the first application of the technique to epilepsy patient data, providing a physically grounded, hardware-efficient analog data augmentation approach conceptually related to reservoir computing.