Time–frequency domain machine learning for detection of epilepsy using wearable EEG sensor signals recorded during physical activities

Baghersalimi S., Teijeiro T., Aminifar A., Atienza D.

Ingolfsson T.M., Benatti S., Wang X., Bernini A., Ducouret P., Ryvlin P., Beniczky S., Benini L., Cossettini A.

AbstractElectroencephalography (EEG) is widely used to monitor epileptic seizures, and standard clinical practice consists of monitoring patients in dedicated epilepsy monitoring units via video surveillance and cumbersome EEG caps. Such a setting is not compatible with long-term tracking under typical living conditions, thereby motivating the development of unobtrusive wearable solutions. However, wearable EEG devices present the challenges of fewer channels, restricted computational capabilities, and lower signal-to-noise ratio. Moreover, artifacts presenting morphological similarities to seizures act as major noise sources and can be misinterpreted as seizures. This paper presents a combined seizure and artifacts detection framework targeting wearable EEG devices based on Gradient Boosted Trees. The seizure detector achieves nearly zero false alarms with average sensitivity values of $$65.27\%$$ 65.27 % for 182 seizures from the CHB-MIT dataset and $$57.26\%$$ 57.26 % for 25 seizures from the private dataset with no preliminary artifact detection or removal. The artifact detector achieves a state-of-the-art accuracy of $$93.95\%$$ 93.95 % (on the TUH-EEG Artifact Corpus dataset). Integrating artifact and seizure detection significantly reduces false alarms—up to $$96\%$$ 96 % compared to standalone seizure detection. Optimized for a Parallel Ultra-Low Power platform, these algorithms enable extended monitoring with a battery lifespan reaching 300 h. These findings highlight the benefits of integrating artifact detection in wearable epilepsy monitoring devices to limit the number of false positives.

Yedurkar D.P., Metkar S., Al-Turjman F., Yardi N., Stephan T.

Cao W., Liu Y., Mei H., Shang H., Yu Y.

Distributed generation and diversified loads increase the uncertainty of district power prediction. Useful prediction requires a highly accurate model, and there are several challenges facing the designers of a new power system with intelligent power distribution. To solve them, we improved an XGBoost model from three aspects: model, data, and performance. This paper proposes an XGBoost model with a windowed mechanism and random grid search (WR-XGBoost model) for self-prediction of short-term district power load. Specifically, a causal sliding window with different strides is introduced into the model optimization stage to process the training and test sets separately. In data optimization, the model initially processes the data organized in forms and then uses discrete difference data as input. Finally, in optimizing the performance, a random grid search method reduces the debugging of hyperparameters. The results show that the WR-XGBoost model outperforms five comparison models in terms of predictive power and generalization, using four datasets and seven statistical indicators.

Gade A., Dash D.K., Kumari T.M., Ghosh S.K., Tripathy R.K., Pachori R.B.

Salafian B., Ben-Knaan E.F., Shlezinger N., De Ribaupierre S., Farsad N.

Hu S., Liu J., Yang R., Wang Y., Wang A., Li K., Liu W., Yang C.

Li C., Huang X., Song R., Qian R., Liu X., Chen X.

Recently, most seizure prediction methods mainly utilize pure CNN or Transformer model, which cannot extract local and global features simultaneously. To this end, we propose an Electroencephalogram (EEG) seizure prediction method based on Transformer guided CNN (TGCNN), which combines the complementary advantages of CNN and Transformer. The proposed method first use short-time Fourier transform (STFT) to extract time–frequency features from EEG signals. Then, these features are fed into the alternating structure to model both local feature and long-distance dependencies, which can overcome both the deficiency of long distance dependence in CNN and the lack of local features in Transformer. Finally, the prediction result is obtained through a global average pooling layer and fully connected layer. The proposed method achieves sensitivity of 91.5%, false prediction rate (FPR) of 0.145/h, and area under curve (AUC) of 93.5% on CHB-MIT database and 82.2% sensitivity, 0.06/h FPR, and 83.5% AUC on Kaggle dataset. • We propose an EEG seizure prediction method based on Transformer guided CNN. • Our method can extract local and global features for EEG seizure prediction. • We propose a new Transformer module for EEG seizure prediction.

Busia P., Cossettini A., Ingolfsson T.M., Benatti S., Burrello A., Scherer M., Scrugli M.A., Meloni P., Benini L.

Hussein R., Lee S., Ward R.

Epilepsy is a neurological disorder that causes recurrent seizures and sometimes loss of awareness. Around 30% of epileptic patients continue to have seizures despite taking anti-seizure medication. The ability to predict the future occurrence of seizures would enable the patients to take precautions against probable injuries and administer timely treatment to abort or control impending seizures. In this study, we introduce a Transformer-based approach called Multi-channel Vision Transformer (MViT) for automated and simultaneous learning of the spatio-temporal-spectral features in multi-channel EEG data. Continuous wavelet transform, a simple yet efficient pre-processing approach, is first used for turning the time-series EEG signals into image-like time-frequency representations named Scalograms. Each scalogram is split into a sequence of fixed-size non-overlapping patches, which are then fed as inputs to the MViT for EEG classification. Extensive experiments on three benchmark EEG datasets demonstrate the superiority of the proposed MViT algorithm over the state-of-the-art seizure prediction methods, achieving an average prediction sensitivity of 99.80% for surface EEG and 90.28–91.15% for invasive EEG data.

Tripathy R.K., Dash S., Rath A., Panda G., Pachori R.B.

In this letter, a promising method is proposed to automatically detect pulmonary diseases (PDs) from lung sound (LS) signals. The modes of the LS signal are evaluated using empirical wavelet transform with fixed boundary points. The time-domain (Shannon entropy) and frequency-domain (peak amplitude and peak frequency) features have been extracted from each mode. The classifiers, such as support vector machine, random forest, extreme gradient boosting, and light gradient boosting machine (LGBM), have been chosen to detect PDs using the features of LS signals automatically. The performance of the proposed method has been evaluated using LS signals obtained from a publicly available database. The detection accuracy values, such as 80.35, 83.27, 99.34, and 77.13%, have been obtained using the LGBM classifier with fivefold cross validation for normal versus asthma, normal versus pneumonia, normal versus chronic obstructive pulmonary disease (COPD), and normal versus pneumonia versus asthma versus COPD classification schemes. For the normal versus pneumonia versus asthma classification scheme, the proposed method has achieved an accuracy value of 84.76%, which is higher than that of the existing approaches using LS signals.

Karhade J., Dash S., Ghosh S.K., Dash D.K., Tripathy R.K.

The damage to the heart valves causes heart valve disorders (HVDs). The detection of HVDs is crucial in a clinical study as these diseases may cause congestive heart failure, hypertrophy, and stroke. The phonocardiogram (PCG) signal reveals information regarding the mechanical activity of the heart. The early detection of HVDs using PCG signal is vital to minimize the chances of cardiac arrest and other cardiac complications. This article proposes the time–frequency-domain deep learning (TFDDL) framework for automatic detection of HVDs using PCG signals. The time–frequency (TF)-domain representations of PCG signals are evaluated using both time-domain polynomial chirplet transform (TDPCT) and frequency-domain polynomial chirplet transform (FDPCT). The deep convolutional neural network (CNN) model is used to detect four types of HVDs using the TF images of PCG signals obtained using both the TDPCT and FDPCT methods. The proposed TFDDL approach is evaluated using PCG signals from public databases. For the detection of HVDs using TDPCT- and FDPCT-based TF images of PCG signals, the suggested approach has achieved overall accuracy values of 99% and 99.48%, respectively. For the classification of normal and abnormal heart sound classes, the proposed TFDDL approach has obtained an accuracy of 85.16% using PCG signals from the Physionet challenge 2016 database. The proposed TFDDL framework is compared with TF-domain transfer learning models such as residual network (ResNet-50) and visual geometry group (VGGNet-16). The overall accuracy values obtained using VGGNet-16 and ResNet-50 are less than the proposed deep CNN model for the detection of HVDs. The proposed TFDDL model can be validated in real-time using heart sound signals recorded from different subjects for automated identification of HVDs.

Shashidhar R., Patilkulkarni S., Puneeth S.B.

Human speech is bimodal, whereas audio speech relates to the speaker's acoustic waveform. Lip motions are referred to as visual speech. Audiovisual Speech Recognition is one of the emerging fields of research, particularly when audio is corrupted by noise. In the proposed AVSR system, a custom dataset was designed for English Language. Mel Frequency Cepstral Coefficients technique was used for audio processing and the Long Short-Term Memory (LSTM) method for visual speech recognition. Finally, integrate the audio and visual into a single platform using a deep neural network. From the result, it was evident that the accuracy was 90% for audio speech recognition, 71% for visual speech recognition, and 91% for audiovisual speech recognition, the result was better than the existing approaches. Ultimately model was skilled at enchanting many suitable decisions while forecasting the spoken word for the dataset that was used.

Dash S., Tripathy R.K., Panda G., Pachori R.B.

In this letter, a novel automated approach for recognizing imagined commands using multichannel electroencephalogram (MEEG) signals is presented. The multivariate fast and adaptive empirical mode decomposition method decomposes the MEEG signals into various modes. The slope domain entropy and $L_1$ -norm features are obtained from the modes of MEEG signals. The machine learning models such as k -nearest neighbor, sparse representation classifier, and dictionary learning (DL) techniques are used for the imagined command classification tasks. The efficacy of the proposed approach is evaluated using MEEG from a public database as input signals. The proposed approach has achieved average accuracy values of 60.72, 59.73, and 58.78% using a DL model and selected features for left versus right, up versus down, forward versus backward based imagined command categorization tasks.

Liang C., Teng Z., Li J., Yao W., Hu S., Yang Y., He Q.

The accurate time-frequency (TF) positioning of power quality (PQ) disturbances is the basis of dealing with PQ problems in power systems. To accurately detect PQ disturbances, this article proposes a Kaiser window-based S-transform (KST) that provides better time resolution at fundamental frequency to detect the amplitude information for voltage swell, sag, interrupt, flicker, and better frequency resolution at higher frequencies to detect the frequency of time-varying harmonics and oscillatory transient. Based on short-time Fourier transform and S-transform, KST uses a Kaiser window with the characteristic of inherent optimal energy concentration as the kernel function. The Kaiser window can be adjusted adaptively according to the detection demand of PQ disturbances by the designed control function. This allows KST to easily accommodate different detection requirements at different frequencies. The utilization of Fourier transform ensures that KST can be realized quickly. The complex TF matrix is generated after a signal is transformed by KST, where the column vector is expressed as the distribution of amplitude and phase with time at a certain frequency, and the row vector represents the distribution of amplitude and phase with frequency at a certain sampling time. Experimental results demonstrate that the proposed KST significantly outperforms the state-of-the-art techniques in TF analysis of PQ signals, especially for the energy concentration and the detection of fundamental wave.