Researchers have demonstrated that deep learning models can help neurologists interpret epileptic episodes during and between seizures from relatively few scalp electroencephalography (EEG) readings.

The best of the models proved out the concept under review: an automated annotation tool needing 142 times less EEG data than human experts would need to comb through—and epilepsy patients would need to log—using digital disease diaries.

The researchers, from IBM in collaboration with Temple University and other academic centers, worked with 87 scientists and software engineers from 14 research centers around the world to develop the models.

The research team analyzed EEG data from 365 patients representing 172,000 ictal (during a seizure) incidents and 2.2 million interictal (between seizures) occurrences. Part of the analysis was conducting a crowdsourced AI challenge.

“Participants were tasked to develop an ictal/interictal classifier with high sensitivity and low false alarm rates,” the study authors explain in a report published March 18 in the Lancet journal EBioMedicine. “We built a challenge platform that prevented participants from downloading or directly accessing the data while allowing crowdsourced model development.”

Led by IBM researchers Stefan Harrer in Australia and Gustavo Stolovitzky in the U.S., the team found their novel automated seizure detector returned sensitivities of up to 91.6%.

“This study demonstrates the utility of deep learning for patient-specific seizure detection in EEG data,” the authors comment. “Furthermore, deep learning in combination with a human reviewer can provide the basis for an assistive data labelling system lowering the time of manual review while maintaining human expert annotation performance.”

The post Deep learning detects, annotates epileptic seizures on scant EEG data appeared first on Artificial Intelligence.

Recent advances in machine learning have enabled the development of techniques to detect and recognize human emotions. Some of these techniques work by analyzing electroencephalography (EEG) signals, which are essentially recordings of the electrical activity of the brain collected from a person’s scalp.

Most EEG-based emotion classification methods introduced over the past decade or so employ traditional machine learning (ML) techniques such as support vector machine (SVM) models, as these models require fewer training samples and there is still a lack of large-scale EEG datasets. Recently, however, researchers have compiled and released several new datasets containing EEG brain recordings.

The release of these datasets opens up exciting new possibilities for EEG-based emotion recognition, as they could be used to train deep-learning models that achieve better performance than traditional ML techniques. Unfortunately, however, the low resolution of EEG signals contained in these datasets could make training deep-learning models rather difficult.

“Low-resolution problems remain an issue for EEG-based emotion classification,” Sunhee Hwang, one of the researchers who carried out the study, told TechXplore. “We have come up with an idea to solve this problem, which involves generating high-resolution EEG images.”

To enhance the resolution of available EEG data, Hwang and her colleagues first generated so-called “topology-preserving differential entropy features” using the electrode coordinates at the time when the data was collected. Subsequently, they developed a convolutional neural network (CNN) and trained it on the updated data, teaching it to estimate three general classes of emotions (i.e., positive, neutral and negative).

“Prior methods tend to ignore the topology information of EEG features, but our method enhances the EEG representation by learning the generated high-resolution EEG images,” Hwang said. “Our method re-clusters the EEG features via the proposed CNN, which enables the effect of clustering to achieve a better representation.”

The researchers trained and evaluated their approach on the SEED dataset, which contains 62-channel EEG signals. They found that their method could classify emotions with a remarkable average accuracy of 90.41 percent, outperforming other machine-learning techniques for EEG-based emotion recognition.

“If the EEG signals are recorded from different emotional clips, the original DE features cannot be clustered,” Hwang added. “We also applied our method on the task of estimating a driver’s vigilance to show its off-the-shelf availability.”

In the future, the method proposed by Hwang and her colleagues could inform the development of new EEG-based emotion recognition tools, as it introduces a viable solution for overcoming the issues associated with the low-resolution of EEG data. The same approach could also be applied to other deep-learning models for the analysis of EEG data, even those designed for something other than classifying human emotions.

“For computer vision tasks, large-scale datasets enabled the huge success of deep-learning models for image classification, some of which have reached beyond human performance,” Hwang said. “Also, complex data preprocessing is no longer necessary. In our future work, we hope to generate large-scale EEG datasets using a generated adversarial network (GAN).”

The post A new deep learning model for EEG-based emotion recognition appeared first on Artificial Intelligence.