Neural network capable of segmenting patterns of epileptiform activity in EEG recordings developed by UNN scientists
The Spike-Wave Index (SWI) is an important quantitative biomarker in epilepsy. It represents the percentage of time during an EEG recording when epileptiform activity patterns are observed, serving as an objective measure of epileptiform load. Traditional methods for calculating SWI involve manually counting spikes and sharp waves over short EEG intervals. Automated analysis of EEG recordings will enable the development of personalised treatment strategies and the objective monitoring of therapy effectiveness.
According to Albina Lebedeva, a senior researcher at the Research Centre for Artificial Intelligence at the UNN Institute of Information Technology, Mathematics and Mechanics (IITMM), manual labelling of EEG by doctors is extremely time-consuming, and visual interpretation is prone to significant variability. Even highly qualified experts often disagree when evaluating the same recordings, especially with ambiguous patterns.
"The neural network does get tired, does not lose focus, and is not prone to subjective interpretations. It produces a stable, mathematically reproducible result that corresponds to the consensus of the expert group on average. The neural network acts as a 'tireless and impartial second expert,'labelling data with an accuracy comparable to the average opinion of several doctors, but doing so instantly and consistently under any conditions,"notes Albina Lebedeva.
A large team of authors participated in the study, including epileptologists from universities in Moscow and Nizhny Novgorod. Lobachevsky University experts have been developing and adapting the architecture of a deep neural network (1D U-Net) for segmenting one-dimensional time series (EEG), training the model, optimising hyperparameters, and comprehensively comparing the results of automatic segmentation with medical experts' assessments.
The 1D U-Net architecture was chosen for the neural network, an adaptation of the classical U-Net for one-dimensional sequential data. This architecture provides competitive overall performance and allows for obtaining SWI estimates calculated using segmentation masks that are in good agreement with expert estimates.
Tatiana Levanova, a senior researcher at the Artificial Intelligence in Cardiology and Neuroscience Laboratory at the UNN IITMM, explains: "The fundamental novelty of our work lies in creating and annotating our own EEG database with the involvement of several independent experts to form a reliable 'gold standard.' We have implemented an approach based on precise segmentation – the precise definition of pattern boundaries over time, rather than simple binary detection of activity presence. This allows for the precise quantification of epileptiform load. We have also conducted a comprehensive comparison of automatic annotations with expert assessments, including direct calculation and comparison of the spike-wave index."
The dataset used for the study consisted of 51 EEG recordings obtained from 45 subjects aged 3 to 13 years. It was divided into training and test subgroups at the individual subject level. The test subgroup included 10 recordings from 10 subjects and was used to obtain and compare expert SWI scores. The remaining 41 recordings were used to train the segmentation model.
Five functional diagnostic experts with at least five years of experience in EEG interpretation independently evaluated the selected 10-minute segments and assessed the SWI. The experts did not know the results of the neural network or each other's assessments, nor the clinical diagnosis of the patients. Four months later, the same records were presented again in mixed order to the same experts. The obtained SWI estimates were compared to determine inter-expert variability. The agreed estimates, defined as the average value among all experts, were compared with the results of the AI model.
"These results confirm that a properly trained deep learning model can achieve a level of consistency with annotations that matches and even surpasses the consistency seen among human experts. A properly trained deep neural network can provide reproducible segmentation of epileptiform activity and high-performance SWI assessment: event-based F1 measures exceed 0.9 in half of the patients in our test cohort. The automatic calculation of SWI strongly correlates with both the reference data and the average opinion of doctors, confirming its suitability as a quantitative biomarker. Thus, we have developed a tool for epileptologists that can serve as an objective, scalable alternative to subjective manual annotation," says Nikolai Gromov, a junior researcher at the Artificial Intelligence in Cardiology and Neuroscience Laboratory at Lobachevsky University’s IITMM.
"The synthesis of competencies in the field of artificial intelligence and clinical medicine allows Lobachevsky University to create high-demand technological solutions. The EEG analysis algorithm developed by our scientists will help ensure high diagnostic accuracy and open up new horizons for personalized medical care,"says Oleg Trofimov, Rector of Lobachevsky University.
The research was performed with the support of the Ministry of Economic Development of the Russian Federation and the results were published in the MDPI Technologies journal.
