Neurosciences - Posters G
Réka Bod1,2, Ágnes Kandrács1, Kinga Tóth1, Estilla Zsófia Tóth1,2, Loránd Erőss3, Attila Bagó3, Gábor Nagy3, Dániel Fabó3, István Ulbert1,3,4 and Lucia Wittner1,3,4
1 Institute of Cognitive Neuroscience and Psychology, Research Center for Natural Sciences, Eötvös Loránd Research Network, Budapest
2 János Szentágothai Doctoral School of Neurosciences, Semmelweis University, Budapest
3 National Institute of Mental Health, Neurology and Neurosurgery, Budapest
4 Faculty of Information Technology and Bionics, Pázmány Péter Catholic University, Budapest
Introduction: Discerning various events on electrophysiological recordings may reveal a fair amount of knowledge about synchrony-generating principles. Considering the vastness of data available for this purpose, as well as the time-intensive and experience-dependent nature of the analysis workflow, application of machine learning-aided technologies is welcome for this task. Although several analogous algorithms were set up for the investigation of interictal events, none of them attempted to detect physiologically occurring hypersynchronous events.
Aims: We ventured on creating artificial neural networks that distinguish spontaneous synchronous population activity (SPA) from background with an accuracy and robustness comparable with manual analysis.
Method: Data were collected by a 24-channel laminar microelectrode from human neocortical slices inferential to patients either or not displaying epileptic signs. Manual analysis identified 53 962 SPAs, based on which 0.1 s-long epochs were generated from 3 neighboring channels where event amplitudes were the highest. Similarly long, although eventless epochs were generated from baseline activity (n= 113 588). Before feeding data in the neural networks, a proper randomization and a 70-20-10% partition of training-validation-testing datasets took place. Neural network architectures relied on 1D- and 2D-convolutional, recurrent (LSTM) and dense layers.
Results: Overall fitness of the artificial neural networks was evaluated by the following metrics: binary accuracy ([true positive nr. + true negative nr.] / total entries), precision (true positive nr./ [true positive nr. + false positive nr.]) and recall (true positive nr. / [true positive nr. + false negative nr.]), the loss function chosen was binary crossentropy. After 30 epochs of training and validation, the neural network employing 1D-convolutional layers performed on the testing dataset as follows: accuracy=0.849, precision=0.752, recall=0.793. We plan to improve performance metrics by applying scheduled learning rates.
Conclusion: By the implementation of artificial neural networks, identification of SPAs benefitted from decimated inter-observer variability and substantial time reduction during analysis. This latter feature encourages our method to be assessed on similarly recorded human in vivo data, with the promise to detect SPAs unprecedentedly in this context.
Funding: This work was supported by the Hungarian National Research Fund OTKA K137886, Hungarian Brain Research Program 3.0, FLAG-ERA VIPattract, OTKA PD143380 grants. Réka Bod is grateful for the SE 250+ Doctoral Scholarship for Excellence.