• machine learning
• epilepsy
• deep learning
• zebrafish

Zebrafish (Danio rerio) is emerging as an important model to understand the cellular basis of epilepsy and cognitive impairment and for treatment screening. This has led to the development of new methods for the identification, quantification and classification of epileptiform events independently from the classic visual scoring of activity. In our laboratory we exploited the availability of a non-pigmented zebrafish expressing the Ca2+ sensor GCaMP6f in neurons to demonstrate the quantitative evaluation of activity and recruitment of neuronal populations by combining local field potential recordings (LFP) with simultaneous two-photon calcium imaging.
We have implemented an automated data pipeline which, through machine learning methods, has allowed us to explore topology and electrophysiological features of seizures, obtaining confirmations and opening up to new questions in the field of neuroscience. We offered new insights to identify characterizing Ca2+ maps through which we have been able to propose a classification of the neurological events as interictal, pre(post)-ictal and ictal, inspired by the clinical classification. Furthermore we have
also been capable to identify the interesting pathophysiological phenomena of synchronous hemispheric activity.
By cross-analysis of electrophysiology and imaging, we tested the possibility to identify the sources of the LFP changes and to dissect the spatiotemporal dynamics of epileptiform activity using the classical electrophysiological features. The partiality of the results obtained and the fact that recently, because of their visual recognition ability, convolutional neural networks are gaining popularity in replacing or complementing the human operator in the interpretation of electrophysiological traces (EEG, LFP, etc..), led us to develop an innovative method which uses calcium imaging to train an electrophysiology-based algorithm for automated seizure detection built on a monodimensional multilayer convolutional neural network (1D-CNN). This deep learning system showed immediately a good
features-extraction ability, easy scaling and therefore the possibility of being a valid instrument both in the cross-over study of calcium imaging and electrophysiology as much as only for the latter once pre-trained.
The accuracy with which the 1D-CNN was able to associate electrophysiology and calcium maps corroborate the neuroscientific thesis according to which different zones of the central nervous system "speak" with a recognizable, characteristic voice (Buzsáki G, Anastassiou CA, Koch C, 2012).