SPIKE DETECTION AND SORTING WITH DEEP LEARNING
01/30/2020
Melinda Rácz1, 2, Csaba Liber1, Erik Németh2, Richárd Fiáth2, 3, János Rokai2, 5, István Harmati1, István Ulbert2, 3, Gergely Márton2, 3, 4
1 Department of Electrical Engineering and Informatics, Budapest University of Technology and Economics, Budapest, Hungary
2 Faculty of Information Technology and Bionics, Pázmány Péter Catholic University, Budapest, Hungary
3 Institute of Cognitive Neuroscience and Psychology, Research Centre for Natural Sciences, Hungarian Academy of Sciences, Budapest,
4 Doctoral School on Materials Sciences and Technologies, Óbuda University, Budapest, Hungary
5 Károly Rácz School of PhD Studies, Semmelweis University, Budapest, Hungary
The extraction and identification of single-unit activities in intracortically recorded electric signals have a key role in basic and applied neuroscience. Besides their importance, these problems are complex and lack universal or definite solutions. Our main objective was the implementation of functional units that ameliorate the efficiency and robustness of invasive brain–computer interfaces. Since machine learning solutions (neural networks, in particular) have been applied successfully for object detection/discrimination and time series prediction problems, we assumed that they can be used for neural data processing, as well. We completed the tasks addressed based on multichannel action potential recordings (n = 9 high-density cortical recordings from anesthetized rats, each containing 23–46 well-isolated single units) using deep learning methods: Convolutional neural networks were applied for sorting and predicting spiking activities. Sorting was based on 2-D and 3-D samples of neuronal activities; as for prediction, only 3-D samples were involved. A combination of recurrent (applying long short-term memory cells) and convolutional neural network was employed for spike detection. In this case, series of multiple datapoints were used. Our algorithms proved to be useful in accomplishing the tasks considered: our detector reached an average recall of 69 %; as for the classification problem, we achieved an average accuracy of 89 %. Our findings support the concept of creating real-time, high-accuracy action potential based brain–computer interfaces in the future, providing a flexible and robust algorithmic background for further development.