TWO PHASE SPIKE DETECTION USING DEEP LEARNING
Automatic identification of the single unit activities is a major part in the analysis of the electrophysiological data recorded from within the central nervous system. Despite multiple unsupervised methods were proposed to detect and sort neural activity, they require hyperparameter tuning for every individual recording.The goal was to build a robust, reliable detector which is unaffected by any parameters of the region the recording was effectuated from. Our proposal is a new detection system, which utilizes deep learning tools to induce generalization. The proposed detection system consists of a pre-detector and a main detector. The electrophysiological data is filtered and thresholded with the pre-detector, ensuring that the information arriving to the main detector has a higher probability being a positive sample. The pre-detector system is built with low computational cost and high operating frequency in mind, while the main detector with moderate computational cost and operating frequency, keeping the option of a future real-time detector open. To evaluate the performance of our model we used 5 different recordings and cross-validated them having at every step training and validation samples from 4 datasets, while the remaining one serving as the test dataset. For evaluating the performance we used the recall, precision and accuracy metrics. We cross-validated our model on small epoch size and selected the best performing configuration which was trained for a longer time, with the following results: 89.59%, 61,40% recall, 88,77%, 50,32% precision and 95%, 81% accuracy for validation and test datasets respectively.