P023 Applying a Machine Learning Method to Detect Miniature EPSCs
Krishan Bhalsod*1, Cengiz Günay1
1Dept. Information Technology, Georgia Gwinnett College, Lawrenceville, Georgia, USA
*kbhalsod@ggc.edu Introduction
This study aims to use machine learning to find miniature excitatory postsynaptic currents (EPSCs) in neurons of aDrosophilato find behavior markers of a seizure. Using MATLAB, we are training a machine learning model on electrophysiological data to recognize patterns of postsynaptic events that show potential seizure activity. We have faced challenges applying this method and we are planning to present these in our poster. The results of this research may help develop a further understanding of seizure mechanisms inDrosophilathat could translate into a more in-depth understanding for neurological disorders in humans.
Methods The particular type of data we are addressing is obtained from intracellular recordings of invertebrate neurons, specifically fromDrosophila(fruit fly) motor neurons [1]. Not only these recordings have low SNR, but also the miniature excitatory postsynaptic current (EPSC) (or ”mini”) event we are looking for come in various magnitudes due to the distance from the event’s origin on the neuron’s morphology. In the present work, our aim is to adapt a novel machine learning and optimal filtering method (MOD) to automatically detect these minis [2]. Results The purpose of MOD is to generate a filter that takes the original data and it removes any noise, turning it into a raw detection trace that closely mirrors the manual scoring trace. The method leverages the Wiener-Hopf equations to derive an optimal filter for detecting post-synaptic events. In the MATLAB code, the optimal-filter equations are directly implemented. First, the program estimates the auto-correlation and cross-correlation from the training data to build a Toeplitz matrix Ry, and then it solves for the filter coefficients a. To correct for any timing differences between the recorded signal and the manual scoring, the algorithm computes filter coefficients for several time shifts and selects the delay that yields the best detection performance (e.g., highest AUC). Finally, a low-pass Hann window filter is applied to smooth the detection trace. Discussion The challenges of machine learning come with filtering noise. Typically, in electrophysiological recordings, the lines are not smooth. Therefore, we applied a bandpass filter of 1-1,000 Hz to reduce the noise. However, we face a problem where the signal oscillates and eventually forms flat lines, which is most likely caused by the filtering algorithm removing low-magnitude events. Because of this, the machine learning model faces difficulties learning from the filtered signal, thus failing to recognize events because the threshold is no longer high enough to flag the event.
Figure 1. Example of recording where blue shaded areas highlight mini events. Time units in seconds. Y-axis units in pA. Acknowledgements The recordings used in this work were provided by Richard Baines from University of Manchester. We are grateful for providing student travel funding to Dr. Joseph Ametepe, Dean of the School of Science and Technology, and Dr. Sean Yang, Chair of the Information Technology Department at Georgia Gwinnett College. Students Jonathan Tran and Niecia Say provided valuable feedback for this project. References 1. C. N. G. Giachello and R. A. Baines. Inappropriate neural activity during a sensitive period in embryogenesis results in persistent seizure-like behavior. Curr Biol, 25(22):2964–2968, Nov 2015. doi: 10.1016/j.cub.2015.09.040. 2. X. Zhang, A. Schlögl, D. Vandael, and P. Jonas. MOD: A novel machine-learning optimal-filtering method for accurate and efficient detection of subthreshold synaptic events in vivo. Journal of Neuroscience Methods, 357:109125, 2021. ISSN 0165-0270. doi: https://doi.org/10.1016/j.jneumeth.2021.109125.
