A spiking neural network (SNN) for detecting high frequency oscillations (HFOs) in the intraoperative ECoG

Figure 1. HFO detection scheme. (a) The wideband ECoG is filtered in the HFO frequency band (250–500 Hz). (b) In the baseline detection stage, the background noise of the signal is used to set the signal-to-spike threshold. (c) The signal-to-spike conversion algorithm converts the analog signal into two streams of digital outputs: UP and DN spikes. (d) The SNN architecture for HFO detection and artifact rejection consists of input neurons (grey) receiving the input UP-DN spikes. These inputs project to a second layer of neurons (green) and to a dis-inhibitory neuron (purple). This neuron projects inhibitory synapses to a global-inhibitory neuron (orange), which is continuously inhibiting the second layer neurons. The synapses of the projections are excitatory (positive, red) or inhibitory (negative, blue).

Figure 2. The SNN distinguishes between an HFO and an artifact (fast transient) in the ECoG. (a,c) ECoG signal and input spike train to the SNN. Consecutive UP-DN spike bursts in the train define an UP-DN cycle. More cycles occurred during the HFO. (b,d) Activity of the neurons in the SNN. (b) During an HFO, the input spike train excited the dis-inhibitory neuron (membrane potential \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_{mem}$$\end{document}Imem in purple), which suppressed the global-inhibitory neuron that became silent (input current \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_{in}$$\end{document}Iin and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt}…

Figure 3. Intraoperative ECoG recording and HFO detection in Patient 6. (a) Position of the high-density grid electrode before resection. (b) Example of an HFO in the pre-resection ECoG, wideband and filtered in HFO band (250–500 Hz), and the time-frequency spectrum. (c) Firing of SNN neurons indicate the occurrence of an HFO. (d) Pre-resection HFO rates for each recording channel. (e) Electrode position after resection. (f) Example of an HFO in the post-resection ECoG. (g) Firing of SNN neurons indicate the occurrence of an HFO. (h) Post-resection HFO rates for each recording channel. In 4 channels (28–29, 29–30, 30–31, 1–2), the HFO rate exceeded 1 HFO/min. The occurrence of these residual HFOs “predict” poor seizure outcome. Indeed, the patient suffered from recurrent seizures (ILAE 3).