In a paper published in the journal Nature Communications, researchers developed a modular spiking neural network (SNN) on a mixed-signal neuromorphic device to process intraoperative electrocorticography (ECoG) in real time. This system efficiently encoded signals with sparsity and decorrelation by utilizing the variability of silicon neurons.
Comparison of DYNAP-SE SNN and Spectrum9 detectors in Patient 5. The rates obtained with the DYNAP-SE SNN for HFO (black) and IED-HFO (orange) were compared with the HFO rates of the Spectrum detector (gray bars). Most of the channels with high rates were the same for both detectors. In this patient, all the channels with high HFO and IED-HFO rates were recorded from tissue that was later resected (red labels). The patient achieved seizure freedom after surgery and was classified as a true negative (TN). This supports our hypothesis that the detected epileptiform patterns indicate the epileptogenic zone. HFO high-frequency oscillation, IED interictal epileptiform discharge. https://www.nature.com/articles/s41467-024-47495-y
The SNN accurately identified interictal epileptiform discharges (IED) and high-frequency oscillations (HFO), particularly focusing on IED-HFO co-occurrences by integrating them into the BCI2000 framework. Validation with pre-recorded data demonstrated consistent HFO rates comparable to offline algorithms. Additionally, a remote analysis successfully compressed ECoG data recorded in Utrecht. It detected IED-HFO patterns in real-time in Zurich, highlighting the potential for automated remote detection in clinical settings.
Related Work
Past work has focused on utilizing intraoperative ECoG to delineate the epileptogenic zone in patients with pharmaco-resistant focal epilepsy. While IEDs have traditionally guided surgical decisions, HFOs have emerged as a more precise epileptiform pattern. Co-occurring IED and HFO (IED-HFO) and fast ripple HFO have shown promise in indicating epileptogenic tissue with high specificity.
Current approaches involve visual annotation by experts or offline application of automated software algorithms, limiting intraoperative HFO analysis to specialized centers. SNNs have shown potential for bio-signal analysis, including HFO detection, but have yet to be widely applied intraoperatively.
Study Overview: Epilepsy Surgery Details
The study included 23 patients with drug-resistant epilepsy, with a median age of 17 years (range: 1–67), and 12 of them were female.
Twenty-two patients underwent epilepsy surgery guided by intraoperative high-density ECoG (hd-ECoG) or standard grid and strip electrodes, and one patient underwent surgery guided by an intraoperative strip electrode. Patient data collection and analysis were conducted in compliance with local research ethics committees' guidelines and regulations, with informed consent obtained from all patients and their parents for investigation use of clinical data.
Anesthesia protocols varied slightly between University Hospital Zurich (USZ) and University Medical Center Utrecht (UMCU). At USZ, anesthesia commenced with the intravenous administration of Propofol and Fentanyl for induction, followed by maintenance with Propofol and Remifentanil.
At UMCU, total intravenous anesthesia (TIVA) was utilized. Propofol was used for induction and maintenance, combined with analgesics and muscle relaxants. The recording setup at USZ involved hd-ECoG recorded using high-density subdural grid electrodes, while standard ECoG electrodes were used in other cases. A needle electrode in the dura served as an electrical reference. The team recorded ECoG data for offline processing, with all data re-referenced to a bipolar montage. At UMCU, ECoG was recorded using a subdural strip electrode with Micromed equipment.
Patients were categorized as true positive (TP), false positive (FP), false negative (FN), or true negative (TN) based on post-resection ECoG results and seizure freedom after surgery in the clinical validation. Positive predictive value (PPV), negative predictive value (NPV), sensitivity, specificity, and accuracy were calculated using standard methods. Statistical analysis was performed using the Scipy Python package, with significance at p < 0.05. Confidence intervals were estimated using the Clopper-Pearson method, while Spearman's was employed for correlation analysis.
Real-Time ECoG Analysis
The study begins by detailing the setup designed for real-time online data analysis, which involves using a dynamic, asynchronous neuromorphic processor (DYNAP-SE) to process intraoperative ECoG data. Specifically, the researchers focused on detecting HFO and IED-HFO patterns within specific frequency bands, describing each step of the processing pipeline in detail.
Signal preprocessing shows the validation of real-time scenarios using pre-recorded data and the streaming of intraoperative data. Filtering procedures and data buffering methods are explained by emphasizing the importance of employing a high filter order for HFO detection.
The process of Delta-modulation encoding (ADM) follows, providing insight into how ECoG data is converted into discrete digital pulses to highlight epileptiform patterns as events of interest (EoI). The description outlines parameters that govern the asynchronous delta modulator (ADM) operation, including threshold levels and refractory periods, along with specific details regarding the tuning phase.
Subsequent paragraphs delve into SNN processing, elucidating how discrete UP/DN pulses are analyzed using a hardware SNN implemented on the DYNAP-SE. The analysts discuss the variability induced by circuit device mismatch factors and the benefits of employing an evolutionary algorithm for training the SNN.
Finally, the detection phase presents simple rules devised to detect epileptiform patterns from SNN activity. Detailed criteria for identifying HFO and IED-HFO patterns and examples of event-based processing are provided. This segment culminates with discussions on ECoG compression, reconstruction, and real-time analysis, including insights from testing and correlation analyses.
Conclusion
To sum up, the study provided a detailed framework for real-time analysis of intraoperative ECoG data, focusing on detecting HFO and IED-HFO. It described signal preprocessing, delta-modulation encoding, SNN processing, and detection algorithms. The findings emphasized parameter optimization, hardware implementation, and algorithm refinement for accurate epileptic activity detection. Additionally, discussions on ECoG compression, reconstruction, and real-time analysis highlighted the potential of neuromorphic computing for improving clinical ECoG monitoring in epilepsy surgery. Overall, the study contributed to advancing real-time neurophysiological monitoring and personalized treatment strategies for epilepsy patients.
AlphaQubit Transforms Quantum Error Correction with Cutting-Edge AI