Poster 5 - A long short-term memory neural network to identify surgical targets in human epilepsy

Saturday, October 26, 2024

10:00 AM - 12:00 PM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

MC

Mira Chu-Shore (she/her/hers)

Student
Boston University Academy
Cambridge, Massachusetts, United States

Co-Author(s)

WS

Wen Shi, PhD

Instructor
Harvard Medical School / Massachusetts General Hospital, United States
XH

Xue Han

Principle Investigator
Boston University, United States
CC

Catherine Chu, MD

Associate Professor, Neurologist
Harvard Medical School / Massachusetts General Hospital, United States

Primary Investigator(s)

MK

Mark A. Kramer, PhD

Professor
Boston University, United States

Introduction: Epilepsy is the most common neurological disorder, characterized by recurring, unprovoked seizures. One third of patients have seizures despite maximal medical therapy, and for these patients, neurosurgery is the most effective treatment. The success of neurosurgical interventions requires accurate identification of the epileptogenic zone (EZ), the brain region responsible for generating seizures. Spike ripples have recently emerged as the most reliable electrographic biomarker for the EZ [1]. The current validated method to detect spike ripples in intracranial electroencephalogram (EEG) data combines a feature-based algorithm and a convolutional neural network detector [1]. Although this detector performs well to detect spike ripples (F1 = 0.68, specificity = 0.42, positive predictive value (PPV) = 0.79), it is computationally inefficient and therefore impractical to use on multiday and multichannel intracranial EEG recordings. Long short-term memory (LSTM) networks are a type of recurrent neural network specifically designed to process and predict data sequences over time and are thus well-suited to EEG [2]. LSTM networks have previously been used to detect epileptic spikes and spike ripples in extracranial EEG. [3] We hypothesized that the LSTM network would have similar accuracy but improved computational efficiency to detect spike ripples in human intracranial EEG compared to the current approach.

Materials and

Methods: We used a previously published multi-institutional intracranial EEG dataset from 18 epilepsy subjects with available intracranial EEG recordings prior to surgery, known resection volumes (e.g., the EZ), and epilepsy cure after resection. In each case, a feature-based spike ripple detector was applied to 10 minutes of data from 10 intracranial EEG channels (5 within and 5 outside of the EZ). All candidate events within the EZ were classified by experts as spike ripples or false positives. All candidate events outside of the EZ were classified as false positives. In total, the training dataset included 1700 true spike ripples and 2550 true negatives. The neural network consisted of 5 layers: (i) a sequence input layer with two inputs (the real and imaginary parts of the Fourier transform of an event at frequencies 1 to 300 Hz), (ii) a bidirectional LSTM layer with 256 hidden units, (iii) a fully connected layer with two outputs, (iv) a softmax layer, and (v) a classification layer that computes the cross-entropy loss for classification. To evaluate the performance of the LSTM network to classify spike ripples, we performed leave-one-out cross-validation. To do so, we trained the neural network on the data of all subjects but one, and tested the performance of the network on the left-out subject’s data. This process was repeated for all subjects. For each iteration, we computed the F1 score, specificity, and positive predictive value (PPV).

Results, Conclusions, and Discussions: Subject characteristics are provided in Table 1. To minimize the impact of sparse events, we tested the LSTM network performance on subjects with ≥50 true spike ripples (n = 8). The optimal mean F1 score was 0.76, which was achieved using a probability threshold of 0.13. At this threshold, the mean PPV was 0.73 and the mean sensitivity was 0.84 (Figure 1). This performance was similar to or improved compared to the currently available detector. In addition, the LSTM detector required ~150s to analyze 10 min of data from 10 channels on a single processor with 2.3 GHz CPU (32 GB RAM). This was ~320% faster than the currently available detector which required ~480s to analyze 10 min of data from 10 channels on a faster processor.

Conclusions:
The LSTM network performed with similar or superior accuracy to the currently available detector and had improved efficiency, making it more suitable to identify the EZ in patients undergoing epilepsy surgery evaluation.

Acknowledgements (Optional): This project is supported by NINDS R01NS119483

References (Optional): 1. Shi W, Shaw D, Walsh KG, et al, 2024. Spike ripples localize the epileptogenic zone best: An international intracranial study. Brain. 2024;147(7):2496-2506. doi:10.1093/brain/awae037

2. Schlafly ED, Carbonero D, Chu CJ, Kramer MA, 2024. A data augmentation procedure to
improve detection of spike ripples in Brain Voltage Recordings. Neuroscience Research.
doi:10.1016/j.neures.2024.07.005

3. Xu Z, Wang T, Cao J, Bao Z, Jiang T, Gao F, 2021. BECT spike detection based on novel EEG sequence features and LSTM algorithms. IEEE Trans. Neural Syst. Rehabil. Eng. 29,
1734–1743. doi:10.1109/TNSRE.2021.3107142.