Abstracts

Rationale: Finding similar electrographic seizure and interictal activity patterns across epileptic patients may facilitate identifying treatment options that worked in patients with brain activity similar to a given patient. Three methods of finding similar cross-patient iEEG (intracranial electroencephalogram) records in a large iEEG dataset were developed. Performance is illustrated by comparing five “similar” iEEG records returned by each method in response to a “query” record.

Methods: ~1 million 4-channel iEEG records (typically 90 seconds in duration) recorded from 256 patients enrolled in NeuroPace RNS System clinical trials were used. A pretrained ResNet50 CNN model with ImageNet weights was used to extract features from spectrogram images of iEEG channels. The resulting features for each of the 4 iEEG channels were concatenated to produce a combined feature matrix for each iEEG record. Patient-specific dimensionality reduction and clustering using PCA (principal component analysis) followed by t-SNE (t-distributed stochastic neighbor embedding) and BGMM (Bayesian Gaussian mixture models) was performed to identify 2,759 clusters of iEEG records within the 256 patients. Training, test, and validation datasets consisted of 200, 31, and 25 randomly selected patients.

Method 1: A pretrained ResNet50 model was used to extract features from iEEG records in the training and test datasets. The resulting features were sent to a PCA mapping function followed by t-SNE. Similar training records to a test record were determined using k-nearest neighbors (kNN).

Method 2: A ResNet50 model with a 256-neuron embedding layer was fine-tuned using a triplet loss function that minimized the Euclidean distance between centroid records in the training dataset and positive examples (non-centroid iEEG records close to their respective centroids), and maximized the Euclidean distance between centroid records and negative examples (non-centroid iEEG records distant from their respective centroids). Validation patients were used to fine-tune the training hyperparameters. Similar training records to a test record were determined using Euclidean distance.

Method 3: A ResNet50 model was fine-tuned by training on an auxiliary task to classifiy iEEG records as seizures and non-seizures using ~130,000 manually labeled iEEG records¹. The trained ResNet50 model (with the final classification layer removed) was used as a feature extractor. Similar training records to a test record were determined using kNN

Results: All 3 methods had comparable performance identifying similar baseline (non-seizure) and interictal spiking iEEG records (Fig 1). For electrographic seizures, the ResNet50 model trained on an auxiliary classification task (Method 3) returned the best matches in most cases (Fig 2). The ResNet50 model fine-tuned using the triplet loss function (Method 2) produced the second-best results, followed by the pretrained ResNet50 model described in Method 1.

Conclusions: Three methods for finding similar cross-patient iEEG records in a large iEEG dataset were developed and compared.

Reference:
1. Barry W et al. Frontiers in Neuroscience, 2021 15, p.697

Funding: Please list any funding that was received in support of this abstract.: None.