Automated sleep detection reveals differences in sleep patterns in an animal model of neocortical epilepsy.

Sleep in epilepsy is best studied in longitudinal preclinical animal models, where state changes can have significant effects on epileptic activities. Voluminous data makes it very difficult to mark sleep stages manually. This demands an automated way to detect sleep and wake states. We developed an approach to characterize sleep-wake states in continuous video-electroencephalography (EEG) recordings in animals. We compared brute force approach based on frequency band-power based thresholding with machine learning algorithms to detect sleep in 600 hours of EEG data from 4 epileptic and 2 control animals. We found that conventional delta and theta band-powers were prominent in sleep; however, this was not sufficient to detect sleep algorithmically. We therefore extracted a set of novel frequency bands to robustly differentiate individual sleep states by using brute-force algorithm and machine learning models, among which k-nearest neighbors (KNN) was the best predictor of sleep with 94% accuracy. We subsequently characterized sleep patterns in animals with chronically induced epileptic spiking in the neocortex from tetanus toxin injections using brute-force algorithm. We found that epileptic spiking animals (without seizures) sleep more frequently, with significantly longer sleep segments and overall daily sleep time, as compared to control animals. This automated algorithm could help expedite sleep studies and help us understand the relationship between sleep and patients with epilepsy.

Engineering nonlinear epileptic biomarkers using deep learning and Benford's law.

In this study, we designed two deep neural networks to encode 16 features for early seizure detection in intracranial EEG and compared them and their frequency responses to 16 widely used engineered metrics to interpret their properties: epileptogenicity index (EI), phase locked high gamma (PLHG), time and frequency domain Cho Gaines distance (TDCG, FDCG), relative band powers, and log absolute band powers (from alpha, beta, theta, delta, low gamma, and high gamma bands). The deep learning models were pretrained for seizure identification on the time and frequency domains of 1 s, single-channel clips of 127 seizures (from 25 different subjects) using "leave-one-out" (LOO) cross validation. Each neural network extracted unique feature spaces that were interpreted using spectral power modulations before being used to train a Random Forest Classifier (RFC) for seizure identification. The Gini Importance of each feature was calculated from the pretrained RFC, enabling the most significant features (MSFs) for each task to be identified. The MSFs were extracted to train another RFC for UPenn and Mayo Clinic's Seizure Detection Kaggle Challenge. They obtained an AUC score of 0.93, demonstrating a transferable method to identify and interpret biomarkers for seizure detection.

Modeling Nerve Compression in Carpal Tunnel Syndrome.

Nerve function loss can result from a variety of conditions that are either sudden onset like head and spinal cord trauma or slowly develop from chronic pressure as in the case of carpal tunnel syndrome. In either case we see compression ofthe nerve ultimately resulting in axon demyelination and loss of signal conduction. For chronic conditions such as carpal tunnel syndrome, treatments focus on alleviating symptoms. Some patients undergo surgery which can be successful in relieving pressure on the median nerve by inflamed surrounding tendons. Symptoms of classical carpal tunnel syndrome have been debated and sometimes patients experiencing similar numbness and pain in the hand do not necessarily have the underlying condition of a compressed median nerve. Therefore, better markers are needed for determining true cases of nerve compression as well as clinical measures to indicate the need for surgical treatment. We have demonstrated a correlation between clinically observed nerve compression derived from MRI slides and clinically observed increases in conduction delay. We have done this by computationally modeling a myelinated axon with various levels of compression and finding the increase in conduction delay from a normal control with no compression. We show that conduction delay measurements could be used as a clinical tool to determine the amount of nerve compression present in a patient of mild carpal tunnel syndrome although further data is required to create a fully predictive model.