Continental generalization of an AI system for clinical seizure recognition

Abstract

A vast majority of epileptic seizure (ictal) detection on electroencephalogram (EEG) data has been retrospective. Therefore, even though some may include many patients and extensive evaluation benchmarking, they all share a heavy reliance on labelled data. This is perhaps the most significant obstacle against the utility of seizure detection systems in clinical settings. In this paper, we present a prospective automatic ictal detection and labelling performed at the level of a human expert (arbiter) and reduces labelling time by more than an order of magnitude. Accurate seizure detection and labelling are still a time-consuming and cumbersome task in epilepsy monitoring units (EMUs) and epilepsy centres, particularly in countries with limited facilities and insufficiently trained human resources. This work implements a convolutional long short-term memory (ConvLSTM) network that is pre-trained and tested on Temple University Hospital (TUH) EEG corpus. It is then deployed prospectively at the Comprehensive Epilepsy Service at the Royal Prince Alfred Hospital (RPAH) in Sydney, Australia, testing nearly 14,590 hours of EEG data across nine years. Our system prospectively labelled RPAH epilepsy ward data and subsequently reviewed by two neurologists and three certified EEG specialists. Our clinical result shows the proposed method achieves a 92.19% detection rate for an average time of 7.62 mins per 24 hrs of recorded 18-channel EEG. A human expert usually requires about 2 hrs of reviewing and labelling per any 24 hrs of recorded EEG and is often assisted by a wide range of auxiliary data such as patient, carer, or nurse inputs. In this prospective analysis, we consider humans' role as an expert arbiter who confirms to reject each alarm raised by our system. We achieved an average of 56 false alarms per 24 hrs.

Adaptive, Unlabeled and Real-time Approximate-Learning Platform (AURA) for Personalized Epileptic Seizure Forecasting

“…The seizure detection model is trained on the Temple University Hospital (TUH) seizure corpus [43] from the U.S., while the prediction model is pre-trained using the European (EU) EPILEPSIAE dataset [44]. The AURA self-learning process is used with the Australian test set from the RPAH where all human-annotated labels have been censored [19], and each patient starts with the same pre-trained prediction model that adapts over the course of their multiple monitoring sessions. Upon completion of all sessions, the sequence of predictions generated by the forecasting network is compared to the uncensored ground truth to provide a performance measure of sensitivity and the number of false alarms.…”

“…(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. *Training the detection model follows the same procedure as in [19]. The TUH seizure corpus provides dedicated validation and test sets, although the labels from the test set are unreleased.…”

“…Detecting a seizure at the time of or immediately after onset has had far more success over forecasting seizures in advance [18]. Several machine learning techniques have been used to augment neurologist readings, leading to faster conclusions whilst main-taining specialist-level EEG-reading performance for the identification of seizures as they happen [19, 20, 21]. Unfortunately, seizure aversion is no longer an option when relying on detection mechanisms alone.…”

“…The world’s largest public seizure database, the Temple University Hospital (TUH) seizure corpus, contains EEG recordings of over 4,000 patients, and is commonly used for testing and validating the performance of seizure detection systems that achieve close to expert clinician performance [4, 19]. The recordings commence between 5 to 35 minutes pre-ictally which limits the length of the SPH.…”

“…Weak supervision has previously been applied by obtaining inaccurate labels from a mix of experts and novices for real-time seizure detection [28]. Our approach can be distinguished as we wholly do away with manual annotations, and instead use a detection model shown to perform similarly to neurologists [19]. These machine-generated detection labels are then used as targets for the prediction model.…”