Authors :
Presenting Author: Ankit Khambhati, PhD – University of California, San Francisco
Joline Fan, MD – Neurology – University of California, San Francisco; Anthony Fong, BS – Neurological Surgery – University of California, San Francisco; Jonathan Kleen, MD, PhD – Neurology – University of California, San Francisco; Edward Chang, MD – Neurological Surgery – University of California, San Francisco
Rationale:
Transient epileptiform events commonly occur during the interictal period between seizures in patients with focal epilepsy. Algorithms that track fluctuation in spatial and temporal emission rates of interictal events may aid in seizure localization, prediction, and control. Expert review of intracranial EEG is the gold-standard approach to identify these events. Yet, this approach is cumbersome, suffers from poor intra and interrater reliability, and is ineffective for real-time applications. In this study, we present a machine learning framework for online mapping of the resting-state epileptic brain network that bridges the two extremes of manual review and unsupervised learning by bringing the human expert into the model optimization loop. Our online method adapts continuously to brain state and facilitates co-learning of epileptic network features between human and machine.
Methods:
We developed an algorithm that monitors the intracranial EEG recording stream and continuously learns epileptic network motifs – defined as the spatiotemporal map of an interictal epileptiform event as it begins and propagates. Specifically, our online machine learning method combines constrained convolutional non-negative matrix factorization called seqNMF1, which discovers network motifs, with a dynamical model called OASIS2, which learns the precise event timing and kinetics (rise and decay time constants) of each motif. Ensemble learning discovers a diverse set of network motifs across a large parameter space. Once a set of network motifs is discovered, backpropagation refines the motifs based on active insight from the human expert during continuous online learning.
- Mackevicius E. L., et al. (2019) Unsupervised discovery of temporal sequences in high-dimensional datasets, with applications to neuroscience eLife 8:e38471
- Friedrich J., et al. (2017) Fast online deconvolution of calcium imaging data. PLoS Comput Biol. 2017 Mar 14;13(3).
Results:
We find in three focal epilepsy patients that interictal epileptiform events follow stereotyped spatiotemporal trajectories that reflect dysfunctional architecture of the
epileptic brain network. Network motifs are learned within the first hour of model training and are stably employed to detect epileptiform events over at least 24h. Moderate backpropagation of human expert ratings during motif learning improved accuracy of event detections (AUC=0.9); however stronger backpropagation penalized accuracy markedly (AUC=0.5). Our results demonstrate a trade-off between unsupervised discovery and expert feedback. Conclusions:
We present a novel, semi-supervised algorithm to map network-level representations of interictal events in real-time during in-hospital epilepsy monitoring. If validated, our algorithm has the potential to augment existing clinical diagnostic workflow by characterizing variability in epileptiform activity, mapping dysfunctional network regions, and shortening the time to surgery.
Funding:
This work was supported by CURE: Taking Flight Award, NIH Awards R01MH123770 and R21MH124759, and the New York Stem Cell Foundation.