Abstracts

Rationale:
The decision-making process in epilepsy surgery for patients undergoing invasive investigations is strongly connected to the interpretation of the intracranial EEG (iEEG). Although deep-learning methodologies have demonstrated high performance in detecting scalp EEG seizures, the limitations of acquiring intracranial ictal events from the Epilepsy Monitoring Unit in traditional clinical practice has not allowed for adequate datasets for training and learning. The recent establishment of closed-loop responsive neurostimulation (RNS) as a treatment option has provided the opportunity to overcome the traditional limitations, and acquire large datasets of intracranially recorded seizures from each individual patient over months and years. In a previous work (Constantino et al., AES, 2019), we showed that a convolutional neural network architecture that processes RNS-derived iEEG ictal events can reach an event detection accuracy of 0.80 - 0.83, which is comparable to the level of agreement among qualified reviewers (0.79; Quigg et al., 2015). In this study, we demonstrate improved machine learning performance in both event detection and event onset timing by using a time-frequency analysis-based deep learning architecture.
Method:
We used the efficientnet-b0 deep neural network architecture (Tan and Le, 2019). Instead of the raw time series data, the network takes as input a short-time-fourier-transform of the EEG time-frequency matrix. Similar to our previous work, the network produces two predictions: the binary outcome of whether the recording contains a seizure and the estimated seizure onset time. Starting from network weights pre-trained on Imagenet (Russakovsky et al., 2015), we train the network’s final layer on a large human intracranial seizure dataset containing 5,380 ictal events recorded from 21 RNS patients. Each recording was manually processed and ictal events were marked and confirmed by two qualified reviewers.
Results:
Using hold-one-patient-out cross validation, we showed that the network achieves a detection accuracy of 0.92 ± 0.07% area under precision-recall curve on previously unseen patients, improving our previous result by 9 percentage points without additional data. In addition, we found the network has a mean absolute accuracy of 9.6 ± 5.1 seconds in identifying the seizure onset time on previously unseen patients.
Conclusion:
Although we have already shown that deep learning networks can provide expert-level accuracy for seizure detection, we hereby demonstrate superior performance in both event and event onset detection when time-frequency plots are used for neural network training instead of the raw time-domain signal. This unique achievement of improved deep learning performance in iEEG seizure and seizure onset detection can also stimulate an important paradigm shift for iEEG evaluations, as spectral analysis is rarely used in routine clinical practice by experts and evaluations rely on time-domain signal changes. We also envision that deep learning methodologies will significantly contribute to the improvement in management of RNS patients and enable the development of tools that take advantage of highly accurate iEEG seizure and seizure onset detection.
Funding:
:This work received no funding.