Using advanced “deep learning” informatics mapping connectome responses in patients with treatment-resistant mesial temporal lobe epilepsy, researchers were able to predict with greater than 20 percent accuracy which individuals would be more likely to benefit from surgery.
Deep learning is an artificial intelligence (AI) computational process that identifies unique individual patterns in large volumes of human connectome data to evaluate neural connectivity activated by specific tasks. Using structural connectivity as the primary input, personalized neural networks showed high accuracy in predicting seizure outcomes after surgery in patients who failed to adequately respond to multiple antiseizure medications, including both ipsilateral and contralateral seizures.
Compared to a standard model using only clinical features that resulted in postsurgery response accuracy of 50 percent or less, the experimental approach predicted surgical success in preventing seizures wih 88 percent to 95 percent accuracy and a negative predictive range of between 79 percent and 87 percent, according to the analysis published in the September issue of Epilepsia.
The researchers attributed the high accuracy to classifications not only in the ipsilateral temporal and extratemporal regions but also in the contralateral hemisphere.
“Deep learning may be a powerful statistical approach to isolating abnormal patterns from complex datasets that could provide highly accurate seizure outcome prediction after surgery,” said corresponding author Ezequiel Gleichgerrcht, MD, a researcher in the department of neurology at the Medical University of South Carolina.
To conduct the analysis, the research team retrospectively studied 50 consecutive patients who were diagnosed with medication refractory TLE; all had undergone T1-weighted magnetic resonance imaging (MRI) and had clinical assessments before surgery. The researchers used the whole-brain diffusion tensor imaging to reconstruct presurgical connectomes.
The team added the information into the experimental network, incorporating multiple variables and data points—including imaging and clinical information—and trained the computer “deep learning” model to identify patterns of network abnormalities associated with becoming either seizure-free or refractory to treatment.
They compared the learning program's predicted surgical treatment outcomes with actual outcomes. The team confirmed either normal MRI or unilateral hippocampal atrophy on the side of seizure onset with prolonged video-electroencephalography (EEG) monitoring, and 3T MRI confirmation of either normal or unilateral hippocampal atrophy, with in-person interviews and assessments at follow-up at least one year after surgery.
Based on one-year patient interviews and clinical assessments, the model's predictive ability was comparable across a wide range of outcomes.
“One of the challenges relating connectivity patterns to clinical phenotypes is that aberrant connectivity is variable across individuals, with different patients exhibiting different foci of abnormalities in limbic and extralimbic networks,” Dr. Gleichgerrcht said. “Therefore, mapping epilepsy brain networks might be improved by statistical approaches that could isolate abnormal individualized patterns in complex datasets.”
He noted that there are nonetheless important limitations in the study that must be considered. The data were retrospective and from a limited sample of patients. Even so, he told Neurology Today, the technique may one day play an important role in identifying complex biomarkers of neural architecture abnormalities to predict distinct clinical phenotypes in epilepsy.
“This research is quite meaningful, because it helps improve surgical-outcome prediction methods,” said Michael R. Sperling, MD, FAAN, professor of neurology and director of the Jefferson Comprehensive Epilepsy Center at Thomas Jefferson University in Philadelphia.
“At present, we select patients based upon a variety of standard clinical measures, generally incorporating the history, examination, interictal and ictal EEG, ictal behavior, MRI, PET, SPECT, and other tests,” Dr. Sperling said.
“We are usually able to do a reasonably good job, but many patients who appear to be excellent surgical candidates experience seizures afterwards. This research expands previously published work by other authors, some of which are cited in the paper, in an attempt at personal prediction of outcome. The methods provide an objective means of achieving this.”
Dr. Sperling said that the next step would be replication by another group and comparing it with other novel individualized prediction methods. Although the benefits include individualized outcome prediction, the potential downside is discouraging patients who are not ideal candidates from having surgery, he said.
“First, no technique is perfect, so some might be seizure-free despite prediction of continued seizures. Second, and most important, surgery causes a significant improvement in seizure frequency and severity in most patients, even if they are not entirely seizure-free,” he said.
Dr. Sperling cited research that shows mortality is substantially reduced in patients who continue to have seizures after surgery, most likely related to abolition or reduction in number of bilateral tonic-clonic seizures.
“So even though some seizures might persist, lives are saved by surgery,” Dr. Sperling said. “This [study] highlights the importance of not viewing epilepsy surgery outcome in a binary fashion—either seizure-free or not. Even not being seizure-free can lead to a meaningful improvement because of reduced risk of dying from epilepsy.”
He also said that the technique, if fully confirmed, may have the ability to detect patterns of connectivity that predict outcome in other types of epilepsy. It is also possible that specific patterns might be associated with response to medical therapy—certainly something worth exploring, he said.
In Dr. Sperling's view, the major limitation is that this requires high-level processing of good-quality scans, which are often marred by motion artifact, and interpretation can be impaired. Assessing connectivity is not an off-the-shelf method, he added, and it requires a highly skilled team, which is probably not present at most sites. For this to become widely accepted, it will need to be either automated or outsourced to a central lab capable of doing the image processing.
Kathryn Davis, MD, MS, assistant professor of neurology and medical director of the epilepsy monitoring unit and epilepsy surgical program at the University of Pennsylvania Hospital, told Neurology Today that the authors achieved a high classification accuracy surpassing a comparison model using clinical variables alone, but the results indicating broad bilateral structural changes in temporal lobe epilepsy are salient and consistent with prior imaging studies such as resting state functional MRI (fMRI).
“This work adds to a growing literature using network neuroscience, which holds great promise for identifying specific epilepsy networks noninvasively. The next logical steps would be to the expand analysis outside of temporal lobe epilepsy, expand the work to multisite larger cohort studies, and incorporate multimodal imaging data such as resting state fMRI.
“These approaches can and should be applied more broadly to other types of focal onset epilepsy, such as extratemporal epilepsy. In addition, machine learning approaches to understanding idiopathic generalized epilepsy may help us understand the underlying network perturbation causing seizures and cognitive deficits in this population.”
The addition of multimodal data may improve or change the predictive ability of their algorithm, she added. “Only including temporal lobe epilepsy patients, although the most common type of localization-related epilepsy, is not broadly applicable to the epilepsy surgical population in general. Finally, the researchers used only one neural network design. As the field of network neuroscience continues to advance, we may identify different approaches and parameters that outperform the authors' model.”
When applied to multimodal noninvasive imaging data, machine learning analysis, including deep learning, can potentially lead to noninvasive seizure onset zones and help guide clinicians in their treatment decisions in the challenging drug-resistant epilepsy patient population, Dr. Davis said.