Article In Brief

Investigators developed an algorithm that uses clinical and genetic data to help determine the treatment response to antiseizure medications in patients with newly diagnosed epilepsy.

Researchers have developed a deep learning model for identifying the best choice of anti-seizure medication for patients with epilepsy, according to findings published online on Aug. 29 in JAMA Neurology.

The model may be the most sophisticated attempt yet to apply artificial intelligence to the tricky realm of choosing the right treatments for these patients, the investigators wrote. While the model can make recommendations that have fared better than random chance and it has the potential for further refining, experts not involved with the paper—as well as the study authors—said the technology is still not ready for clinical use.

Nonetheless, they said the latest model is a significant step forward in ongoing efforts to improve a decision-making process that often leaves patients with failed attempts on their initial medications.

“We think it will make a huge difference for both the patient and the clinician,” said Patrick Kwan, MD, PhD, professor of neurology at Monash University in Melbourne, Australia. “It will give the patient much more confidence in the likelihood of whether a medication will work or not. Then they can organize their life around it.”

The Deep Learning Model

Researchers developed a deep learning model, a type of machine learning, that is “attention-based” and can focus on certain parts of the input more than others, allowing for more subtle patterns to be observed.

“This type of model can extract correlations between latent variables better than other non-attention-based deep learning models,” said Zongyuan Ge, an associate professor at Monash eResearch Center and a deep learning specialist.

Because the model is better at making sense of unlabeled input, it is better suited for using clinical reports and gene sequence information, which researchers incorporate into the model.

Researchers trained and validated the model using five cohorts of patients, all with newly diagnosed epilepsy: 1,065 patients from the Western Infirmary in Glasgow, Scotland; 242 from the University of Malaya Medical Centre in Kuala Lumpur, Malaysia; 191 at First Affiliated Hospital of Chongquing Medical University in China; 189 at WA Adult Epilepsy Service in Perth, Australia; and 111 at First Affiliated Hospital, Sun Yat-Sen University, in Guangzhou, China.

The model considered lamotrigine, valproate, carbamazepine, levetiracetam, oxcarbazepine, topiramate, and phenytoin. The treatment goal was to achieve complete seizure freedom for at least a year on the first medication.

The researchers fed these variables into the model: sex; age at the start of treatment; traits in their history, such as whether they’d experienced febrile convulsions or significant head trauma; the presence of cerebrovascular disease or intellectual disability; whether they’d had either five fewer or five more seizures; whether the epilepsy was focal, generalized, or unclassified; and three categories of EEG and brain imaging findings (normal, abnormal epileptiform, or abnormal nonepileptiform).

Among the validation cohorts, the sensitivity of the transformer model for predicting success ranged from 0.41 to 0.61, and specificity ranged from 0.55 to 0.66. The mean area under the receiver operating characteristic curve (AUROC) ranged from 0.52 to 0.60. Dr. Kwan said the goal is for this number to reach the 0.70 to 0.75 range.

“That’s what we hope,” he said.

The addition of genetic information and full imaging data—either with a language description or the images themselves—likely could help with this, he said. The vision for the model is that information can be pulled from the electronic medical record and a recommendation then given to the clinician, Dr. Kwan said. But he added that the model doesn’t currently account for many of the nuances in how medications are chosen, such as drug-to-drug interactions or side effects.

“At this stage, we’re not suggesting it will replace the clinical decision-making entirely,” he said, but rather, it would be a tool clinicians could use along with other considerations.

But the approach could be helpful in a process that now often leaves physicians with two or three medications from which to choose after considering factors such as seizure type and drug interactions.

“It’s trying to avoid the wrong drug rather than finding the right drug,” Dr. Kwan said.

If the model can be refined, improved choices for first antiseizure medications could mean a better quality of life for patients. Unlike patients with hypertension, for instance, who can see their blood pressure fall and know a drug is working, patients with epilepsy are forced to take a “passage of time” approach and hope for an absence of seizures, he said.

“If we have an accurate prediction, that will provide that certainty, the likelihood of responding or not responding,” Dr. Kwan said.

If the model predicts that none of the pharmaceutical options likely would work, then non-pharmacological options, such as surgery, might be considered earlier, he said.

The tool could also help primary care physicians in their prescribing, he said, especially in rural areas where access to neurologists can be slim.

Input from Independent Experts

Andreas Alexopoulos, MD, MPH, a staff physician at the Epilepsy Center at Cleveland Clinic, applauded the effort but said it would be some time before any possibility of clinical use.

“Given that this study is informed based on retrospective data and diverse data from different centers with variable populations and given—as the authors acknowledge—the modest performance of the algorithm, I don’t think we are ready for something that is a paradigm shift,” he said. “There is a lot of enthusiasm, obviously, in the research community for artificial intelligence and big data analytics to help with decision-making in clinical practice. The enthusiasm has to be tempered by how applicable this is.”

Going forward, inputting data on pathophysiology will be essential to the success of this kind of model, he said.

“It is worthwhile to show that this is feasible and that this is doable and continue to improve and inform these models as knowledge advances,” Dr. Alexopoulos said.

Joyce Cramer, a pharmaceutical and medical device consultant and retired associate research scientist at Yale University who studied epilepsy, said she doesn’t expect this model to make its way into clinical use.

“I’ve been through this many times over the years,” she said. “They keep coming out with some kind of an algorithm, and they’re sure this is going to be it. It never gets picked up because they really don’t work any better than gut instinct.”

Despite the large amount of research done in epilepsy treatment, not much has focused on guiding clinical choices. There has been “an awful lot of data but very little way to make a definitive treatment algorithm,” Cramer said. So far, the accuracy of this model, at about 60 percent, she noted, is not impressive.

“That’s very close to 50/50,” she said.

Her skepticism about the model is based, in part, on the input for it so far. For instance, she said, the patient history data likely would not help identify the right treatment.

“Whether somebody’s epilepsy is due to a stroke or some other kind of head injury, that’s not going to make any difference in how they respond to medication,” she said.

The real test would be a comparison of machine recommendations compared to the clinician’s own decision. “If you compare it to non-machine, human judgment, what would a good neurologist come up with in this situation versus what the machine could come up with?” Cramer said.

She also noted that the drugs used in the model so far don’t include the newest options. “If you included certain other drugs, the newer drugs, that would make a difference,” Cramer said. “But these are all very standard drugs.”



Dr. Kwan has received fees for consulting/speaking from Angelini, Eisai, LivaNova, and UCB Pharma outside the submitted work.



Source:, Thomas R. Collins