, according to authors of a recent review. Clinical prediction models can help neurologists identify which patients could benefit from more aggressive early treatment, authors added, although concerns over bias and model applicability leave room for improvement.
Triggering Aggressive Treatments
“These models are helpful because if you can predict that someone is going to do well with one or two medications, that’s great,” said Aatif M. Husain, MD. “But if you know early on that someone likely will not do well, will need many medications, and still not have their seizures under control, you’re much more likely to be more aggressive with their management, such as closely refer them to a specialist epilepsy center and evaluate them for surgical treatment options. This could minimize the amount of time their seizures are inadequately controlled.” Dr. Husain is an epileptologist, neurologist, and sleep medicine specialist at Duke University Health System in Durham, North Carolina. Dr. Husain was not involved with the study, which was published in Epilepsia.
“But the other important finding is that these models so far have not been that great,” he added.
Dr. Aatif M. Husain
Prognosis Predictors
Investigators Corey Ratcliffe of the University of Liverpool in England and colleagues systematically searched MEDLINE and Embase for relevant publications, ultimately analyzing 48 models across 32 studies. The strongest predictors of seizure remission were history and seizure types or characteristics, the authors wrote, followed by onset age.
Regarding seizure history, a March 2018 JAMA Neurology study and a December 2013 BMC Neurology study linked factors such as history of seizures in the year pre-diagnosis, family history of epilepsy, and history of febrile seizures and of migraines with lower chances of seizure remission. Seizure types with increased chances of poor outcomes in the review included status epilepticus and seizures with complex or mixed etiologies. Additional seizure types associated with poor control include tonic-clonic seizures, frequent focal seizures, and seizures stemming from certain genetic predispositions, said Dr. Husain.
Although the roles of many of the foregoing factors are easily explained, he added, other variables’ impact is less clear. Younger onset often signals more refractory seizures, for example, while data regarding older onset are mixed. “Sometimes older individuals will have mild epilepsy due to a stroke, tumor, or something that can be relatively easily treated,” said Dr. Husain. Conversely, epilepsy can become more complicated if such patients take several medications and/or have coexisting medical problems that seizures or antiseizure medications exacerbate. “So sometimes it’s not so obvious.”
Incorporating Imaging, AI
Dr. Husain found it surprising that very few of the selected models incorporated EEG and MRI findings. “Subsequent research should look at those, since they are important diagnostic tests.” Moreover, he recommended including more sophisticated quantitative and connectivity analyses of EEG and MRI data. These analyses might provide additional prognostic information beyond a simple visual analysis of these tests, Dr. Husain explained, although their potential here remains unproven.
As for factors not represented in the review, he said, future studies will help clarify AI’s role in predicting newly diagnosed epilepsy outcomes. A study published in Epilepsia showed that among 248 potential pediatric surgical candidates, those whose providers received alerts based on machine learning analysis of prior visit notes were more likely to be referred for presurgical evaluation (9.8% versus 3.1%). Future clinical models will use AI to examine not only established elements of neurologic history, said Dr. Husain, but also other types of history such as socioeconomic characteristics, geographic location, and other such data.
Additionally, study authors recommended a standardized approach to prediction modeling, using Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines. Using consistent definitions, outcomes, and reporting requirements will facilitate communication among researchers, reduce bias, and support systematic between-study comparisons, Mr. Ratcliffe and colleagues wrote.
Reaching General Neurologists
Epilepsy specialists are generally aware of reliable outcome predictors, Dr. Husain said, though they do not use models per se. “But the vast majority of patients with epilepsy are seen by general neurologists.” And the lack of awareness among these physicians and primary care practitioners drives a need for education to facilitate appropriate referrals to subspecialty centers, he said.
The stakes for timely referrals can be high. Although using appropriate outcome models improves patients’ quality of life sooner, said Dr. Husain, allowing seizures to go untreated or undertreated results in neuroplastic changes that hinder long-term seizure control.
The fact that all 32 included studies reflected a high risk of bias, and 9 studies raised high applicability concerns, raises questions regarding the models’ validity, he added. Mr. Ratcliffe and colleagues attributed both types of concerns to the fact that 20% of included studies used baseline treatment response data as outcome predictors.
Nevertheless, Dr. Husain cautioned against dismissing prediction models in newly diagnosed epilepsy. “Practicing neurologists need to realize that the perfect model has yet to be developed. But the current tools can be used to help manage patients with epilepsy and predict who will do well and not as well,” he said.
Dr. Husain is a member of the American Epilepsy Society. He has been a consultant and researcher for Marinus Pharmaceuticals, PranaQ, and UCB, and a consultant for Eisai, Jazz Pharmaceuticals, Merck, and uniQure. Study authors reported no funding sources or relevant conflicts of interest.