By Steven Karceski, MD
Director of Clinical Trials, Weill Cornell Epilepsy Center
SYNOPSIS: A new deep learning artificial intelligence algorithm was able to identify the most effective initial drug to treat newly diagnosed epilepsy, compared to the physicians’ clinical judgment. The algorithm required prospective, carefully collected clinical data for its success.
SOURCE: Hakeem H, Feng W, Chen Z, et al. Development and validation of a deep learning model for predicting treatment response in patients with newly diagnosed epilepsy. JAMA Neurol 2022;79:986-996.
Epilepsy is an illness in which a person experiences recurrent seizures. Although it sounds like epilepsy is “one disease,” there are many kinds of epilepsy. And there are many causes of epilepsy: prior brain injuries, prior brain infection, and genetic mutations are just a few. In 2022, in the United States, there are many types of treatments for epilepsy: medications, implanted devices, special diets, and brain surgery. With advances in clinical research, there now are more than 30 medications that can be used to treat seizures. The increasing number of medications offers many treatment options for patients. However, the growing list of choices presents a challenge to physicians. With such a long list of options, which one is best?
Studies have shown that the first medicine a person tries is important. If a person’s seizures stop with the first medicine, they often do well over a long period of time. People who continue to have seizures while taking the first medication are more likely to continue to have seizures. This raises many questions. Does the choice of the first medicine make a difference? If true, is there a way that doctors can make better choices? For this reason, Hakeem and colleagues performed this study of the application of artificial intelligence (AI) techniques to help in the selection of the most effective anti-epileptic medications. The authors hypothesized that machine learning algorithms might find patterns linking treatment outcomes to patients’ health data.
To evaluate whether AI could help with doctors’ selection of the best medication to treat epilepsy, the investigators first had to collect information from a group of people with newly diagnosed epilepsy. They gathered information on 1,798 adults (older than 18 years of age; the average age was 34 years) in five medical centers in four countries (Glasgow, United Kingdom; Kuala Lumpur, Malaysia; Chongqing, China; Guangzhou, China; and Perth, Australia). All were newly diagnosed with epilepsy and started on a single medication. Each person was followed for one year to see how they responded to their first medicine. Were they completely seizure-free? Did they have side effects? Did they continue to have seizures and need to change medication?
To understand the factors that influence a response to medicine, the investigators looked at a large series of clinical variables, collected prospectively from the cohort. The clinical variables were each person’s age at initiation of treatment, sex, medical history (febrile convulsions, infections, head trauma, substance abuse, alcohol abuse, stroke, psychiatric disorder), type of epilepsy, number of pre-treatment seizures, electroencephalogram and brain imaging findings, and drug used for initial treatment. These clinical data were entered into a database and then processed with a new AI deep learning program to analyze these clinical factors. The goal was to show that the AI program would be able to identify associations and connections that the doctors’ human brain did not. The study hypothesis was that AI would help to select a better medication that would be more likely to prevent epileptic seizures.
First, investigators compared the new AI deep learning model to five other available programs and showed that the new deep learning model outperformed the others. Second, they used this new method to see if they could predict which medication choice was going to be best. Since they had five groups of people from five different medical centers, they applied the new method to each of these groups and found that the AI model was able to identify medicines that were more likely to help the person’s epilepsy.
COMMENTARY
There are several limitations to the study. The investigators recognized the limited amount of information that was available for each person with epilepsy. The groups were taking seven anti-seizure medications. With more than 30 anti-seizure medicines worldwide, these results need to be validated in a larger group of people taking the other anti-seizure medicines. Second, the group of people with epilepsy were adults. Would this AI deep learning model work for children? Would it work for infants with seizures? The authors stated that there may be other medical test results that could influence the AI decision-making. As an example, the model may be better if the results of genetic testing were added.
There are several important conclusions to be drawn from this study. First, this is a new method of deep learning. If further studies agree, this method could lead to the development of a very helpful tool that physicians could use in addition to their medical knowledge to personalize and optimize the treatment of epilepsy.
With further study, we may discover that AI-assisted decision-making is superior to human decision-making alone. If so, using this tool, neurologists might be able to arrive more quickly at a correct decision regarding initial medication for their patients with epilepsy. This could result in less illness, and improved long-term outcomes, and will help many people with epilepsy.
