UCLA

OBJECTIVE: Despite rapid advances in understanding the disease spectrum and its progression, little is known about which antiseizure medications (ASMs) are likely to be beneficial or detrimental as first-line therapies for patients with SCN8A-related epilepsy (SCN8A-RE). This is a critical issue given low rates of seizure freedom and treatment resistance rates that exceed 75%. In this study, we test the hypothesis that machine learning (ML) algorithms can improve selection of ASMs that are likely to benefit patients with SCN8A-RE. METHODS: We leverage comprehensive medical data in the International SCN8A Patient Registry to construct a neural network that recommends ASMs based on a caregiver-centered composite measure incorporating improvements in seizure control, alertness, and side effects. We directly compare the recommendations of the algorithm to preferences of clinician experts through a follow-up survey and evaluate how ASM selection is influenced when informed by the ML algorithm. RESULTS: Despite challenges resulting from the prevalent use of polypharmacy and frequent suboptimal treatment responses, the algorithm identified ASMs likely to be beneficial in 76% ± 3% of cases while never recommending a detrimental ASM in 1100 trials. Clinician experts independently recommended beneficial ASMs in 22% (16/72) of cases, a rate that increased to 46% (11/24) when choices were given based on algorithm recommendations. SIGNIFICANCE: The results indicate that ML algorithms can improve selection of ASMs that are likely to be beneficial in the early treatment of SCN8A-RE patients, with little risk of recommending ASMs with detrimental effects-a particular hazard for patient populations requiring long-term maintenance on polypharmacy. The results also expand the number of recommended ASMs from two sodium channel blockers (SCBs) identified in a recent consensus process to five SCBs and a γ-aminobutyric acidergic drug. The algorithm lays the groundwork for incorporating composite measures that include both seizure control and quality of life metrics.