Rationale:
SCN8A-related disorders (SCN8A-RD) are characterized by large heterogeneity in developmental outcomes and high rates of pharmacoresistant epilepsy. A significant challenge in treating these patients is the lack of evidence on “best” first line antiseizure medications (ASMs), especially given the known variable response to current standard of care ASMs. Here we explore the utility and feasibility of developing a machine learning (ML) tool to aid in selecting the first ASM for an individual based on clinic phenotype, genotype, and output of previously developed ML algorithms.
Methods:
Caregiver-reported data were collected in the SCN8A Patient Registry on clinical phenotype for 220 individuals. These individuals were classified into 4 phenotypic subgroups (Hack et al. 2024b Epilepsia) using previously developed ML classifiers (Hack et al. 2023 Neurol. Genet.; Hack et al. 2024a BioOpen). Input features included clinical profile and predicted likelihood from previous models. ASM data were collected as “best” and “worst” ASMs determined by impact on seizure frequency and duration, alertness, and side effects and transformed into 2 output matrices of 18 columns. Two neural networks (NN) were constructed to predict best and worst ASM. Predictions for best ASMs were split into four categories: ASMs that were reported as best, ASMs that were never tried, ASMs tried but not among best or worst, and ASMs that were detrimental.
Results:
The constructed NNs perform with 51.5 % and 45.16% accuracy for best and worst ASM, respectively. The categorical cross entropy values for each model are 4.05 and 2.68, respectively. When evaluating the top 3 predictions for individuals in the training set (n=154) 88 were correct, 253 were possibly correct, and 121 were wrong. In the test set (n=66), 31 were correct, 119 were possibly correct and 48 were wrong. The worst ASMs were never among the top 3 predictions.
Conclusions:
This study serves as a proof of concept that ML tools can be applied to rare disease registries and aid clinicians in choosing the best first line ASMs. The models provide an accurate prediction ~50% of the time in each case, which is a significant improvement over random trial of ASMs. This model succeeds in identifying a best ASM for 50% of patients and a potentially beneficial ASM for up to 75% of patients within the first 3 recommended prescriptions. Most importantly, this tool is effective at limiting the risk of prescribing an ASM that will be detrimental to the health of the individual. The limitations are largely data related, particularly in how an ASM is determined to be among the best or worst ASMs. Developing infrastructure to collect high quality longitudinal data on seizure frequency merged with electronic medical records will improve these models. Another significant limitation is the different prescription practices between clinicians, which reduces the signal of effective ASMs. While the use of ML tools in clinical settings requires significant development, testing, and validation, this study provides a foundation for constructing more effective ML models as data and methods improve.
Funding: None