Rationale:
Mutations in the SCN1A gene are the most frequent monogenic cause of epilepsy and can result in very different outcomes: Dravet syndrome (DS), a highly treatment resistant severe epilepsy with profound learning disability and significant mortality, or genetic epilepsy with febrile seizures plus (GEFS+), a well-controlled epilepsy syndrome with normal cognition. Distinction on clinical grounds alone is challenging in the first two years of life, with the diagnosis of Dravet syndrome often delayed for several years. There is a need for early transformative treatments in DS including gene therapy, and objective prediction models are required to enable early recognition of a child’s risk of developing DS vs GEFS+.
Method:
We developed a DS prediction model using clinical and genetic data that are easily accessible to clinicians in any young child presenting with a pathogenic SCN1A variant. The model includes age at seizure onset and a newly developed SCN1A-specific genetic score combining paralog conservation of the mutated amino acid position with physicochemical properties of the observed substitution. We retrospectively reviewed clinical and genetic data for 975 SCN1A-positive Dravet syndrome and GEFS+ patients ascertained from seven countries. Data from the UK, France, Italy, Netherlands and Denmark formed the training and initial testing cohort (n=698), whereas two independent blinded cohorts from Australia (n=205) and Belgium (n=72) were used to validate the model.
Results:
A high SCN1A genetic score, as well as young age at seizure onset, were individually associated with DS (p-value=1.60x10-27; p-value=1.70x10-36). We trained a supervised machine learning model on the DS or GEFS+ clinical outcomes. Patients with probability values above 50% were predicted to be Dravet syndrome and patients with a probability below 50% were predicted to be GEFS+. The model was effective in separating DS from GEFS+ (AUC=0.895) and outperformed other models we developed based on age at seizure onset combined with established pathogenicity scores such as CADD (AUC=0.823), REVEL scores (AUC=0.808) or a model solely based on age at seizure onset (AUC=0.853). Among the 277 cases from the blinded validation cohorts (208 DS and 69 GEFS+), a total of 206 DS patients were correctly predicted, achieving a 99% accuracy. Table 1 describes the sensitivities and specificities observed at different thresholds. We deployed the model into an online tool designed to calculate a patient's probability (%) of developing severe DS vs mild GEFS+, which will be released at the meeting.
Conclusion:
Accurate prediction whether a young child with a pathogenic SCN1A variant will develop severe DS or mild GEFS+ is important, however, there are currently no validated models that combine a child’s multiple risk factors into a single outcome predictor. Our clinico-genetic prediction model allows an objective estimation whether a child is likely to develop DS vs GEFS+ at a very young age. This approach will assist clinicians with prognostic counselling and decision making about the initiation of treatment earlier than previously possible.
Funding:
:None
FIGURES
Figure 1