Abstracts

Rationale: Accurate seizure prediction in persons with epilepsy would create opportunities for both precautionary measures and pre-emptive treatment. Herein, we develop models to predict seizures in one sample and assess its value in predicting future seizures in the same individuals and in a similar small, independent sample. We also test the hypothesis that prediction models improve as follow-up time increases.Methods: We conducted a 2-phase analysis of prospective daily seizure diary data. In Phase 1, using the first 60 days of study, we fit multilevel logit-normal binary response models using previously identified predictors (stress, anxiety, hours of sleep). In Phase 2, we examined how the results from that model predicted the occurrence of seizures over the next 30 days of the study in the same individuals and in a small independent sample. To assess the contribution of a longer observation period, multi-level logit-normal binary response models for the first 11 months of follow-up were predicted and tested for predictions in month 12.Results: In Phase 1, 16 subjects contributed 585 days of observation, including 92 days with seizures, over the first 60 calendar days of diary data collection. Seizure occurrence was modeled as the outcome. In Phase 2, 15 of the 16 subjects who had contributed data to Phase 1 and 4 additional subjects contributed data to days 61 to 90 of diary data collection. During Phase 2, seizures occurred on 59 out of 519 days. For days 61 to 90, a predicted probability of seizure was computed using the Phase 1 model, including the individual level predicted values of the random intercept for any subject who contributed data to both Phase 1 and 2. For the 4 newly enrolled subjects we used only population average effects. For 80% specificity of seizure prediction, the model achieved 63% sensitivity. ROC curves were computed; the area under the curve was 0.75. Similar models were applied to the first 11 calendar months of the study in 43 subjects contributing 6276 days, including 641 with seizures. Using those 11 months of data to model seizures in the 12th month (29 Ss, none newly enrolled, with seizures in 56 out of 816 days) resulted in improved estimation of variance components for sleep and self-prediction, an increase in sensitivity to 71%, and an increase in area under the ROC curve to 0.84.Conclusions: Development of individual-level models for prediction of future seizures based on diary data may be possible. As follow-up time increases, model predictions improve. The possibility of clinically relevant prediction should be examined in a study with electronic data capture, more specific and more frequent sampling, and with subject training and feedback based on model results.