Abstracts

Rationale: Antiepileptic drug (AED) monotherapy is recommended for most patients with newly diagnosed epilepsy. However, not all patients respond adequately to the first or subsequent regimens. Machine learning (ML) models using big data could help select AEDs with higher likelihood of success from the start. Here we describe the development of AED-specific algorithms that will serve as the engine for a TDSS. Methods: Longitudinal US claims data aggregated from multiple channels, from Jan 2006 to Dec 2015, were obtained from IQVIA. Data for 1,376,756 patients with epilepsy aged >16 years were used to develop cohorts of first-time use for 15 of the most commonly used AEDs. Features from patients’ history–including demographics, diagnoses, comorbidities, hospitalizations, procedures, medications–were used to build tree-based classifiers (random forest) that predict AED treatment stability for a given patient. Stability was defined as continuous coverage of the AED for ≥12 months after AED initiation without addition of another AED (dose changes allowed). Independent predictive algorithms were developed for each AED to provide an AED-specific score. Results: In a sample of 96,006 initial monotherapy decisions for the top 15 AEDs, 19,408 (20.22%) and 39,493 (41.14%) matched the AEDs with the highest and top 3 highest predictive scores, respectively. Matched AEDs had median treatment durations of 21.55 and 21.47 vs 18.81 and 17.95 months for unmatched AEDs, respectively. In the same sample, 1,377 (1.43%) and 5,934 (6.18%) decisions matched AEDs with the lowest and 3 lowest predictive scores, with median treatment durations of 12.00 and 14.30 months, respectively. Relative to 58.24% treatment stability for all initial decisions, use of AEDs within the top 1 and 3 predictive scores led to 10.67% and 7.73% improvement in treatment stability; conversely, use of AEDs within the bottom 1 and 3 predictive scores led to 22.36% and 15.65% reduction in treatment stability, respectively. The C-statistic, a measure of the accuracy of each AED-specific algorithm in predicting treatment stability, ranged from 0.545 to 0.693, with higher values indicating stronger predictions. Conclusions: AEDs attributed higher predictive scores by the algorithms were associated with better treatment stability than those with lower scores. These results suggest that using ML models can potentially help clinicians select the optimal AED for a given patient at diagnosis, allowing them to achieve the best possible outcomes without having to trial through multiple regimens over the years. Ongoing work will determine whether the prediction proxy (1-year treatment stability) adequately approximates clinical outcomes, and whether AEDs matching the model’s predictions yield positive clinical outcomes. To increase applicability of the predictions to as wide a patient population as possible, future models will include interfaces enabling users to apply different features such as generally-accepted guidelines (eg, AED teratogenicity for women of child-bearing age). Funding: UCB Pharma-sponsored