Abstracts

Rationale: Patients diagnosed with epilepsy (PWE) face major uncertainties in terms of outcomes to taking anti-epileptic drugs (AEDs). While 65% of them become seizure free, the remaining 35% do not respond to medication and are subject to drug resistant epilepsy (DRE). Existing guidelines to prescribe AEDs do not always lead to the best outcome for a given patient as it is often implemented as a trial and error process increasing the healthcare cost and journey length. Patients who do not respond to their first AED may take up to 10 different regimens and spend years before finding an effective one [1]. Thanks to the availability of clinical data, there has been a recent interest [2] in developing decision support (DS) algorithms (e.g. based on machine learning) for treatment recommendations. The overall clinical impact of such algorithms and the required accuracy is not fully understood. In this study, we have developed a computational model to assess the effect of treatment choice made using DS algorithms in terms of seizure freedom rate and treatment duration for PWE.

Methods: We have developed a stochastic framework to simulate patient journey and quantify the impact of hypothetical DS algorithms and their accuracy in AED regimen selection and detection of DRE. The model is formalized as a Markov decision process represented by a series of transition probabilities between patient outcomes through taking each AED regimen. Based on clinical findings, we assumed: (1) 35% of the PWE are DRE; (2) in the clinical (baseline) model, the probability of seizure-freedom after each regimen is based on [1] (Fig. 1); (3) the duration of each regimen follows a uniform distribution between 6 to 18 months [1, 2]. We assumed that an algorithm exists that recommends the optimum AED regimen and flags the DRE. If patients do not become seizure free, the same algorithm is re-applied. We simulated the model with variable accuracy of the algorithm and used the Monte-Carlo approach to conduct probabilistic analysis.

Results: Simulations reveal that the patient journey based on clinical data is comparable to that guided by a DS algorithm with an accuracy of 70%. Higher accuracy improves the journey, e.g. fewer AED switches and higher seizure freedom rate per AED regimen (Fig. 2-A) and shortens the duration (Fig. 2-B) significantly (pval< < 0.05) compared to clinical data.