Abstracts

Rationale: The first antiseizure medication (ASM) is ineffective or intolerable in 50% of epilepsy cases and about two-thirds become seizure-free after a trial-and-error of ASMs. Patient characteristics such as age, sex, other medical conditions, and previous ASMs may guide selection, but more than 25 ASMs are available, making the selection of treatment challenging. While randomized control trials are the gold standard for understanding treatment effect of ASMs, register data is easier to obtain and can be both updated and retrieved automatically, which are especially useful properties as data for machine learning. The goal of this project is to develop a machine learning algorithm and train it on data from registers to help clinicians to select ASM.

Methods: We used register data of ASM prescriptions from 40015 patients with epilepsy collected between 2005 and 2020. Since patients use multiple treatments over time, the cohort could be divided into 59960 patient instances in total. A lower limit of 100 patient instances was set for each treatment, leading to 8 ASMs. ICD-codes of the patients of 13 different comorbidities were retrieved, including stroke, trauma, and dementia. The length of treatment use for a patient was utilized as outcome of the treatment, which is a surrogate marker of seizure suppression and how severe the side effects are. Patients under the age of 25 were removed to avoid the problems of ASM withdrawal in childhood and juvenile epilepsies. Discontinuation of treatment was defined as more than 12 months without a new prescription and set at 3 months after the last dispensation. From a machine learning perspective, there are two major obstacles. The first is the censored nature of the data, which needs to be handled with a time-to-event approach to avoid inducing bias. The second problem is that the data is observational; there is selection bias since clinicians prescribe different treatments to different kinds of patients. We developed a novel neural network architecture that utilizes representation learning to mitigate selection bias for any number of treatment groups, while also being able to learn from censored data. The network takes age, sex, comorbidities, previous ASMs, and suggested ASM as input, and outputs an estimation of the time-to-event i.e. the termination of ASM use.

Results: Our machine learning algorithm was compared to four time-to-event machine learning baselines by computing the concordance index (CI) on the test set. Our method had a higher performance and achieved a CI of 0.677. The baseline methods and their corresponding CI: Gradient boost: 0.657, Deep survival machines: 0.642, Random survival forest: 0.630, and DeepSurv: 0.489.

Conclusions: We demonstrate that data from registers is a viable data source that can be modeled and analyzed by machine learning algorithms to help in clinical decision-making. Our novel machine learning algorithm based on causal inference shows promising results on real-world observational data compared to the established practices.

Funding: Please list any funding that was received in support of this abstract.: This work was supported by the Knut and Alice Wallenberg foundation.