Machine learning and big data for personalized epilepsy treatment

Abstract

Finding an effective anti-seizure medication (ASM) with minimal side effects is a challenge. Patient characteristics are used to guide treatment selection, but about half of the patients with epilepsy do not achieve seizure freedom with their first ASM. While randomized controlled trials are the gold standard for estimating treatment efficacy, they may not always be clinically relevant, especially for rare conditions. Registers are valuable sources of data because they can contain many patients, are accessible, and are updated regularly. The aim of the present research is to evaluate registers and develop machine learning algorithms for personalized medicine in epilepsy.

We used prescriptions, in- and outpatient data, and mortality data from national Swedish registers to model ASM use of patients. As a bundled estimation of efficacy and tolerability, retention rate was used as the measure of outcome.

The results indicate that using register data to estimate retention of ASMs is feasible and personalized ASM selection can potentially improve patient outcomes. Retention rates from registers are similar to that of RCTs and meta-analyses of RCTs. In an analysis of patients with epilepsy and comorbidities, there was a potential improvement of 14-21% of the 5-year retention rate for the initial ASM (Paper I). Ranking of ASMs for patient cases based on retention rates from register data is similar to suggestions based on expert advice (Paper II). We also studied ASM use in children, a group with limited evidence (Paper III). Specialized machine learning algorithms can potentially be a useful source of information for doctors for selecting ASMs (Paper IV).

In conclusion, this research highlights the potential of registers as a data source for personalized medicine. Machine learning trained on register data can be used to predict the efficacy of ASMs, but the methodology needs further development and clinical verification.