Predicting drug-resistant epilepsy - A machine learning approach based on administrative claims data.

Area of Science:

• Computational Medicine
• Neurology
• Health Informatics

Background:

• Drug-resistant epilepsy (DRE) poses significant risks, but early identification and specialist referral are often delayed.
• Predicting DRE risk at treatment initiation is crucial for timely, personalized interventions and improved patient outcomes.
• Current methods lack the ability to proactively identify patients at high risk for DRE.

Purpose of the Study:

• To assess the feasibility of developing machine learning algorithms for predicting DRE.
• To identify patients at high risk of DRE using longitudinal healthcare claims data.
• To enable earlier intervention and personalized treatment strategies for DRE.

Main Methods:

• Analysis of longitudinal US pharmacy, medical, and hospital claims data from 1,376,756 patients (2006-2015).
• Development and training of machine learning models (including random forest) using 1270 patient features.
• Validation of predictive models against a benchmark model using age and sex.

Main Results:

• The best random forest model achieved an AUC of 0.764, significantly outperforming the benchmark (0.657).
• The model predicted DRE risk up to two years before patients met criteria for drug resistance (failure of ≥3 antiepileptic drugs).
• Predicted probabilities for DRE were well-calibrated with observed frequencies in the test dataset.

Conclusions:

• Machine learning models utilizing claims data can effectively predict patients at risk of DRE.
• Early prediction enables timely referral to specialist care and more aggressive therapeutic interventions.
• This approach has the potential to reduce the severe morbidity and mortality associated with DRE.