A New Artificial Intelligence Guided Method for Identifying Add-on and Switch in Secondary Data Sources: A Case Study on Antiseizure Medications in Danish Registries

Background: In a recent review evaluating the quality of reporting of exposure in pharmacoepidemiological studies, biases have been found in algorithms developed for identifying therapeutic switching and add-ons in secondary data sources.

Objectives: 1) To develop a new algorithm for identifying switches and add-ons in secondary data sources that overcomes the limitations identified in other algorithms;
2) To test the algorithm on antiseizure medication users aged 65 or older during their first 2 years of treatment after developing epilepsy (new user design).

Methods: Antiseizure medication users aged 65 or older were identified in Danish registers between 1996 and 2018. For each individual and separately for each antiseizure medication, the algorithm assessed the start and the end of each redeemed prescription of antiseizure medications (medication events) during the follow-up period. After, it calculated co-exposure time as the number of days from the first day of overlapping of medications events of different antiseizure medications (i.e., different ATC codes) to the last day in treatment or the end of the follow-up whichever came first and accounting for censoring. ICD codes of epilepsy diagnosis, duration of epilepsy hospitalization, duration of medication events, the year of inclusion in the cohort, the direction of the switch/add-on, and time to the first co-exposure were also considered by the algorithm, including pharmacological considerations on which combinations of antiseizure medications can be combined based on recommendations from clinical guidelines and expert consensus. All these variables were used to train a classification model benchmarking six different machine learning/deep learning techniques: 1) Linear Regression, 2) Naïve Bayes, 3) Support Vector Machine, 4) Neural Network, 5) Classification and Regression Tree, and 6) Random Forest. Overall accuracy in 8-fold cross-validation has been used to evaluate the algorithm performance in training (75% of the data) and test sets (25% of the data).

Results: 15870 individuals were included in the study population of which, 988 had 1485 co-exposure to 2 or more antiseizure medications during the follow-up. The algorithm had an overall accuracy of 88% and 92% for classifying 296 and 1189 co-exposure of antiseizure medications as switches and add-ons, respectively.

Conclusions: The algorithm correctly classified 9 out of 10 co-exposure of antiseizure medications as switches or add-ons resulting in a promising tool for data-driven identification of switches/add-ons in secondary data sources.