(166) A Novel, Hypothesis-free Approach to Detecting Safety Signals by Clustering Distributed Representations of Adverse Events Based on Claims Databases

Background: Traditional signal detection approaches based on spontaneous reporting systems suffer from biased reporting, resulting in a high false-positive rate, and cannot effectively utilize information of patients’ backgrounds, other (non-)exposures and co-occurring symptoms. Several new approaches of signal detection based on large-scale database of electronic health records usually require defining pre-specified drug-event pairs to be tested, potentially limiting their applications.

Objectives: We constructed a novel hypothesis-free approach to safety signal detection by clustering distributed representations of co-occurring adverse events (AEs) based on claims records.

Methods: A predictive model of AEs after starting an antidiabetic drug was constructed by each cohort of antidiabetic drug class, based on Bidirectional Encoder Representations from Transformer (BERT). Target antidiabetic drug classes included insulin, DPP-4 inhibitors, SGLT-2 inhibitors, biguanides, and GLP-1 receptor antagonists (RA). Input variables were patient's age, gender, medical history, concomitant/previous treatments with diabetes medications and co-occurring AEs coded in ICD-10 in claims databases (JMDC and DeSC Healthcare) and embedded into a high-dimensional vector space of distribution representation through BERT, considering chronological order of the features. Masked language model was applied to acquire distributed representation of each AE. AEs with the number of occurrences less than 20 were excluded. Ward's method clustered the vector embeddings of AEs derived from the trained BERT model. Silhouette index (SI) and Davies-Bouldin index (DBI) were calculated for formed clusters. A medical review was conducted to validate potential signal clusters with SI≥0, in reference to the package inserts.

Results: The maximum precision micro was 0.34 for 94,395 insulin users, 0.44 for 64,183 DPP-4 inhibitor users, 0.4 for 23,866 SGLT-2 inhibitor users, 0.4 for 40,167 biguanide users, and 0.67 for 20,070 GLP-1 RA users. DBI was approximately 2 for all drug classes. More than 600 AE clusters identified in total and ≥15% of them had the AEs listed in the package inserts. A cluster of genital infection and frequent urination was formed for SGLT-2 inhibitors (SI=0.081). A cluster containing pemphigoid was formed for DPP-4 inhibitors (0.0086). For DPP-4 inhibitors, clusters containing arthritis (0.13), dermatitis such as exfoliative dermatitis (0.11), lung diseases such as interstitial pneumonia and asthma (0.053) were formed. Lactic acidosis cluster was identified for biguanides (0.071).

Conclusions: Distributed representation-based clustering of AEs is a promising approach for a hypothesis-free signal detection method based on electronic health records.