Liraglutide and Unexpected Psychiatric Adverse Events in Patients with Obesity: A Pharmacovigilance Study with Machine Learning Approach

Background: Liraglutide (Saxenda®) demonstrated its superior efficacy for weight management and well-tolerated safety in clinical trials. However, clinical trials involve studying in a relatively small number of selected individuals for a short period. Unexpected adverse drug reactions (ADRs), not identified in clinical trials, can emerge once liraglutide has been used by a large population and over a long period.

Objectives: To detect unexpected adverse events (AEs) that may be attributed to liraglutide in the Korea Institute of Drug Safety & Risk Management-Korea adverse event reporting system database (KIDS-KD)

Methods: We investigated the safety signals of liraglutide by implementing a gradient boosting machine (GBM) algorithm in the KIDS-KD from 2017 to 2020. To identify candidates of unexpected AEs, our investigators reviewed drug labels of liraglutide and other drugs in the same therapeutic class (GLP-1 agonists and DPP-4 inhibitors) in the Ministry of Food and Drug Safety. Based on the reviewed labels, we classified 189 AEs reported with liraglutide into three ADR categories: (a) label-positive; (b) label-negative; (c) unknown. Given label-positive and negative ADRs, we developed prediction models based on demographic and clinical characteristics, causality assessment by clinical experts, and statistical indices. Then, we validated the performance of prediction models by measuring the area under the receiver operating characteristics curve (AUROC) and precision-recall curve (AUPRC). The optimal model, achieving the highest scores of AUROC and AUPRC, was implemented to detect safety signals among unknown ADRs.

Results: Among a total of 972,193 reports, 1,201 reports were reported with liraglutide in South Korea. In the phase of model development and validation, we generated the optimal prediction model with the comparable performance on the classification of safety signals (AUROC: 0.8558, AUPRC: 0.9538). Our model detected 33 of 56 unknown ADRs as safety signals including psychiatric disorders, infections, and musculoskeletal disorders. Six safety signals indicated that liraglutide may attribute to psychiatric disorders (suicide attempts, depression, aggression, affect lability, apathy, and derealisation disorder).

Conclusions: Although pre-approval clinical trials showed that liraglutide is associated with a relatively lower psychiatric risk than other anti-obesity drugs, our study identified six potential ADRs related to psychiatric events during post-approval use of liraglutide. Given that liraglutide is a centrally-acting appetite suppressant that may affect mental health, further investigations are needed to confirm the psychiatric safety of liraglutide.