Screening of drug-related adverse events by sequence symmetry analysis with spline-based self-controlled case-series design for dose-response relationship: a proof of concept study

Background: Application of sequence symmetry analysis (SSA) to large-scale healthcare databases has been implemented to identify signals of drug-related adverse events. However, SSA could produce a great number of signals, and it may not be feasible to examine all the signals by conducting comparative studies with fine-tuned confounding adjustment. Therefore, it is warranted to narrow down the signals, for example, by testing dose-response relationship.

Objectives: In this study, we proposed SSA combined with spline-based self-controlled case-series design (SCCS) that examines dose-response relationship and potentially reduces the number of false-positive signals, and we aimed to compare its performance with that of SSA alone, using antidiabetic medications as examples.

Methods: We used the healthcare claims database provided by the DeSC Healthcare Inc., Japan (2014-2021). New-users of the 7 classes of antidiabetic medications were included in the study. Exposure was defined as the initiation of each medication, and the outcomes were all diagnoses recorded in the database, defined as the first three digits of the ICD-10 codes.

We first conducted SSA alone, with 6-month intervals, and estimated crude sequence ratios and 95% confidence intervals. Outcomes with lower limits > 1.0 were considered signals. Then, in addition to SSA, we assessed the dose-response relationship using spline-based SCCS that models time-varying exposure effects using a linear combination of cubic M-splines. The outcomes with a gradually increasing trend of exposure effect over the risk period were considered signals. We calculated positive predictive values (PPV), defined as the number of identified outcomes already on the drug label in Japan divided by the total number of outcomes identified by each method.

Results: PPVs in SSA alone and SSA with spline-based SCCS, respectively, were 16.3% (7/43) and 27.3% (3/11) for alpha-glucosidase inhibitors, 14.3% (13/91) and 6.3% (2/32) for biguanides, 14.4% (22/153) and 4.3% (1/23) for DPP-4 inhibitors, 24.1% (7/29) and 30.0% (3/10) for GLP-1 receptor agonists, 19.8% (24/121) and 16.0% (8/50) for SGLT2 inhibitors, 13.7% (7/51) and 6.3% (1/16) for sulfonylureas, and 23.1% (9/39) and 20.0% (3/15) for thiazolidinediones, meaning that the number of signals decreased by SSA with spline-based SCCS in all 7 classes, whereas PPV increased in only 2 classes (alpha-glucosidase inhibitors and GLP-1 receptor agonists).

Conclusions: Compared to SSA alone, SSA with spline-based SCCS led to a decreased number of signals. However, the performance in terms of PPV did not improve, suggesting that important signals could be missed by this method. Further studies are warranted to explore various settings of spline-based SCCS to improve the performance.