Abstract
This high-throughput screening effectively identified both known and potential new safety signals for SGLT2i. The use of semi-Bayesian shrinkage provides a robust framework for post-marketing surveillance in large healthcare databases.
Introduction
When a new pharmacological agent like a Sodium-Glucose Cotransporter-2 (SGLT2) inhibitor enters the market, its safety profile is initially established through Randomized Controlled Trials (RCTs). While RCTs remain the gold standard for efficacy, they are frequently limited by sample size and duration, often failing to detect rare or long-term adverse events. Traditional post-marketing surveillance has historically relied on spontaneous reporting systems, which suffer from significant underreporting and a lack of a clear denominator for risk calculation.
Large-scale Electronic Health Records (EHRs) offer a powerful alternative for real-world evidence. However, they introduce a significant statistical hurdle: the "multiple testing" problem. When screening thousands of potential side effects simultaneously, the probability of encountering false positives (statistical noise) increases dramatically. This study presents a rigorous, automated high-throughput pipeline designed to scan EHR databases for safety signals while maintaining strict control over false discoveries.
Methodology
The researchers implemented a high-throughput target trial emulation pipeline, designed to mimic the rigor of a clinical trial using observational data. The methodology consisted of four primary components:
- Cohort Design: The study utilized a "new-user" design, comparing patients initiating SGLT2 inhibitors to those starting other second-line diabetes medications. This approach minimizes "confounding by indication," where the reason for prescribing a specific drug might itself be linked to the outcome.
- Propensity Score Weighting: To ensure an "apples-to-apples" comparison, the team applied Inverse Probability of Treatment Weighting (IPTW). This balanced the study and comparator groups across dozens of baseline characteristics, including age, BMI, and various comorbidities.
- Outcome Screening: Rather than focusing on a single pre-specified adverse event, the pipeline screened all ICD-10-CM 4-digit diagnosis codes available in the record.
- Semi-Bayesian Shrinkage: To address the noise of massive multiple testing, the authors applied a semi-Bayesian shrinkage method. This technique "shrinks" extreme hazard ratio (HR) estimates toward a global mean, particularly for rare outcomes with sparse data. A signal that remains statistically significant after this adjustment is far more likely to represent a true safety concern than a random fluctuation.
Data Sources and Population
The study leveraged the OneFlorida+ Data Trust, a massive repository of longitudinal EHR data from Florida and the surrounding regions. The source data spanned 2012–2023, with the specific study cohort focused on 2016–2023. The final analysis included 29,425 adult patients with Type 2 Diabetes, consisting of 10,277 SGLT2i initiators and 19,148 initiators of other second-line medications.
Key Insights and Results
Validation of Known Risks
The pipeline successfully identified genital mycotic infections (candidiasis), a well-documented side effect of SGLT2 inhibitors. SGLT2i use was strongly associated with an elevated risk of candidiasis of the vulva and vagina (HR 1.64; 95% CI 1.42–1.90). This served as a critical "positive control," validating the pipeline's ability to detect established safety signals.
Discovery of New Potential Signals
Beyond known risks, the analysis uncovered several potential new safety signals that warrant further clinical investigation:
- Viral conjunctivitis: HR 1.58 (95% CI 1.09–2.31)
- Melanocytic nevi of the trunk (moles): HR 1.34 (95% CI 1.12–1.60)
- Bacterial intestinal infections: HR 1.40 (95% CI 1.10–1.78)
Addressing Surveillance Bias
The system also flagged outcomes such as "screening for respiratory neoplasms." The authors identified this not as a physiological side effect of the drug, but as surveillance bias. Patients prescribed newer, high-profile medications like SGLT2 inhibitors often receive more frequent medical attention, leading to a higher rate of incidental screenings and findings compared to those on older therapies.
Statistical Robustness
The effectiveness of the semi-Bayesian shrinkage was evident in the data stability. For common outcomes like candidiasis, the shrinkage factor remained high (>0.89), indicating stable raw data. Conversely, for rare outcomes, the factor dropped to approximately 0.41, demonstrating how the method aggressively filters out potential false positives in low-prevalence conditions.
Discussion and Limitations
While this high-throughput approach represents a significant advancement in drug safety monitoring, several limitations remain. As an observational study, residual confounding from unmeasured factors cannot be entirely ruled out. Furthermore, the reliance on ICD-10 codes—primarily used for billing—may lack the clinical nuance of adjudicated outcomes found in formal trials.
It is essential to view these findings as "signals" or hypotheses rather than definitive proof of causality. Future research is required to investigate the biological mechanisms behind findings like viral conjunctivitis. Nevertheless, this study provides a robust, scalable framework for real-time post-marketing surveillance, allowing researchers to move beyond reactive reporting toward proactive, data-driven safety monitoring.
Summary updated