1 in 8
People worldwide lived with obesity in 2022.
A chronic condition that often requires sustained care.
WHO, Obesity and overweight (2025)BIOMEDICAL INFORMATICS · PHARMACOVIGILANCE
Safety signals.
Clinical context.
What can global safety reports and longitudinal health records tell us about GLP-1/GIP treatment?
Tracking the Dynamic Trajectories:
A Global-to-Local Pharmacovigilance Analysis of GLP-1 Receptor Agonists
Accepted at the 2026 AMIA Annual Symposium
Explore the evidence ↓Start with the question ↗THE EVIDENCE PATH2014—2025
CONNECTED BY REPRESENTATIONS
not individual patient linkage↗
not individual patient linkage↗
A RETROSPECTIVE, TWO-SOURCE STUDY
286,461FAERS reports
+71,316EHR patients
One informatics question:
How do we connect the evidence?
How do we connect the evidence?
01 / THE QUESTIONWHY THIS STUDY
Better efficacy is only
part of the story.
Obesity care needs treatment that people can access, tolerate, and continue.
GLP-1/GIP therapies changed expectations for weight loss. In routine care, however, treatment trajectories vary. This study asks how safety signals look when placed in the context of observed treatment duration.
15–21%
Mean weight reduction in pivotal trials.
Separate semaglutide and tirzepatide trials; not a head-to-head comparison.
Wilding et al., NEJM (2021); Jastreboff et al., NEJM (2022)54.9%
Discontinued within one year.
One health-system cohort of 8,184 initiators. Stopping has multiple causes.
Gasoyan et al., Obesity (2025)More clinical context
A continuum of care
Nutrition, physical activity, behavioral support, pharmacotherapy, and metabolic or bariatric procedures can contribute to treatment. Patient response, access, and persistence differ.
Tolerability matters
Nausea, vomiting, diarrhea, constipation, and abdominal discomfort can affect continuation. Other safety concerns include pancreatitis, gallbladder events, and heart-rate changes; labeling varies by drug.
Discontinuation is multifactorial
Among 288 chart-reviewed discontinuers in the cited study, 14.6% stopped because of intolerable side effects. This uses a different denominator from the 54.9% figure.
Presentation slides 3–6. WHO; NIDDK treatment guidance; FDA prescribing information; Gasoyan et al. (2025), as cited in the supplied slides.
THE RESEARCH QUESTION
Which signals recur across drugs—and how do the corresponding EHR outcomes differ across completed treatment-duration groups?
02 / THE EVIDENCE LABFOLLOW A SIGNAL THROUGH THE WORKFLOW
Two data sources.
Three translation steps.
Start with a reporting signal. Follow its clinical representation. Then see how fragmented medication records become a treatment episode.
FAERS · 2014–2025Broad screening for disproportionate reporting
→ SEMANTIC + TEMPORAL REPRESENTATION →
UT Physicians EHR · 2015–2025Diagnosis and medication histories in clinical context
Study design, source roles & cohort definitions
Different sources, different questions
Trials offer structured follow-up and stronger attribution. FAERS prioritizes reporting patterns without a population exposure denominator. EHR data add clinical context but retain confounding and incomplete capture.
Medication scope
The targets are liraglutide, semaglutide, and tirzepatide. The broader baseline includes core AOMs, metabolic medications, and background treatments. These do not all share the same indication or regulatory status for obesity treatment.
The analytic EHR cohort
The supplied diagram starts with 365,091 obesity and 260,987 overweight patients. After the stated exclusions, 610,372 remain; 71,316 have AOM exposure.
Presentation slides 7–14 and 21; manuscript Methods and Tables 1–2. The sources are not linked at the individual-patient level.
03 / THE FINDINGSLOOK AT THE PATTERNS
Duration changes
the context.
Short observed treatment episodes were common.
Duration groups contain different observed episodes. They are not randomized treatment assignments or repeated snapshots of the same patients.
OBSERVED TREATMENT DURATION
51.6%
of GLP-1/GIP episodes fell in the two shortest duration groups.
This is not a patient-level adverse-effect discontinuation rate.
Manuscript Figure 4; presentation slide 23. Source interval labels and rounded percentages are retained.
EVENT EXPLORER
Inspect one phenotype at a time.
Choose an outcome, then select a duration group.
Counts describe recorded events—not incidence.
RECORDED EVENT COUNT
GLP-1-involved rows transcribed from manuscript Figure 5. No new rates or PRR estimates are calculated.
How to read these counts
Completed duration
The x-axis groups episodes by completed treatment duration. It does not show when an event occurred within the same patients' follow-up.
Prior recorded diagnoses
The study applies a phenotype-specific 365-day lookback before episode start. An absent prior code is not proof of absent prior disease.
Denominator definitions
The text reports 28,961 filtered GLP-1/GIP records, while Figure 5 column Ns sum to 33,647. This demo preserves source counts and does not derive new rates from these denominators.
THE MODEL ADDS A COMPARISON
Patterns differ across outcomes.
Nausea and vomiting
An elevated reference-stratum treatment contrast accompanies lower interaction estimates in longer-duration strata.
TOLERABILITY PATTERNHepatic steatosis
A higher interaction estimate in the longest-duration group is exploratory and requires replication.
HYPOTHESIS-GENERATINGAn interaction is not the same as a treatment odds ratio.Open a mathematical illustration of the model contrast.
ILLUSTRATION · NOT FITTED STUDY ESTIMATES
Combine the two terms.
A same-duration treatment contrast combines the reference-stratum treatment term and the relevant interaction.
ORsame duration = exp(β₁ + β₃)
= ORreference × ORinteraction
= ORreference × ORinteraction
The source model includes treatment, categorical duration, and their interaction. It does not include dose.
COMBINED CONTRAST
1.9 × 0.6
An interaction below 1 can coexist with a treatment OR above 1. A confidence interval requires coefficient covariance; none is reconstructed here.
Presentation slides 25–27; manuscript regression section and Figure 6. The results remain associative.
04 / INTERPRETATIONWHAT THE PATTERNS CAN—AND CANNOT—TELL US
One pattern.
Several explanations.
A lower association in longer-duration groups does not, by itself, tell us why it is lower.
These are competing hypotheses, not mechanisms established by the study. The hepatic finding does not establish drug-induced liver disease.
Presentation slides 27–29 and 31; manuscript Discussion and Limitations.
05 / BIOMEDICAL INFORMATICSWHAT THIS WORK CONTRIBUTES
The contribution
is the connection.
Existing tools supply components. Study-specific definitions and integration decisions turn heterogeneous records into interpretable analysis units.
Presentation slide 30; manuscript Methods.
Cross-source evidence architecture
Keep reporting signals and EHR associations distinct while defining a traceable handoff between them.
Semantic interoperability
Review candidate MedDRA–UMLS–ICD mappings, exclusions, and consolidations to construct 14 EHR phenotypes.
Temporal exposure phenotyping
Convert fragmented medication records into 289,655 treatment episodes using explicit reconstruction rules.
Interpretable model contrasts
Locate differences across completed-duration strata while distinguishing treatment terms from interactions.
Where uncertainty remains
Reporting & normalizationUnderreporting, stimulated reporting, and classification errors affect the global screen.
Semantic granularitySymptoms and coded diagnoses are not one-to-one; mappings can exclude or consolidate concepts.
Signal scopeThe shared intersection omits drug-specific signals.
Clinical captureOne health system and out-of-network care limit generalizability.
Residual confoundingIndication, concurrent treatments, dose, and selection complicate causal interpretation.
Presentation slide 31; manuscript Limitations.