A new study published in Obesity Science & Practice set out to test whether a surprisingly low-tech solution could close that gap: asking study participants to photograph their medication. The research, conducted by a team at the University of Michigan, evaluated the feasibility of a photo-review protocol for online GLP-1RA studies, in which participants upload deidentified images of their injector pens, vials, labels, or prescription documents. The goal was to determine whether such images could be collected reliably, coded systematically by trained reviewers, and compared against what participants simply claimed to be taking. The answer, on all three counts, was yes, and the discrepancies the method uncovered are striking enough to matter for anyone interpreting online survey data on these blockbuster drugs.

The study design was deliberately rigorous for an online survey. Researchers recruited through Prolific, an online research platform, applying prescreening filters that restricted access to United States residents with a 99 percent or higher approval rate, at least 50 prior submissions, and self-reported current use of Ozempic or another GLP-1RA. Prospective participants were told upfront that they would need to upload a photo of their medication. A screening survey assessed whether they used an injectable semaglutide or tirzepatide product, took it once weekly, and had been on it for at least three months. Eligible participants then completed a longer main survey covering demographics, anthropometrics, prescriber type, payer type, and self-reported medication identity.

The privacy procedures were a central feature of the protocol rather than an afterthought. Participants were instructed to show the medication label while hiding personal identifying information such as names and dates of birth. A study author then reviewed every uploaded image, edited out any identifiers that slipped through, deleted the unredacted originals, and stored only redacted copies in a secure folder. For the coding stage, two trained research staff members who had no role in eligibility screening and were blinded to participants’ self-reports independently classified each photo into one of five categories: Ozempic, Wegovy, Mounjaro, Zepbound, or a compounded semaglutide or tirzepatide product. Any disagreement between the two coders went to an independent adjudicator who had not seen the images before and was also blinded to self-reports.

The mechanics of classification were straightforward but consequential. A photo was coded as a branded, FDA-approved formulation when the medication itself, its labeling, or associated records displayed the brand name. It was coded as compounded when the image showed semaglutide or tirzepatide as the active ingredient but no brand name appeared anywhere. Of the 419 completed screening responses, 47 participants were excluded for reasons ranging from insufficient duration of use to uploading photos of an ineligible drug, most commonly Trulicity, or submitting images that appeared to be stock photos. Twenty-two participants were excluded on the basis of their medication photos alone, a demonstration that the review process had real screening power. The final analytic sample comprised 322 adults with a mean age of about 45 years and a mean pre-treatment body mass index of roughly 40.

On the feasibility question, the results were emphatic. Uploaded photos generally contained enough information for coders to make a determination, and intercoder agreement before adjudication was almost perfect, with a Cohen’s kappa of 0.98. Only five discrepancies arose between the two primary coders, including one foreign-language injector pen label and several branded products, and all were resolved through adjudication. In an era when online health surveys are often dismissed as noisy or unverifiable, the demonstration that ordinary participants can and will submit usable photographic documentation of their medications is itself a methodological advance.

The more provocative findings emerged when photo-based classifications were compared with self-report. Overall agreement between the two methods was substantial, at 82.61 percent with a kappa of 0.78, but the pattern of disagreement was highly uneven. Among the 83 participants whose photos were coded as compounded semaglutide or tirzepatide products, more than a quarter of the entire sample, 65 percent self-reported using a branded medication such as Ozempic, Wegovy, Mounjaro, or Zepbound. Only 29 of the 83 compounded users, or about 35 percent, gave answers consistent with compounded use. By contrast, discrepancies between two branded products were almost nonexistent, amounting to just two cases in the whole sample.

When medication types were collapsed into branded versus compounded groups, the diagnostic statistics told a sharper story. Self-report showed perfect specificity and perfect positive predictive value, meaning that everyone who claimed a compounded product genuinely had one, and no branded user was misclassified as compounded. But sensitivity was only about 35 percent, and the negative predictive value was roughly 82 percent, indicating that a large share of compounded users were invisible to self-report. In practical terms, an online study relying purely on participant claims would have underestimated compounded GLP-1RA use by nearly two-thirds within this sample, potentially biasing any conclusions about effectiveness, side effects, or user experience.

The authors are careful to note that the survey instrument itself may share some blame. The medication-type item offered only the four FDA-approved brand names, an unsure option, and a write-in field labeled none of the above; there was no explicit response option for compounded products. Because all participants had already confirmed injectable semaglutide or tirzepatide use, write-in answers were treated as concordant with compounded classification. The researchers recommend that future studies include explicit compounded response options, pilot-test how participants understand terms like compounded and non-branded, and use cognitive interviewing or randomized wording experiments to refine the instruments. They also advise warning participants in advance about photo requirements, listing acceptable documentation forms, and considering photographs taken with the survey visible in the background to deter fraudulent submissions.

Several limitations temper the conclusions. The stringent prescreening filters identified only about 1,200 eligible participants from a pool of more than 300,000 verified platform users, and highly experienced survey takers may be unusually attentive and compliant, meaning the observed discordance could be conservative. Photo-based classification also cannot confirm the authenticity or chemical contents of a medication; it verifies only what is visible on labels and documentation, and it was not validated against dispensing records or laboratory assays. Still, the protocol draws on established clinical traditions, including brown bag medication reviews and the practice of verifying contraceptive pill packs in reproductive health research, and adapts them to a scalable online format. As GLP-1RA use continues to spread across telehealth, med spas, and compounded channels, the study suggests that a simple photograph, reviewed blindly and coded carefully, may become an essential tool for keeping online medication research honest.

Subject of Research: Feasibility of a medication photo-review protocol for classifying GLP-1 receptor agonist use in online survey research

Article Title: Improving Medication Classification in GLP‐1RA Research: Feasibility of a Photo‐Review Protocol

Article References: Katz, J. M., Gearhardt, A. N., & Griauzde, D. H. (2026). Improving Medication Classification in GLP‐1RA Research: Feasibility of a Photo‐Review Protocol. Obesity Science & Practice, 12(5), Article e70200. https://doi.org/10.1002/osp4.70200

Image Credits: AI Generated

DOI: 10.1002/osp4.70200

Keywords: GLP-1 receptor agonists, semaglutide, tirzepatide, compounded medications, medication photo review, online surveys, self-report validity, obesity research, telehealth, medication classification, research methods, Ozempic