Reading the notes, fewer GLP-1 patients stayed on the drug than scripts implied
A prescription is not a swallowed pill, and a refill order is not proof anyone is still taking anything. That gap is where a new study went looking, and it found that the standard way of counting who stays on GLP-1 drugs overstates the number by roughly a quarter.
Researchers at nference built a large language model pipeline to read the free-text clinical notes doctors type. That is the part of the record that says whether a patient actually started a drug, hit a snag, or quit. The team drew on a de-identified US records network covering more than 29 million patients. They ran the pipeline across 553,073 adults who had a semaglutide ↗ or tirzepatide ↗ prescription and a baseline weight on file, then checked the machine against physician-adjudicated records. It matched the human reviewers 98.4 percent of the time on sorting the notes. On the harder call, whether someone was on, off, or starting a drug, it hit 88.2 percent. The work was published June 24 in Biology Methods and Protocols ↗.
The headline gap is in persistence, whether people are still on the drug months later. Read from prescription records, 55.4 percent of semaglutide patients and 58.3 percent of tirzepatide patients looked like they were still on treatment at 18 months. Read from the notes, those numbers fell to 42.3 and 43.1 percent. The two methods draw materially different survival curves from the same patients, and the prescription-based one sits higher the whole way. A script keeps counting a patient as persistent after they have quietly stopped. The note catches the stop.
Where the two agree is at the start. Among patients with a documented initiation, most began without recorded trouble. Frictionless starts made up 70.1 percent of semaglutide and 77.9 percent of tirzepatide cases. When people did fall off, the notes gave reasons that no prescription line carries: insurance and cost barriers, holds before surgery, and gastrointestinal side effects were the ones that recurred.
The odder finding is about compounded versions, the copies mixed by pharmacies during the shortage years. The pipeline flagged compounded exposure in 1,696 patients, and in about half of them the first sign of compounding showed up on or before the first structured prescription. Some patients were on a compounded GLP-1 drug before the record shows them getting a real one.
None of this changes what the drugs do. It changes how much to trust the real-world numbers built on top of them. Payers, safety monitors, and health systems lean on persistence, adherence, and access figures pulled from claims and prescription feeds. Those figures decide who is counted as benefiting and who is being left on the shelf. If the feeds run about 13 points high on persistence, the picture of how well GLP-1 therapy actually sticks is rosier than the ground truth. Reading the notes by hand is slow and expensive. The argument here is that an audited model can do it across tens of millions of patients, and that the corrected number is worth the cost.