A new predictive model uses early data to forecast long-term GLP-1 weight loss, exposing a massive real-world retention crisis.

Why do we evaluate weight-loss drugs using only the patients who finish the trial? In clinical practice, most people quit. By ignoring those who drop out, standard medical literature creates a distorted picture of how these drugs actually perform in the wild.

This disconnect is the real story. Clinicians need to make decisions at the two-month mark, yet they lack data on who will actually benefit. A new study tackles this head-on by building a predictive tool designed for the exact moment a patient sits down for their week-eight visit. It shifts the clinical focus from trial perfection to real-world pragmatism.

The real-world dropout crisis

The researchers analyzed data from 237,800 adults enrolled in a US telehealth program. To ensure data quality, they filtered the group down to an analytic cohort of 22,538 patients who had documented week-eight weights, refill-confirmed doses, and reported ethnicity. Among those who actually completed six months of treatment, the results were strong. Patients lost an average of 11.7% of their body weight on semaglutide and 14.1% on tirzepatide.

But the data also revealed a stark reality. A staggering 66% of patients disengaged from the program before the six-month mark. This high dropout rate means traditional clinical trials, which focus almost entirely on completers, fail to reflect the typical patient experience. Evaluating drug efficacy without accounting for this attrition is a major clinical blind spot.

Why week eight matters

The researchers chose the week-eight mark because 80% of slow responders reach their post-titration decision point by this visit. Earlier weeks are too volatile. Predictions at weeks two through six yielded weak R2 values between 0.48 and 0.61. While later weeks offered higher accuracy, waiting until week 10 or 12 means the clinical decision has already been made. Week eight is the sweet spot where the data becomes predictive and the clinician can still act.

• The week-eight model achieved a test R2 of 0.65 with a mean absolute error of 2.76 percentage points.
• The 80% quantile-regression interval covered 76% of test patients, while the 95% interval covered 93%.
• A logistic regression model predicted patient dropout with an AUC of 0.79, outperforming gradient boosting at 0.74.
• For patients who remained engaged, the median time to reach a weight plateau was 387 days.

The limits of prediction

This tool is a major step toward realistic clinical expectations, but it is not a silver bullet. The researchers are honest about its limitations. The model is not yet externally validated. Because it was built on data from a single telehealth platform, other clinics should treat it as a target for recalibration rather than a plug-and-play solution. Furthermore, two of the twenty subgroup cells showed reduced predictive accuracy, and another two were too sparse to validate at all.

Ultimately, this tool is prognostic, not therapeutic. It can help clinicians identify patients who are likely to struggle or drop out, but it cannot make the final clinical decision for them. It simply gives providers the data they need to have honest conversations about expectations early in the treatment cycle.

Read the full preprint at medRxiv.