Predicting who responds would be commercially valuable to every seller on this desk, and clinically useful to everyone else. A study trained seven machine learning approaches on real-world records to try [1]. How much response actually varies is documented in the super-responder brand comparison.
The headline is a weight-loss model with discrimination around 0.94 and accuracy of 0.89 to 0.90, which in most clinical prediction contexts would be exceptional. The same methods on the same data predicted glycemic control at around 0.79 discrimination and 0.73 accuracy — respectable, and far short of the first figure. Expressed as plain accuracy that is 89% vs 73% of predictions correct.
The gap between the two models is the more informative result. Glycemic control is the harder target and the less self-referential one, and it is where the honest ceiling of this approach shows: roughly three predictions in four correct, using duration of diabetes, baseline HbA1c and existing sulfonylurea or insulin use as the main features. That is a useful tool and not a crystal ball — and a far more honest number than the weight model's, for the same reason the derived figures in the prediabetes screening analysis need their predictive values read rather than their sensitivities.
One methodological limit matters commercially. The models were assessed with ten-fold cross-validation, which resamples the same dataset, and no external validation on a separate population is described. A model can score highly on the data it grew up in and fail on patients from a different health system — which is precisely what would happen if a seller took a published model and applied it to their own customers.
There is also nothing here a buyer can act on. The predictors are baseline characteristics, so learning that your starting BMI predicts your response does not give you anything to change. The measurements that do shift with treatment, and what they mean, are handled in responding is not one thing, and the question of how early a real response becomes visible in slow responders at week eight.