Personalized predictions to identify individuals most likely to achieve 10% weight loss with a lifestyle intervention.

Abstract

Objective: The objective of this study is to generate an algorithm for making predictions about individual treatment responses to a lifestyle intervention for weight loss to maximize treatment effectiveness and public health impact.

Methods: Using data from Action for Health in Diabetes (Look AHEAD), a national, multisite clinical trial that ran from 2001 to 2012, and machine-learning techniques, we generated predicted individual treatment effects for each participant. We tested for heterogeneity in treatment response and computed the degree to which treatment effects could be improved by targeting individuals most likely to benefit.

Results: We found significant individual differences in effects of the Look AHEAD intervention. Based on these predictions, two-thirds of the sample was predicted to experience a treatment effect within ±2% weight loss from the average treatment effect. If the treatment was targeted to the 69% of patients expected to meet a 7% weight-loss target at 1-year follow-up, the average treatment effect increases, with 10% average observed weight loss in the intervention group.

Conclusions: The Look AHEAD intervention would achieve a 10% average weight reduction if targeted to those most likely to benefit. Future research must seek external validation of these predictions. We make this algorithm available with instructions for use to demonstrate its potential capacity to inform shared decision-making and patient-centered care.