Predicting successful long-term weight loss from short-term weight-loss outcomes: new insights from a dynamic energy balance model (the POUNDS Lost study)

Abstract

Background: Currently, early weight-loss predictions of long-term weight-loss success rely on fixed percent-weight-loss thresholds.

Objective: The objective was to develop thresholds during the first 3 mo of intervention that include the influence of age, sex, baseline weight, percent weight loss, and deviations from expected weight to predict whether a participant is likely to lose 5% or more body weight by year 1.

Design: Data consisting of month 1, 2, 3, and 12 treatment weights were obtained from the 2-y Preventing Obesity Using Novel Dietary Strategies (POUNDS Lost) intervention. Logistic regression models that included covariates of age, height, sex, baseline weight, target energy intake, percent weight loss, and deviation of actual weight from expected were developed for months 1, 2, and 3 that predicted the probability of losing <5% of body weight in 1 y. Receiver operating characteristic (ROC) curves, area under the curve (AUC), and thresholds were calculated for each model. The AUC statistic quantified the ROC curve's capacity to classify participants likely to lose <5% of their body weight at the end of 1 y. The models yielding the highest AUC were retained as optimal. For comparison with current practice, ROC curves relying solely on percent weight loss were also calculated.

Results: Optimal models for months 1, 2, and 3 yielded ROC curves with AUCs of 0.68 (95% CI: 0.63, 0.74), 0.75 (95% CI: 0.71, 0.81), and 0.79 (95% CI: 0.74, 0.84), respectively. Percent weight loss alone was not better at identifying true positives than random chance (AUC ≤0.50).

Conclusions: The newly derived models provide a personalized prediction of long-term success from early weight-loss variables. The predictions improve on existing fixed percent-weight-loss thresholds. Future research is needed to explore model application for informing treatment approaches during early intervention.

Trial registration: ClinicalTrials.gov NCT00072995.

Keywords: dynamic model; energy balance; likelihood function; mathematical model; receiver operating characteristic; regression; weight loss.

© 2015 American Society for Nutrition.

Figures

FIGURE 1

The ROC curves in month 1 (A), month 2 (B), and month 3 (C) based on prediction of a binary outcome of 5% weight loss in 1 y. The solid dark curves use a probability function derived from a logistic regression model that includes covariates of age, baseline weight, height, sex, target intake, distance between actual and expected weight, and percent weight loss available by each month. The long-dashed medium dark curves represent the ROC curves derived from the leave-out-one cross-validation. The dashed curves are based solely on percent weight loss. All dashed curves have an AUC below 0.50. The AUC for the probability function–derived curves in months 1, 2, and 3 were 0.68, 0.75, and 0.79, respectively. The AUCs for the leave-one-out cross-validation ROC curves were 0.66, 0.73, and 0.77 respectively. ROC, receiver operating characteristic.