Predicting weight loss success on a new Nordic diet: an untargeted multi-platform metabolomics and machine learning approach

Abstract

Background and aim: Results from randomized controlled trials indicate that no single diet performs better than other for all people living with obesity. Regardless of the diet plan, there is always large inter-individual variability in weight changes, with some individuals losing weight and some not losing or even gaining weight. This raises the possibility that, for different individuals, the optimal diet for successful weight loss may differ. The current study utilized machine learning to build a predictive model for successful weight loss in subjects with overweight or obesity on a New Nordic Diet (NND).

Methods: Ninety-one subjects consumed an NND ad libitum for 26 weeks. Based on their weight loss, individuals were classified as responders (weight loss ≥5%, n = 46) or non-responders (weight loss <2%, n = 24). We used clinical baseline data combined with baseline urine and plasma untargeted metabolomics data from two different analytical platforms, resulting in a data set including 2,766 features, and employed symbolic regression (QLattice) to develop a predictive model for weight loss success.

Results: There were no differences in clinical parameters at baseline between responders and non-responders, except age (47 ± 13 vs. 39 ± 11 years, respectively, p = 0.009). The final predictive model for weight loss contained adipic acid and argininic acid from urine (both metabolites were found at lower levels in responders) and generalized from the training (AUC 0.88) to the test set (AUC 0.81). Responders were also able to maintain a weight loss of 4.3% in a 12 month follow-up period.

Conclusion: We identified a model containing two metabolites that were able to predict the likelihood of achieving a clinically significant weight loss on an ad libitum NND. This work demonstrates that models based on an untargeted multi-platform metabolomics approach can be used to optimize precision dietary treatment for obesity.

Keywords: machine learning; metabolomics; new Nordic diet; obesity; precision nutrition.

Figure 1

Percentage weight change of the participants completing the 26 weeks intervention following a New Nordic diet. Responders had a weight loss ≥5% (green area) and non-responders had a weight loss <2% (pink area).

Figure 2

Data sets used in this study: label information corresponding to each subject (0 = non-responder and 1 = responder), metadata including clinical variables, metabolomics measurements of LC–MS (both positive and negative ionization mode) and NMR analysis.

Figure 3

(A) Model signal path for the success of weight loss on an NND diet, (B) 2D response of the model predictions with training and test data overlaid. The decision boundary separates the response areas. The dots represent the subjects of the classified responders (green) and non-responders (pink) whereas the background represents the model’s prediction (1 = responder, 0 = non-responder), and (C) Receiver operator characteristic (ROC) for training and test set.

Figure 4

Heatmap of adipic acid and argininic acid correlation with baseline levels of clinical variables using Pearson correlations. Intensity of the blue and red colors indicate the strength of negative and positive correlations, respectively. *p < 0.05 and **p < 0.005. HOMA, homeostatic model assessment for insulin resistance; HDL, high-density lipoprotein; LDL, low-density lipoprotein.

Figure 5

Levels of adipic acid and argininic acid for responders (green) and non-responders (pink) at baseline, week 12 and week 26. Error bars represent SEM. *p < 0.05 between responders and non-responders at baseline. ††p < 0.01 and †††p < 0.0001 are changes from baseline to week 26 within each group over the time of intervention.

Figure 6

Changes in body weight (in percent from initial weight) for the responders (green) and the non-responders (pink) during 26 weeks of intervention (full line) followed by an additional 52 week follow-up period (dashed line). Error bars represent SEM.