Predictive Modeling of Weight Loss and Metabolic Health Outcomes: A Retrospective Predictive Modeling Study

Abstract

Background

Obesity is a chronic, complicated, and progressive disease that significantly affects mortality, quality of life, and overall health in nearly 13% of the adult population worldwide. Thus, solutions like a hypocaloric diet with a Mediterranean diet pattern aim to control this and other metabolic problems.

Objectives

This study developed and measured the performance of different machine learning (ML) models designed to predict body weight loss and/or metabolic syndrome (MetS) change after a 3-month hypocaloric diet with a Mediterranean pattern in obesity-diagnosed patients.

Methods

The data set was provided by a clinical trial of 893 obese patients. Five machine learning architectures were implemented: Logistic Regression, Decision Tree Classifier, Random Forest Classifier, eXtreme Gradient Boosting Classifier (XGBoost), and Support Vector Classifier. Performance metrics such as accuracy, precision, recall, F1-score, and ROC curve were used to assess the prediction models. The influence of the predictors was also evaluated in every case.

Results

For body weight loss prediction, Stacking and Random Forest models outperformed the other models, with accuracies of 81.37% and 76.44% and AUC of 86.79% (95% CI: 82.9%–90.4%) and 86.25% (95% CI: 82.3%–89.9%), respectively. For MetS change, Stacking had the best performance, with an accuracy of 85.90% and an AUC of 83.65% (95% CI: 76.9%–89.8%). For the prediction model of body weight loss and MetS change, Stacking was the best algorithm, with an accuracy of 94.74% and an AUC of 95.35% (95% CI: 88.7%–99.9%). Furthermore, variables associated with metabolic and inflammatory markers exhibited stronger correlations with the outcomes.

Conclusion

Machine learning models, especially ensemble methods like Stacking and XGBoost, effectively predict body weight loss and MetS improvement in obese patients following a Mediterranean diet. Key predictors include age, insulin resistance markers, and inflammatory biomarkers. Integrating these predictive tools can significantly enhance personalized dietary interventions, optimizing treatment outcomes in clinical practice.