Scholarlink

Purpose <p>The relationship between dietary nutrient intake and metabolically healthy obesity (MHO) remains poorly understood. This study aimed to construct machine learning models to predict MHO based on dietary nutrient profiles and to identify the most influential nutrients contributing to this phenotype.</p> Methods <p>Data were derived from the U.S. National Health and Nutrition Examination Survey (NHANES) 2005–2018. Forty-five dietary nutrients, along with demographic and lifestyle variables, were included in two predictive frameworks: a dietary-only model and a complete model. Feature preprocessing involved assessing mixture effects, removing multicollinear variables, addressing class imbalance, and selecting important predictors. Six machine learning algorithms—random forest (RF), light gradient-boosting machine, k-nearest neighbor, Naive Bayes, support vector machine, and eXtreme Gradient Boosting (XGBoost)—were developed and benchmarked to compare performance. Model interpretability was examined using SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME).</p> Results <p>A total of 8914 participants, including 475 classified as having MHO, were analyzed. The Random Forest model exhibited the best predictive performance in the complete model, achieving training and validation AUCs of 0.986 and 0.991, respectively. In contrast, XGBoost demonstrated superior performance in the dietary-only model, with AUCs of 0.971 and 0.988. SHAP and LIME analyses revealed that added vitamin B12, lycopene, caffeine, theobromine, and lutein/zeaxanthin were the strongest positive predictors in the complete model. When only dietary factors were considered, lycopene, lutein/zeaxanthin, magnesium, potassium, and selenium emerged as the most influential nutrients.</p> Conclusions <p>RF and XGBoost models provided the highest predictive accuracy for MHO using complete and dietary feature sets, respectively. The consistent findings from SHAP and LIME analyses emphasized lycopene and lutein/zeaxanthin as reliable and biologically relevant key predictors of metabolically healthy obesity.</p> <p><i>Level of Evidence</i>: Level III, well-designed cohort or case–control analytic study.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Prediction of metabolically healthy obesity based on dietary nutrients: a comparative analysis of six machine learning models with SHAP and LIME interpretation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月9日

查看全文
  • Chenyi Ji,
  • Zhijian Qin,
  • Yucheng Yang,
  • Yao Shen,
  • Jie Gao,
  • Fangrun Zhu,
  • Fang Liu

摘要

Purpose

The relationship between dietary nutrient intake and metabolically healthy obesity (MHO) remains poorly understood. This study aimed to construct machine learning models to predict MHO based on dietary nutrient profiles and to identify the most influential nutrients contributing to this phenotype.

Methods

Data were derived from the U.S. National Health and Nutrition Examination Survey (NHANES) 2005–2018. Forty-five dietary nutrients, along with demographic and lifestyle variables, were included in two predictive frameworks: a dietary-only model and a complete model. Feature preprocessing involved assessing mixture effects, removing multicollinear variables, addressing class imbalance, and selecting important predictors. Six machine learning algorithms—random forest (RF), light gradient-boosting machine, k-nearest neighbor, Naive Bayes, support vector machine, and eXtreme Gradient Boosting (XGBoost)—were developed and benchmarked to compare performance. Model interpretability was examined using SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME).

Results

A total of 8914 participants, including 475 classified as having MHO, were analyzed. The Random Forest model exhibited the best predictive performance in the complete model, achieving training and validation AUCs of 0.986 and 0.991, respectively. In contrast, XGBoost demonstrated superior performance in the dietary-only model, with AUCs of 0.971 and 0.988. SHAP and LIME analyses revealed that added vitamin B12, lycopene, caffeine, theobromine, and lutein/zeaxanthin were the strongest positive predictors in the complete model. When only dietary factors were considered, lycopene, lutein/zeaxanthin, magnesium, potassium, and selenium emerged as the most influential nutrients.

Conclusions

RF and XGBoost models provided the highest predictive accuracy for MHO using complete and dietary feature sets, respectively. The consistent findings from SHAP and LIME analyses emphasized lycopene and lutein/zeaxanthin as reliable and biologically relevant key predictors of metabolically healthy obesity.

Level of Evidence: Level III, well-designed cohort or case–control analytic study.