Görmez Yasin, Yagin Fatma Hilal, Yagin Burak, Aygun Yalin, Boke Hulusi, Badicu Georgian, De Sousa Fernandes Matheus Santos, Alkhateeb Abedalrhman, Al-Rawi Mahmood Basil A, Aghaei Mohammadreza
Department of Management Information Systems, Faculty of Economics and Administrative Sciences, Sivas Cumhuriyet University, Sivas, Türkiye.
Department of Biostatistics, Faculty of Medicine, Malatya Turgut Ozal University, Malatya, Türkiye.
Front Physiol. 2025 Jul 16;16:1549306. doi: 10.3389/fphys.2025.1549306. eCollection 2025.
This study aims to build a machine learning (ML) prediction model integrated with explainable artificial intelligence (XAI) to categorize obesity levels from physical activity and dietary patterns. The inclusion of XAI methodologies facilitates a comprehensive understanding of the risk factors influencing the model predictions and thus increases transparency in the identification of obesity risk factors.
Six ML models were used: Bernoulli Naive Bayes, CatBoost, Decision Tree, Extra Trees Classifier, Histogram-based Gradient Boosting and Support Vector Machine. For each model, hyperparameters were tuned by random search methodology and model effectiveness was evaluated by repeated holdout testing. SHAP (SHapley Additive Annotations) and LIME (Local Interpretable Model Independent Annotations) interpretability methods were used to generate local and global feature importance measures.
The CatBoost model exhibited the highest overall performance and achieved superior results in accuracy, precision, F1 score and AUC metrics. Nonetheless, other models such as Decision Tree and Histogram-based Gradient Boosting also yielded strong and competitive results. The results also highlighted age, weight, height and specific food patterns as key predictors of obesity. In terms of interpretability, LIME showed superior in fidelity, whereas SHAP showed improved sparsity and consistency across models, facilitating a comprehensive understanding of trait importance.
This research demonstrates that ML algorithms, when integrated with XAI technologies, can accurately predict obesity levels and explain important contributing risk factors. The use of SHAP and LIME increases model transparency, facilitating the identification of specific lifestyle patterns linked to obesity risk. These findings help to formulate more precise intervention techniques guided by a reliable and understandable predictive framework.
本研究旨在构建一个集成可解释人工智能（XAI）的机器学习（ML）预测模型，以根据身体活动和饮食模式对肥胖水平进行分类。纳入XAI方法有助于全面理解影响模型预测的风险因素，从而提高肥胖风险因素识别的透明度。
使用了六种ML模型：伯努利朴素贝叶斯、CatBoost、决策树、极端随机树分类器、基于直方图的梯度提升和支持向量机。对于每个模型，通过随机搜索方法调整超参数，并通过重复留出测试评估模型有效性。使用SHAP（SHapley加法解释）和LIME（局部可解释模型无关解释）可解释性方法来生成局部和全局特征重要性度量。
CatBoost模型表现出最高的整体性能，在准确性、精确性、F1分数和AUC指标方面取得了优异的结果。尽管如此，其他模型如决策树和基于直方图的梯度提升也产生了强劲且具有竞争力的结果。结果还突出了年龄、体重、身高和特定食物模式是肥胖的关键预测因素。在可解释性方面，LIME在保真度方面表现出色，而SHAP在模型间的稀疏性和一致性方面有所改进，有助于全面理解特征重要性。
本研究表明，ML算法与XAI技术集成时，可以准确预测肥胖水平并解释重要的风险因素。SHAP和LIME的使用提高了模型透明度，便于识别与肥胖风险相关的特定生活方式模式。这些发现有助于制定更精确的干预技术，以可靠且可理解的预测框架为指导。
