Abstract：
Stress is a psychological condition resulting from the body\'s response to challenging situations, which can negatively impact physical and mental health if experienced over prolonged periods. Early detection of stress is crucial to prevent chronic health problems. Wearable sensors offer an effective solution for continuous and real-time stress monitoring due to their non-intrusive nature and ability to monitor vital signs, e.g., heart rate and activity. Typically, most existing research has focused on data collected in controlled environments. Yet, our study aims to propose a machine learning-based approach for detecting stress in a free-living environment using wearable sensors. We utilized the SWEET dataset, which includes data from 240 subjects collected via electrocardiography (ECG), skin temperature (ST), and skin conductance (SC). We assessed four machine learning models, i.e., K-Nearest Neighbors (KNN), Support Vector Classification (SVC), Decision Tree (DT), Random Forest (RF), and XGBoost (XGB) in four different settings. This study evaluates the performance of various machine learning models for stress classification using the SWEET dataset. The analysis included two binary classification scenarios (with and without SMOTE) and two multi-class classification scenarios (with and without SMOTE). The Random Forest model demonstrated superior performance in the binary classification without SMOTE, achieving an accuracy of 98.29 % and an F1-score of 97.89 %. For binary classification with SMOTE, the K-Nearest Neighbors model performed best, with an accuracy of 95.70 % and an F1-score of 95.70 %. In the three-level classification without SMOTE, the Random Forest model again excelled, achieving an accuracy of 97.98 % and an F1-score of 97.22 %. For three-level classification with SMOTE, XGBoost showed the highest performance, with an accuracy and F1-score of 98.98 %. These results highlight the effectiveness of different models under various conditions, emphasizing the importance of model selection and preprocessing techniques in enhancing classification performance.
摘要：
压力是由身体对挑战性情况的反应引起的心理状况，如果长时间经历，会对身心健康产生负面影响。早期发现压力对于预防慢性健康问题至关重要。可穿戴传感器由于其非侵入性和监测生命体征的能力，为连续和实时的压力监测提供了有效的解决方案。例如，心率和活动。通常,大多数现有的研究都集中在受控环境中收集的数据。然而，我们的研究旨在提出一种基于机器学习的方法，用于使用可穿戴传感器检测自由生活环境中的压力。我们利用SWEET数据集，其中包括通过心电图（ECG）收集的240名受试者的数据，皮肤温度(ST),和皮肤电导（SC）。我们评估了四种机器学习模型，即，K-最近邻居(KNN),支持向量分类(SVC)决策树(DT)随机森林(RF),和XGBoost（XGB）在四个不同的设置。本研究使用SWEET数据集评估了各种机器学习模型对压力分类的性能。该分析包括两个二元分类方案(有和没有SMOTE)和两个多分类方案(有和没有SMOTE)。随机森林模型在没有SMOTE的二元分类中表现出优越的性能，准确率为98.29%，F1评分为97.89%。对于使用SMOTE的二元分类，K-最近邻居模型表现最好，准确率为95.70%，F1评分为95.70%。在没有SMOTE的三级分类中，随机森林模型再次出类拔萃，准确率为97.98%，F1评分为97.22%。对于使用SMOTE的三级分类，XGBoost表现出最高的性能，准确率和F1评分为98.98%。这些结果突出了不同模型在各种条件下的有效性，强调模型选择和预处理技术在提高分类性能方面的重要性。
