Thity Nishat Tasnim, Rahman Atikur, Dulmini Adisha, Yasmin Mst Nilufar, Rois Rumana
Department of Statistics and Data Science, Jahangirnagar University, Dhaka, Bangladesh.
Department of Business and Law, University of Wollongong, New South Wales, Australia.
PLoS One. 2025 Jul 21;20(7):e0325855. doi: 10.1371/journal.pone.0325855. eCollection 2025.
The internet is one of the essential tools today, and its impact is particularly felt among university students. Internet addiction (IA) has become a serious public health issue worldwide. This multi-class classification study aimed to identify the potential predictors of IA by four severity levels among university students in Bangladesh. We used cross-sectional survey data from 424 university students from different universities in Bangladesh. Data was collected using a self-reported questionnaire, along with an IA test to assess addiction levels. We identified the important features related to IA using the Boruta algorithm. Predictions were made using different machine learning (ML) (decision tree (DT), random forest (RF), support vector machines (SVMs), and logistic regression (LR)) models. Their performance was assessed using confusion matrix parameters, receiver operating characteristics (ROC) curves, and k-fold cross-validation techniques for multi-class classification problems. The prevalence of severe IA was 3.77% among the participating university students in Bangladesh from July 15 to July 22, 2024. University students' backgrounds, depression, anxiety, stress, participation in physical activity, misbehaving with family members, memory loss symptoms, and being COVID-19-positive were selected as significant features for predicting IA. Overall, the RF (accuracy = 0.531, sensitivity = 0.200, specificity = 0.986, precision = 1.00, k-fold accuracy = 0.4858, micro-average area under curve (AUC) = 0.7798) more accurately predicted IA compared to other ML techniques. The ML framework for multi-class classification study can reveal significant risk factors and predict this behavioral addiction more precisely. It can help policymakers, stakeholders, and families better understand the situation and prevent this severe crisis by improving policy-making strategies, promoting mental health, and establishing effective university counseling services. Therefore, raising awareness among the younger generation and their parents about the predictors of IA is important.
互联网是当今重要的工具之一，大学生对其影响感受尤深。网络成瘾（IA）已成为全球范围内一个严重的公共卫生问题。这项多类别分类研究旨在确定孟加拉国大学生中按四个严重程度级别划分的网络成瘾潜在预测因素。我们使用了来自孟加拉国不同大学的424名大学生的横断面调查数据。数据通过自我报告问卷收集，并进行了一项网络成瘾测试以评估成瘾程度。我们使用博鲁塔算法确定了与网络成瘾相关的重要特征。使用不同的机器学习（ML）（决策树（DT）、随机森林（RF）、支持向量机（SVM）和逻辑回归（LR））模型进行预测。使用混淆矩阵参数、受试者工作特征（ROC）曲线以及针对多类别分类问题的k折交叉验证技术评估了它们的性能。在2024年7月15日至7月22日参与调查的孟加拉国大学生中，严重网络成瘾的患病率为3.77%。大学生背景、抑郁、焦虑、压力、参与体育活动、与家庭成员行为不端、记忆丧失症状以及新冠病毒检测呈阳性被选为预测网络成瘾的显著特征。总体而言，与其他机器学习技术相比，随机森林（准确率 = 0.531，灵敏度 = 0.200，特异性 = 0.986，精确率 = 1.00，k折准确率 = 0.4858，微平均曲线下面积（AUC） = 0.7798）能更准确地预测网络成瘾。多类别分类研究的机器学习框架可以揭示重大风险因素，并更精确地预测这种行为成瘾。它可以帮助政策制定者、利益相关者和家庭更好地了解情况，并通过改进决策策略、促进心理健康和建立有效的大学咨询服务来预防这场严重危机。因此，提高年轻一代及其父母对网络成瘾预测因素的认识很重要。
