Scholarlink

<p>One in 25 to 35% of students is affected by mental disorders in universities around the world, but 60 to 80% will not seek help from a mental health professional. This crisis has been made worse by the COVID-19 pandemic, and post-pandemic studies have found that there is a 40–60% rise in the prevalence of depression and anxiety among university students worldwide, which is why scalable, proactive screening methods are urgently needed. The paper developed and experimented on BD-MHAM (Big Data-Mental Health Assessment Model), a multi-modal machine learning framework that incorporates six complementary sources of data, including psychological assessments, academic records, digital behavioral indicators, physiological measurements, social-environmental factors, and ecological momentary assessments. Three new elements were created: (1) Multi-Modal Temporal Feature Fusion (MMTFF) to learn dynamic trends across data types; (2) Stability-Weighted Ensemble Feature Selection (SWEFS) to identify predictors that are stable; and (3) Fairness-Constrained Stacking Ensemble (FCSE) to make fair predictions between demographics. Large-scale studies on 14,604 participants in 24 institutions showed that BD-MHAM was significantly better than seven baseline methods with an AUC of 0.972, an accuracy of 95.6%, and an F1-score of 0.928 (p &lt; 0.001). High transportability (AUC = 0.947, degradation of only 2.57%) was also validated on a separate cohort (n = 2,156). The multi-institutional dataset was collected across 24 geographically distributed universities spanning five provinces in China, encompassing urban and rural institutions, research-intensive and teaching-focused universities, with standardized data collection protocols coordinated through a centralized center. SHAP-based interpretability analysis identified sleep quality, academic trajectory, and social interaction frequency as the most influential predictors. Algorithmic fairness evaluation demonstrated equitable performance across gender, age, and academic-level subgroups (demographic parity difference = 0.032, equal opportunity difference = 0.028). These findings carry implications not only for institutions in high-income contexts but also for adaptation to diverse educational systems in low- and middle-income countries, where mental health resources are even more limited.</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

A multimodal fairness aware machine learning framework for mental health risk prediction in university students

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月3日

查看全文
  • Zhu Tian,
  • Jian Zhang,
  • Xia Wang

摘要

One in 25 to 35% of students is affected by mental disorders in universities around the world, but 60 to 80% will not seek help from a mental health professional. This crisis has been made worse by the COVID-19 pandemic, and post-pandemic studies have found that there is a 40–60% rise in the prevalence of depression and anxiety among university students worldwide, which is why scalable, proactive screening methods are urgently needed. The paper developed and experimented on BD-MHAM (Big Data-Mental Health Assessment Model), a multi-modal machine learning framework that incorporates six complementary sources of data, including psychological assessments, academic records, digital behavioral indicators, physiological measurements, social-environmental factors, and ecological momentary assessments. Three new elements were created: (1) Multi-Modal Temporal Feature Fusion (MMTFF) to learn dynamic trends across data types; (2) Stability-Weighted Ensemble Feature Selection (SWEFS) to identify predictors that are stable; and (3) Fairness-Constrained Stacking Ensemble (FCSE) to make fair predictions between demographics. Large-scale studies on 14,604 participants in 24 institutions showed that BD-MHAM was significantly better than seven baseline methods with an AUC of 0.972, an accuracy of 95.6%, and an F1-score of 0.928 (p < 0.001). High transportability (AUC = 0.947, degradation of only 2.57%) was also validated on a separate cohort (n = 2,156). The multi-institutional dataset was collected across 24 geographically distributed universities spanning five provinces in China, encompassing urban and rural institutions, research-intensive and teaching-focused universities, with standardized data collection protocols coordinated through a centralized center. SHAP-based interpretability analysis identified sleep quality, academic trajectory, and social interaction frequency as the most influential predictors. Algorithmic fairness evaluation demonstrated equitable performance across gender, age, and academic-level subgroups (demographic parity difference = 0.032, equal opportunity difference = 0.028). These findings carry implications not only for institutions in high-income contexts but also for adaptation to diverse educational systems in low- and middle-income countries, where mental health resources are even more limited.