TY  - JOUR
AU  - Bhukya, Ramesh K.
AU  - Singh, Nirbhay
AU  - Tiwari, Aniket
PY  - 2026
DA  - 2026/10/01
TI  - A Multimodal Ensemble Learning Framework for Early Detection of Adolescent Depression Using Multi-Source Real-World Data
JO  - ICCK Transactions on Educational Data Mining
T2  - ICCK Transactions on Educational Data Mining
JF  - ICCK Transactions on Educational Data Mining
VL  - 2
IS  - 2
SP  - 63
EP  - 84
DO  - 10.62762/TEDM.2026.244050
UR  - https://www.icck.org/article/abs/TEDM.2026.244050
KW  - educational data mining
KW  - student depression
KW  - academic stress
KW  - student well-being
KW  - learning analytics
KW  - early warning systems
AB  - Depression and academic stress undermine students' attendance, performance, and retention, yet educational institutions lack reliable early-alert tools. Because about three-quarters of lifetime psychiatric conditions emerge before age 25, timely automated screening is critical for early intervention. This paper proposes a multi-source ensemble learning framework for distinguishing {\it depression} (DP) from {\it non-depression} (ND) using 1,547 harmonized samples drawn from six real-world sources: occupational surveys, student academic stress records, Twitter sentiment data, Reddit depression posts, fine-grained emotion corpora, and behavioural monitoring signals. Academic stress and behavioural records serve as the core educational sources, while the remaining sources broaden linguistic and affective coverage. Six {\it machine learning} (ML) classifiers were benchmarked under 10-fold cross-validation: {\it decision tree} (DT), {\it support vector machine} (SVM), {\it random forest} (RF), {\it adaptive boosting} (AdaBoost), {\it extreme gradient boosting} (XGBoost), and {\it light gradient boosting machine} (LightGBM). RF performed best, achieving 97.996% accuracy, the highest sensitivity (96.741%), only 6 false positives, and 25 missed DP cases. LightGBM recorded the highest specificity (99.615%), making it suitable for precision-oriented screening, whereas the boosting ensembles performed comparably to the linear SVM and all ensembles clearly outperformed the DT baseline. These findings support ensemble learning, particularly bagging, for heterogeneous, multi-source student mental health data and suggest practical value for counsellors and student support services in identifying at-risk students early.
SN  - 3070-5843
PB  - Institute of Central Computation and Knowledge
LA  - English
ER  - 

@article{Bhukya2026A,
  author = {Ramesh K. Bhukya and Nirbhay Singh and Aniket Tiwari},
  title = {A Multimodal Ensemble Learning Framework for Early Detection of Adolescent Depression Using Multi-Source Real-World Data},
  journal = {ICCK Transactions on Educational Data Mining},
  year = {2026},
  volume = {2},
  number = {2},
  pages = {63-84},
  doi = {10.62762/TEDM.2026.244050},
  url = {https://www.icck.org/article/abs/TEDM.2026.244050},
  abstract = {Depression and academic stress undermine students' attendance, performance, and retention, yet educational institutions lack reliable early-alert tools. Because about three-quarters of lifetime psychiatric conditions emerge before age 25, timely automated screening is critical for early intervention. This paper proposes a multi-source ensemble learning framework for distinguishing {\it depression} (DP) from {\it non-depression} (ND) using 1,547 harmonized samples drawn from six real-world sources: occupational surveys, student academic stress records, Twitter sentiment data, Reddit depression posts, fine-grained emotion corpora, and behavioural monitoring signals. Academic stress and behavioural records serve as the core educational sources, while the remaining sources broaden linguistic and affective coverage. Six {\it machine learning} (ML) classifiers were benchmarked under 10-fold cross-validation: {\it decision tree} (DT), {\it support vector machine} (SVM), {\it random forest} (RF), {\it adaptive boosting} (AdaBoost), {\it extreme gradient boosting} (XGBoost), and {\it light gradient boosting machine} (LightGBM). RF performed best, achieving 97.996\% accuracy, the highest sensitivity (96.741\%), only 6 false positives, and 25 missed DP cases. LightGBM recorded the highest specificity (99.615\%), making it suitable for precision-oriented screening, whereas the boosting ensembles performed comparably to the linear SVM and all ensembles clearly outperformed the DT baseline. These findings support ensemble learning, particularly bagging, for heterogeneous, multi-source student mental health data and suggest practical value for counsellors and student support services in identifying at-risk students early.},
  keywords = {educational data mining, student depression, academic stress, student well-being, learning analytics, early warning systems},
  issn = {3070-5843},
  publisher = {Institute of Central Computation and Knowledge}
}