JMIR Publications

ABSTRACT

Background:

Depression is a major contributor to the global burden of disease and is particularly prevalent among young people, including university students. In Bangladesh, limited mental health resources, stigma, and gaps in early identification can delay timely support for students who are at risk. At the same time, digital health and machine learning approaches create an opportunity to scale screening and improve triage in resource-constrained settings, but these tools must be clinically aligned, interpretable, and fair to be trusted in real services. Against this backdrop, this study develops a human-in-the-loop, explainable, and calibrated machine-learning decision-support framework that translates predicted depression risk into clear tiers (Normal, Mild, Moderate, Severe) to assist counselors in prioritizing care while also providing understandable feedback for students

Objective:

**Objective of the study** The overarching objective of this study is to develop and evaluate a human-in-the-loop, explainable, and clinically usable machine-learning decision-support framework for depression risk screening among Bangladeshi university students, with emphasis on reliable probability calibration and actionable risk communication. **Specific objectives** 1. To adapt and assess a cross-population depression screening approach by leveraging relevant external survey evidence alongside local Bangladeshi student data where appropriate. 2. To implement and compare probability calibration strategies and identify clinically meaningful thresholds to enhance trust and decision alignment. 3. To convert model probabilities into four intuitive risk tiers (Normal, Mild, Moderate, Severe) to support counselor triage and student-facing understanding. 4. To integrate explainability and subgroup fairness checks to ensure transparent and equitable screening outcomes across demographic groups.

Methods:

This study used a quantitative, cross-population transfer-learning design for depression risk screening among Bangladeshi university students. The pipeline leveraged NHANES 2017–2018 as a source dataset and a Bangladeshi student dataset as the target, retaining only shared variables (e.g., age, gender, marital status, smoking, alcohol use, weekly physical activity). Preprocessing included missing-value imputation, feature standardization, one-hot encoding, and SMOTE to address class imbalance. A model zoo (logistic regression, random forest, and gradient boosting variants) was evaluated, with the best model selected using ROC-AUC and PR-AUC criteria, then calibrated using Platt scaling and isotonic regression. Clinically meaningful thresholds were tuned (F1/Youden), and calibrated probabilities were mapped into four risk tiers (Normal, Mild, Moderate, Severe). Explainability (SHAP), subgroup fairness checks, decision-curve analysis, and bootstrap-based validation were applied to ensure transparent, reliable, and counselor-usable outputs.

Results:

The calibrated Random Forest model showed excellent performance on the Bangladeshi holdout set, achieving **ROC–AUC = 0.993**, **PR–AUC = 0.997**, **Accuracy = 0.977**, **F1 = 0.981**, and strong calibration with a **Brier score = 0.021** (with supportive bootstrap CIs reported). The evaluation was conducted on a clearly defined split of the Bangladeshi student cohort (**1,400** in the calibration set and **600** in the holdout set), where severe depressive symptoms were notably prevalent at about **61.3%** in both subsets. These results suggest the framework can provide reliable, tiered risk outputs suitable for counselor triage and student-facing screening support.

Conclusions:

This study concludes that a calibrated transfer-learning approach can feasibly adapt NHANES-based depression screening signals for use among Bangladeshi university students. The findings highlight three key contributions: maintained discrimination and calibration after transfer and probability scaling, clinically meaningful cut-offs through threshold optimization, and conversion of predicted probabilities into four intuitive risk tiers (Normal, Mild, Moderate, Severe) that are understandable for both counselors and students. Practically, the framework can reduce counselor workload via triage while offering a simple, potentially less stigmatizing self-understanding for students; however, it is intended as decision support rather than diagnosis and requires broader external validation and real-world co-design before routine clinical deployment.