Mental Health Detection from Open-Ended Survey Responses Using NLP: A Study on Private University Students in Bangladesh

Abstract

Mental-health challenges among university students have become a significant global concern, with rising rates of depression, anxiety, and academic burnout reported across higher-education sectors. In Bangladesh, these issues are particularly acute, yet research rarely examines the deeper emotional and linguistic expressions through which students articulate psychological distress. This study investigates mental-health risk detection among private university students using a hybrid Natural Language Processing (NLP) and machine-learning (ML) framework applied to open-ended survey responses. A mixedmethod dataset of 304 students was collected through both online (Google Forms) and printed questionnaires, containing structured psychometric indicators alongside seven narrative questions capturing emotional experiences. The textual data underwent a comprehensive NLP pipeline including tokenization, stopword removal, lemmatization, TF–IDF vectorization, sentiment analysis (VADER), and Latent Dirichlet Allocation (LDA) topic modeling. Four classical ML classifiers Logistic Regression, Support Vector Machine (SVM), Gaussian Naïve Bayes, and Random Forest were developed using a combination of linguistic features, sentiment polarity, academic stress scores, financial stress scores, and PHQ-9–inspired symptom scores. Among these, Gaussian Naïve Bayes achieved the highest accuracy (0.80) and F1-score (0.80), demonstrating strong performance on short, sparse, and mixed-language student text. Topic modeling revealed three dominant psychological themes: exam-related anxiety, emotional exhaustion, and coping strategies. Sentiment polarity showed clear alignment with mental-health risk categories, with high-risk students expressing significantly more negative emotional tone. The findings highlight the potential of NLP-based approaches for early mental-health screening in Bangladeshi universities and provide the first openended mental-health dataset curated specifically for private university students. This study establishes a foundation for AI-supported counseling interventions and contributes to the emerging field of computational mental health in low-resource academic environments.