Depression Detection Using a Hybrid Convolutional Neural Network and Recurrent Neural Network Approach

Abstract

Depression is a widespread psychological condition that disrupts how people feel, think, and function, and recognising it early allows support to begin before symptoms worsen. Most automated approaches rely either on conventional machine learning, which struggles with interacting risk factors, or on single deep learning models that capture only one type of pattern. Here we propose a hybrid network that joins one-dimensional convolutional layers (CNN) to a bidirectional long short-term memory layer (BiLSTM) in order to separate depressed from non-depressed students. The novelty lies in treating structured demographic, academic, and lifestyle attributes as an ordered sequence so that convolutional and bidirectional recurrent layers can learn their interactions jointly. An open Kaggle dataset of 27,901 student records with 16 attributes was used. Missing values were imputed, categorical attributes were encoded, numerical features were standardised, and class weighting was applied; 80% of the records were used for training, while a stratified 20% was kept aside for testing. Evaluation relied on six measures: accuracy, precision, recall, F1-score, area under the ROC curve, and Matthews correlation. On the unseen test records the network reached 84% accuracy, with F1 = 0.84, ROC-AUC = 0.91, and MCC = 0.67, outperforming the classical models reported in comparable studies. The framework offers a reliable and scalable basis for automated mental-health screening in educational settings.