Predicting Mental Health with Neural Networks: An Open-Source Project for Questionnaire-Based Depression, Anxiety, and Stress Assessment

This article introduces the open-source project named mental-health-project on GitHub, released by developer tasneem33355. It uses neural network models to predict users' depression, anxiety, and stress levels based on responses to the standardized DASS-21 questionnaire. The project provides a complete pipeline from data processing to model training and an interactive prediction interface, aiming to address issues such as traditional mental health assessments relying on professionals, long cycles, and poor accessibility, thus offering practical references for the intelligentization of mental health screening.

Over 280 million people worldwide are affected by depression, and the prevalence of anxiety disorders is equally severe. Traditional assessments rely on face-to-face interviews, facing bottlenecks such as a shortage of professionals. The project builds a model based on the DASS-21 scale (21 questions, divided into three dimensions: depression, anxiety, and stress, with a four-level rating). The core hypothesis is that the pattern of questionnaire responses contains regularities that can be captured by the model, and prediction is achieved by learning the mapping relationship of labeled samples.

Data Processing: Clean missing/abnormal values, feature encoding (retain rank order or one-hot encoding), split into 80/20 training and test sets, and cross-validation can be introduced. Model Design: Use a Multi-Layer Perceptron (MLP), where the input layer receives preprocessed features, the hidden layer uses ReLU activation, and the output layer performs multi-task learning to predict three dimensions simultaneously; Adam optimizer is used for training, and Dropout and early stopping are applied to prevent overfitting. Interactive Interface: Provide a user-friendly interface, allowing non-technical users to answer the questionnaire and get intuitive prediction results, lowering the threshold for use.

Evaluation metrics include accuracy, precision, recall, and F1 score, among which recall is crucial for screening scenarios (the consequences of missed detection are more serious). Since mental health levels are continuously distributed, it is necessary to examine the performance at different levels; the confusion matrix can show the classification performance at each level, helping to optimize the weak links of the model.

This tool is only for preliminary screening and cannot replace professional diagnosis; it is necessary to strictly comply with data privacy regulations to ensure that user data collection and storage comply with the principle of informed consent; attention should be paid to fairness issues caused by training data bias to avoid low prediction accuracy for some groups.

Application Scenarios: Can be deployed on corporate health platforms, school psychological systems, and community health institutions as a preliminary screening tool. Improvement Directions: Introduce Transformer to capture the dependency between questions; combine NLP to analyze free text; integrate multi-modal data such as voice/facial expressions/physiological signals; develop longitudinal tracking functions to monitor state changes.

The project demonstrates the complete process of building a mental health prediction tool using neural networks and provides clear and reproducible practical references. Although it faces challenges such as accuracy verification, ethical compliance, and fairness assurance, it opens up an exploration path for the intelligentization and universalization of mental health services, and is a valuable starting point for AI applications in the public health field.