Evaluating current trends in stress and depression detection using artificial intelligence and machine learning

Stress and depression are two major mental health-related concerns that affect people of varying ages, ranging from teenagers to working class people to senior citizens. The detection of these conditions helps the patients recover effectively which improves their quality of life. As behavioral, physiological, and digital trace data continues to be more accessible, computational models (especially AI-based ones) have become considered a promising solution to stress and mental health assessment. The current paper is a review of the recent research on AI-based stress and depression detection and tries to describe the way the field is developing. The diversity of the works is categorized broadly based on model type, input modality, data source, and evaluation strategy. Classical applications for machine learning remain quite prevalent, particularly for structured data such as EEG signals and questionnaire measurements. DL models and transformer architectures use of transformer-based architectures seems to be effective at physiological signal analysis and text inputs, but reported performance fluctuates. Vision-based models, particularly those using Convolutional Neural Networks (CNNs) on static image data such as facial scans or psychological drawings, are widely used. The use of Large Language Models (LLMs) and multimodal systems remains limited. Performance across reviewed studies was primarily evaluated using Accuracy, F1-score, and Root Mean Square Error (RMSE). Reported accuracies generally ranged from 74% to 98%, though results vary significantly depending on whether subject-independent validation was used. Some of the limitations in the literature we synthesized are - limited interpretability, inconsistent data standards, virtually no longitudinal modeling, and very little manifestation or clinical validation. This article unites these findings and identifies common gaps and recommends where the field might want to go next - including more standardized datasets, increased interpretability, and practical validation.