Mental Health & AI

Early AI Applications & Foundations

Before the pandemic, AI models primarily utilized textual data and Natural Language Processing (NLP) to identify linguistic markers of depression on platforms like Reddit and Facebook. Studies such as Islam et al. [41] and Tadesse et al. [38] established foundational methods with reasonable accuracy.

Islam et al. achieved an F-measure of 0.73, identifying psycholinguistic indicators from Facebook comments. Tadesse et al. demonstrated 91% accuracy using MLP classifiers on Reddit posts. Ricard et al. [40] explored user-generated and community-generated Instagram data, yielding an AUC of 0.72, showcasing early predictive capabilities.

These initial efforts laid the groundwork, emphasizing text-based analysis and basic machine learning, but faced limitations in data diversity and real-time applicability.

AI During COVID-19: Adaptations & Accelerations

The COVID-19 pandemic significantly increased social media usage and the prevalence of mental health issues, creating a surge of emotionally charged data. This period saw a shift towards more advanced AI, particularly transformer-based Deep Learning (DL) models, to monitor depression trends.

Zhang et al. [36] used BERT, RoBERTa, and XLNet on 5150 Twitter users, achieving 78.9% accuracy and highlighting improvements with larger datasets. Zogan et al. [33] introduced the Hierarchical Attenuation Convolutional Neural Network (HCN) to detect increased depression rates during the pandemic, reporting an impressive 93.4% accuracy and 89.7% F1-score. This era marked a critical acceleration in AI model sophistication to handle the complex, evolving nature of online communication.

Post-Pandemic Advances: Multimodal & Real-time AI

The post-pandemic period has seen a continued evolution in AI for depression detection, with a growing emphasis on multimodal approaches—integrating visual, textual, and interaction data for more nuanced analysis. Real-time detection capabilities have also become a key focus.

Anshul et al. [35] developed a novel DL framework integrating Visual Neural Networks (VNN) with textual analysis, achieving high accuracies of 93.1% on the Tsinghua dataset and 91.7% on a COVID-19 dataset. Chatterjee et al. [34] created a real-time depression detection model using advanced ML (SVM classifier) combined with sentiment analysis, achieving 89% accuracy. These advancements underscore the potential for comprehensive, dynamic AI tools in mental health monitoring.

Addressing Technical & Ethical Challenges

Across all temporal phases, AI implementation for mental health detection has faced consistent challenges. Key technical hurdles include data selection bias, limited generalization, and issues with self-reporting reliability. Demographic, linguistic, and cultural biases remain significant concerns, potentially affecting model fairness and applicability across diverse user groups [48].

Ethical considerations, particularly regarding user privacy, data security, and responsible AI practices, are paramount [47]. While acknowledged in many studies, concrete strategies for enhancing transparency, accountability, and user-centered approaches are still evolving [49]. Future research must prioritize interdisciplinary collaboration and robust ethical frameworks to ensure AI solutions are both effective and responsible.