Machine Learning Approaches to Anxiety Detection: Trends, Model Evaluation, and Future Directions

Meruyert TaskynbayevaAlina Gutoreva^*

• Kazakh-British Technical University, Almaty, Kazakhstan

Background: Anxiety is a widespread mental health disorder with profound implications for individual well-being and societal functioning. Its increasing prevalence, particularly among young people in digitally saturated environments, underscores the urgent need for predictive tools that can enable early intervention. Machine learning (ML) has emerged as a promising approach in this domain, yet the literature lacks a consolidated synthesis that critically evaluates current applications, methodological trends, and translational challenges. Objectives: This systematic review synthesises studies applying ML to anxiety prediction, with the aims of (1) identifying the most frequently used algorithms and diagnostic instruments, (2) evaluating model performance across diverse datasets, and (3) highlighting methodological limitations, gaps, and future directions. Methods: We systematically searched Scopus and Google Scholar for studies published between 2018 and July 2025, using combinations of the terms “anxiety prediction,” “machine learning,” and “mental health.” Eligible studies employed ML to predict anxiety levels or severity based on clinical or behavioural datasets. Exclusion criteria encompassed non-English, review articles, older publications, and studies not targeting anxiety prediction. Screening occurred in two phases, with potential bias noted in studies relying solely on self-reported metrics or with unclear algorithmic transparency. Results: Nineteen studies, including 44,608 participants, met inclusion criteria. Commonly used diagnostic instruments included GAD-7 and DASS-21. Random Forest and Gradient Boosting algorithms consistently demonstrated high predictive performance, with some studies reporting up to 98% accuracy. Performance metrics most frequently reported were F1-score, AUC, and specificity. However, methodological weaknesses were evident, including reliance on small or homogeneous samples, inconsistent evaluation metrics, lack of external validation, and the predominance of “black-box” models that limit interpretability and clinical trust. Conclusions: The review confirms the promise of ML in predicting anxiety, particularly in enabling early intervention. Yet the scarcity of studies addressing anxiety onset before symptom manifestation highlights a critical gap. Future research should prioritise explainable AI approaches, integration of multimodal and clinically validated datasets, and rigorous external validation to improve generalizability and foster clinical adoption. Together, these efforts can advance the transition from reactive classification toward proactive risk forecasting in mental health.