<![CDATA[Machine learning is now woven into mental health assessment, and the predictive models it produces are impressively accurate. Yet the black-box nature of those models keeps raising the same worries: transparency, clinician trust, and whether they can be adopted safely in practice. Explainable AI, or XAI, is often proposed as the answer, though the evidence on where it helps and how much is still scattered. We ran a systematic review following PRISMA 2020 to pull that evidence together. A structured search of Scopus turned up 403 records. After 18 duplicates were removed, 385 went into screening; 75 full texts were assessed, and 15 studies, all published between 2021 and 2025, made it into the final pool. Study design, mental health condition, modality, XAI technique, interpretability, trust, and clinical utility were all captured and synthesised thematically. Three things stood out. First, SHAP and other post-hoc feature-attribution techniques dominated the field, appearing across work on depression, suicide risk, stress, and anxiety alike. Second, explanation quality did shape trust and acceptance, but remarkably few studies actually involved human participants. Third, heterogeneity, small samples, geographic bias, and a lack of standardised metrics all limited how far the results could be compared or translated into practice. Where should the field go next? Priorities include clinician-in-the-loop evaluation, standardised reporting of explanation quality, and equitable external validation.]]>
Copyrights © 2025
