Gender differences in the prevalence of anxiety and post-traumatic stress disorder and associated factors among adults in sub-Saharan Africa: a systematic review and meta-analysis.

The integration of artificial intelligence into medicine has accelerated markedly over the past decade. Psychiatry, unsurprisingly, is no exception. From automated diagnostic classification to AI-assisted clinical interviews and smartphone-based passive monitoring, the scope of potential application is striking. A systematic review encompassing studies on diagnosis, monitoring, and intervention identified support vector machines, random forests, and AI-based conversational agents as the most frequently deployed approaches (). Machine learning models have achieved classification accuracies exceeding 0.90 for anorexia nervosa, bulimia nervosa, and major depressive disorder (MDD). Furthermore, deep learning applied to nationwide register and genetic data has enabled cross-diagnostic predictions for schizophrenia, bipolar disorder, attention-deficit/hyperactivity disorder (ADHD), autism spectrum disorder (ASD), and MDD with AUC-ROC values between 0.71 and 0.82 (; ; ). These figures are technically noteworthy but must be interpreted within their methodological context. These metrics reflect performance in retrospective or internally validated research settings, and should not be interpreted as evidence of clinical readiness, external generalizability, or diagnostic validity in real-world psychiatric practice. At the same time, a growing body of evidence points to important methodological weaknesses, an absence of external validation that is near-universal in certain domains [particularly passive sensing, where it is present in only 2% of studies ()], and a fragmented and inconsistently enforced governance landscape that may expose patients - among the most vulnerable populations - to risks they cannot meaningfully consent to. A further, underappreciated risk is conceptual: AI diagnostic systems are built upon a notion of normality that has never withstood rigorous epistemological scrutiny (; ). Importantly, when statistical averages derived from non-representative populations become the implicit reference for machine learning classifiers, pathologization of diversity becomes structurally likely - not merely possible - particularly when models employ diagnostic labels that conflate difference with disorder, or are deployed without cultural and functional contextualization. When a machine-learning model predicts a suicide risk on the basis of a pattern it cannot explain, the traditional dyad of care is fundamentally challenged. While AI offers growing technical capability, it does so within a governance landscape that has yet to establish adequate and universally enforceable patient protections - and, for some commercial actors, within incentive structures that treat sensitive mental health data as a monetizable asset.