Review Finds Artificial Intelligence Depression Detection Performance Drops In Large Studies

Artificial intelligence models designed to detect depression demonstrate substantially lower accuracy in large populations compared to small benchmark tests, according to a narrative review published in Discover Mental Health. While smaller case-control investigations involving 30 to 200 participants frequently report diagnostic accuracies between 93 percent and 100 percent under internal validation, performance drops markedly in studies evaluating several hundred to nearly 15,000 individuals. In these larger cohorts, models analyzing language, speech, and clinical records achieved areas under the curve or balanced accuracies of approximately 0.69 to 0.78, while neuroimaging models recorded accuracies of only 54 to 61 percent.

The review investigated artificial intelligence approaches across multiple modalities, including text, clinical records, large language models, speech and acoustic signals, facial behavior, electroencephalography, and multimodal combinations. Researchers conducted comprehensive searches across PubMed/MEDLINE, Embase, PsycINFO, Scopus, IEEE Xplore, and the ACM Digital Library, alongside the ACL Anthology, arXiv, and citation searching. The synthesis evaluated papers published from January 2015 through August 2026, appraising each study for sample size, data modality, reference standard, and validation strategy.

Although reported figures across modalities frequently exceeded 80 percent accuracy, with multimodal fusion often outperforming single modalities, the authors identified pervasive methodological limitations across the literature. Most models are trained against self-report scales such as the PHQ-8 rather than diagnostic interviews. Furthermore, external validation across independent datasets remains rare, widely used public benchmarks exhibit demographic and interviewer-prompt biases, and no study to date has demonstrated that implementing artificial intelligence detection improves patient outcomes.

Individual technologies presented distinct technical and clinical constraints. Large language models broaden text-based evaluation capabilities but display variable, prompt-dependent performance, especially when restricted to locally deployable systems. Similarly, electroencephalography provides objective neurophysiological biomarkers, but the near-perfect detection figures reported in small sample sizes have failed to replicate at scale. Diagnosis in traditional settings continues to rely on clinical interviews and self-reports that remain subjective, stigma-sensitive, and resource-intensive, leaving depression under-recognised, particularly in low-resource environments where specialist capacity is constrained.

Given these constraints, the authors conclude that artificial intelligence currently offers plausible utility for scalable pre-screening, continuous monitoring, and triage rather than serving as an autonomous diagnostic replacement. Achieving genuine clinical readiness will require establishing larger and demographically diverse datasets, utilizing interview-based reference standards, completing rigorous external and prospective validation, and adopting transparent reporting, fairness auditing, and privacy-preserving system designs under clinician supervision.

Subject of Research: Clinical readiness and accuracy of artificial intelligence methods for depression detection

Article Title: A narrative review of the accuracy and clinical readiness of artificial intelligence methods for detecting depression

Article References: Younas, S., Paripati, L. M. R., Niharika, A., Panda, A., Mwita, R. P., Alghzawi, H. M., Sahu, S., Modi, N. M., & Aminpoor, H. (2026). A narrative review of the accuracy and clinical readiness of artificial intelligence methods for detecting depression. Discover Mental Health. https://doi.org/10.1007/s44192-026-00653-8

Image Credits: AI Generated

DOI: 10.1007/s44192-026-00653-8

Keywords: artificial intelligence, depression detection, narrative review, machine learning, clinical readiness, diagnostic accuracy, large language models, electroencephalography, narrative, review, accuracy, clinical