Tailoring care to the mind: The emerging role of artificial intelligence in precision psychiatry. Part II: Validating, regulating and integrating artificial intelligence into care

Encephale. 2026 Sep 19:S0013-7006(26)00182-X. doi: 10.1016/j.encep.2026.08.007. Online ahead of print.

ABSTRACT

OBJECTIVES: Precision psychiatry seeks to tailor care to the clinical, biological, cognitive, behavioral, and contextual characteristics of each patient. Artificial intelligence may help achieve this goal by combining multimodal data, identifying complex patterns, following individual trajectories, and refining prognostic or therapeutic stratification. Yet strong statistical performance does not automatically translate into meaningful clinical benefit. This second part of Tailoring Care to the Mind: The Emerging Role of Artificial Intelligence in Precision Psychiatry explores the conditions required for an algorithm to become a reliable, acceptable, equitable, and responsible clinical innovation.

MATERIALS AND METHODS: We conducted a narrative and conceptual review organized around three main questions. First, how should predictive models be evaluated from a clinical perspective? Second, under what conditions can they be integrated into routine psychiatric care? Third, what ethical, regulatory, and relational issues arise when these tools are used in practice? Particular attention was given to discrimination, calibration, decision usefulness, external validation, bias, acceptability, interpretability, responsibility, and human oversight.

RESULTS: Statistical performance alone is not enough to establish the clinical value of a model. A useful system must not only distinguish between patients with different outcomes, but also provide risk estimates that match what is actually observed in practice. It should demonstrate added value compared with clinical judgment, existing scores, or standard care, while taking into account the consequences of false-positive and false-negative predictions. External validation is essential because a model that performs well in one population or center may be less reliable elsewhere. Performance must also be examined across clinically relevant subgroups to identify unequal or potentially discriminatory effects. Once deployed, models require ongoing monitoring, as changes in populations, data collection, or clinical practice may lead to model drift and loss of calibration. Their integration into care also depends on practical factors: the information provided must be understandable, timely, useful, and compatible with clinical workflows. Acceptability among clinicians and patients is shaped by perceived usefulness, reliability, transparency, preservation of professional roles, maintenance of human contact, and confidence in the protection of sensitive psychiatric data. Ethical legitimacy further requires clear rules for data access, secondary use, accountability, and the possibility of questioning or overriding an algorithmic recommendation.

DISCUSSION: Artificial intelligence should be seen neither as a replacement for the psychiatrist nor as a neutral technical device. It becomes a third component within the clinical relationship, capable of influencing how problems are framed, risks are estimated, and treatment options are discussed. Its outputs must therefore be interpreted in light of the patient’s history, context, preferences, values, and lived experience. A hybrid model, combining algorithmic information, clinical expertise, and patient participation, appears most consistent with the nature of psychiatric care.

CONCLUSION: Moving from algorithmic performance to clinical innovation requires rigorous validation, continuous monitoring, robust governance, practical and social acceptability, and preservation of the therapeutic alliance. The future of precision psychiatry will depend less on autonomous systems than on the careful integration of artificial intelligence into shared, contextualized, and accountable clinical decision-making.

PMID:42763258 | DOI:10.1016/j.encep.2026.08.007