Application of Artificial Intelligence in psychiatry

AI is a branch of computer science comprising wide-ranging methodologies that have proliferated through our daily lives in recent years. The breadth of this field is exemplified by the European Parliament’s definition of AI systems as specified under Article 3 of Regulation (EU) 2024/1689, also known as the European Union (EU) AI Act:

Figure 1 illustrates the domains that AI systems can impact within current psychiatric care. These systems could enhance the quality of patient care, professional development, and mental healthcare systems. To date, the primary field of AI that has been applied to psychiatry is machine learning, which can be utilised to discover patterns within datasets to inform the development of models that make predictions in unseen data; the majority of these prediction models have been trained using supervised learning. These models learn using labelled data, where the labels represent a pre-specified phenomenon, such as a diagnosis. This paradigm underpins many of the emerging clinical decision support systems that are being applied in psychiatry [Reference Ghelfi, Healy, Piacenza, French, McNamara and Rabbi3], such as PsyMetRiC: a cardiometabolic risk calculator that has been developed to assist care for young people with psychotic disorders [Reference Perry, Osimo, Si, Hitchins, Lewis and Laws4]. Unsupervised learning has also been utilised to uncover patterns in unlabelled datasets, allowing for the identification of latent psychiatric phenotypes across multiple diagnoses [Reference Hofman, Lier, Ikram, van Wingerden and Luik5]. This methodology could allow for symptom-level precision in psychiatric treatment and even the reorganisation of diagnostic constructs.

Figure 1.

We conceptualise the primary domains of AI integration into psychiatric care on a continuum of clinical sensitivity according to their proximity to direct patient care. Leftward domains are primarily system-level or professional-facing, whereas rightward domains are increasingly patient-facing and clinically sensitive.

The recent advent of deep learning – a category of machine learning that leverages multi-layer artificial neural networks – has enabled a qualitative shift in terms of models capable of learning rich, generalisable representations of language at scale. This has ultimately culminated in large language models, which are specifically designed to process, generate, and interpret complex linguistic information. These models underpin many popular conversational agents, such as ChatGPT and Claude. As psychiatry is fundamentally a language-based discipline, in which both the diagnostic process and therapeutic interventions largely constitute structured verbal exchange, large language models may be particularly well suited to assume or augment selected roles, especially tasks that revolve around language-mediated interaction. Indeed, autonomous scribes have already demonstrated potential to reduce administrative burden and clinician burnout [Reference Olson, Meeker, Troup, Barker, Nguyen and Manders6].

Barriers to progress

Despite its transformative potential and increased intellectual interest, the evidence base for AI in psychiatry remains preliminary, with few examples of clinical implementation. Most published studies are early-stage investigations, often characterised by small sample sizes, algorithmic bias, heterogeneous methods, a lack of replication, and utilisation of costly input features or data not routinely collected [Reference Ghelfi, Healy, Piacenza, French, McNamara and Rabbi3, Reference Şahin, Kambeitz-Ilankovic, Wood, Dwyer, Upthegrove and Salokangas7]. These research endeavours are emblematic of what appears to be a fragmented research landscape characterised by siloed projects, which attract scientific curiosity, but seldom progress towards real-world clinical implementation. Indeed, this phenomenon has been observed in other fields of medical AI research [Reference Berkhout, van Wijngaarden, Workum, van de Sande, Hilling and Jung8].

Meaningful adoption into clinical practice will depend not only on model performance, but also on broader contextual factors including digital infrastructure, regulatory preparedness, robust data privacy policies, and cultural readiness to engage with AI tools among both clinicians and service users. Involving key stakeholders, such as experts by experience, clinicians, and policymakers, will be essential to drive real-world impact. Successful implementation will also require attention to core issues of AI such as interpretability, transparency, and responsible handling of sensitive data. Central to this is the challenge of “black box” models, where model accuracy may be high, yet predictive rationale remains cryptic. Addressing this through Explainable AI frameworks, such as SHAP (SHapley Additive exPlanations) values, can improve model transparency. Enabling human oversight through the utilisation of human-in-the-loop approaches is also essential to ensure ethical AI development and implementation.

Forming an international psychiatry and artificial intelligence consortium

The initiation of a large-scale cross-continental international collaborative effort is one route to accelerating the impact of AI in psychiatry. Indeed, consortium approaches have shown benefit in other fields of scientific enquiry that were previously constrained by similar difficulties, such as genomics research [Reference Seminara, Khoury, O’Brien, Manolio, Gwinn and Little9]. Recent activity appears to acknowledge the need for efforts of this nature in psychiatric AI development, exemplified by the inception of the European College of Neuropsychopharmacology’s AI in Mental Health Network (https://www.ecnp.eu/research/ecnp-networks/list-of-ecnp-networks/ai-in-mental-health/). This European-focused consortium has goals of improving data and AI infrastructure, developing and validating individualised prediction models, designing and implementing AI-enabled therapeutics, establishing frameworks for evaluating predictive models, and ensuring the responsible development and deployment of AI systems across the continent. Other large research collaborations are also in operation, such as the TRUSTING consortium (https://trusting-project.eu/), which aims to develop a speech-based AI monitoring system for relapse prediction of psychotic disorders. This research consortium provides a template for successful navigation of key implementation challenges, such as compliance with regulatory and legal frameworks across EU and non-EU countries.

Establishing a robust governance structure, stable funding model, and network culture would be key initial priorities for a broad transcontinental AI and psychiatry consortium, with the long-standing Psychiatric Genomics Consortium (https://pgc.unc.edu/) providing a useful model in this regard. Governance structures could include a central steering committee, data access committee, expert-by-experience advisory panel, and ethics and governance board. Input from lived experience collaborators should be central to the governance of the consortium, particularly given the safety and ethical concerns that would need to be navigated to progress widespread use of AI in psychiatry. The unique commercial potential of AI systems may also open the door to diverse funding opportunities extending beyond the grant-based tradition of academic research. From a clinical research perspective, the leftmost domains of impact (Figure 1) could serve as initial priorities, as their comparably lower clinical risk could allow for faster implementation and benefit.

While such a consortium may address many research and cultural challenges, addressing regulatory barriers and facilitators would require coordinated government action. Working across international boundaries with differing legal and regulatory frameworks may prove challenging in the future; however, progress in this area should be shared across all continents where possible. Positively, two recent legislative developments in the EU provide the appropriate foreground for such an international psychiatry and AI consortium to emerge: the European Health Data Space Regulation (EHDS) and the EU AI Act. The EHDS will enable the reuse of health data through a coordinated framework for electronic health record systems, providing the digital backbone and scale necessary for high-quality AI research and implementation. Furthermore, the EHDS’s wide range of data modalities, expansive demographic coverage, and alignment with FAIR (Findable, Accessible, Interoperable, Reusable) principles could enable the development of globally generalisable AI systems. Simultaneously, the EU AI Act provides the regulatory governance necessary to ensure ethical AI system development. The Act outlines a risk-based classification for AI systems, with high-risk systems being subjected to strict compliance requirements during their development and operation. This legislation will be rolled out over the next 5 years, with the core rules governing the use of AI in psychiatry coming into effect from 2029 onwards. These progressive regulations provide an optimal environment for developing a highly impactful and ethical international consortium in AI and psychiatry.