The cutting edge of predictive modelling

In recent years, AI has gained increasing prominence in psychiatry, offering powerful tools to address digital health challenges such as early detection, risk prediction, treatment response and personalisation (Sheu Reference Sheu, Magdamo and Miller2023; Hansen Reference Hansen, Bernstorff and Enevoldsen2025). To fully harness these opportunities, it is essential to maintain a critical awareness of the specific roles and goals of various AI techniques. Aim, design and data source are not discrete elements and cannot be meaningfully disentangled: rather, they must be aligned to ensure methodological validity and interpretability.

The most widespread application of AI in mental healthcare lies in deploying its discriminative and predictive abilities, particularly through the branch that predicts outcomes such as treatment response, relapse or hospital readmissions among people with a diagnosed mental disorder, as well as the likelihood of disorder onset in at-risk populations, based on input data (e.g. clinical features, electroencephalography, magnetic resonance imaging, speech, and multimodal techniques integrating small changes in data modalities that shed light on the underlying pathophysiology of mental illness). These are most commonly operationalised through machine learning (ML) algorithms, capturing patterns in training data to enable inference and prediction on unseen data, while classifying clinically meaningful structures within relevant subgroups. Deep learning (DL), using multiple layers to uncover multifaceted relationships across variables, further expands these possibilities by enabling the analysis of more complex data, such as neuroimaging or individual speech patterns within large, unstructured data-sets (Graves Reference Graves, Mohamed and Hinton2013).

Beyond predictive abilities

However, predictive models may offer an overly simplified view of clinical care, often reducing it to binary classifications, which limits their practical utility in real-world contexts. This gap highlights the need for a better alignment with the complexity of clinical decision-making. Taken together, these considerations call for approaches of explainable AI (xAI) that not only provide more interpretable input–output relationships, but also enhance transparency in how inputs relate to the professional judgement of mental health practitioners. This would bring the links between algorithms’ behaviour and their inputs in line with the cognitive strategies underpinning clinical decision-making (Joyce Reference Joyce, Kormilitzin and Smith2023).

By providing a sufficient qualitative understanding of the process leading to algorithmic outputs, xAI approaches enhance both understandability and interpretability. Within the xAI framework, interactive visualisations may serve as key instruments: when combined with clinicians’ ability to detect potential errors or biases, they demonstrate that design features are integral to responsible AI development, with effective deployment inherently linked to the core AI application.

Consequently, interface design consistent with real-world clinical workflows, together with robust external validation, remains crucial to ensure safe, transparent and trustworthy implementation (Jia Reference Jia, McDermid and Lawton2022).

Such design considerations may also function as safeguards against the risks of AI misuse, including overreliance on ‘opaque’ models or misinterpretation of predictions in mental health contexts. When explainability mechanisms are weak or absent, AI systems in mental healthcare become vulnerable to misuse, thereby undermining clinical safety and eroding trust in AI-based digital mental health interventions.

Separately, the growing availability of multiple observations per person requires appropriate handling to address longitudinal coherence, which is particularly relevant for passive sensing and continuous monitoring purposes (Schick Reference Schick, Rauschenberg and Ader2023).

Moreover, predictive AI by itself can make only a limited contribution to interventional recommendations, since observed associations may not reflect cause-and-effect relationships. Although a causal framework is not strictly required for prediction per se, incorporating causal reasoning can substantially enhance interpretability and clinical actionability. Consequently, to enable the personalisation of digital mental health interventions, disentangling the relationships among predictors, outcomes and confounders may contribute to estimating and comparing potential outcomes under alternative interventions, thereby facilitating translation into clinical practice (Varidel Reference Varidel, An and Hickie2025). Therefore, in the context of mental healthcare, the focus on large-scale clinical trials, as presented by Lewin and colleagues, captures only a partial dimension of the wider scenario.

Although the training of algorithms frequently requires massive amounts of data, an investigative approach to AI may outline how the interaction of individual model components and data inputs can give rise to emergent collective behaviours, whose properties may diverge significantly from the patterns observed at the level of single variables or cases (Buscema Reference Buscema, Lodwick and Massini2025). This may underpin the transdiagnostic potential of AI in psychiatry and digital mental health, enabling the identification of clinically relevant features beyond the constraints of current categorical diagnostic systems in mental healthcare such as high comorbidity, intra-diagnostic heterogeneity and overlapping mechanisms across conditions. Multimodal AI frameworks that integrate traditional clinical assessments with objective digital signals (e.g. speech and language patterns, sleep–wake activity, mobility or social interaction metrics) can reveal quantifiable variations serving as digital signature markers. By capturing cross-cutting symptoms, these approaches can better account for the high prevalence of comorbidities, supporting more adaptive pathways of care. Uncovering latent dimensions through AI may ultimately support earlier risk detection, more precise stratification and personalised interventions.

On the other hand, enhancing the clinical relevance of applications remains crucial, as meaningful insights often depend on capturing disorder-specific patterns that may guide diagnosis, prognosis and treatment planning.