Challenges facing this approach
Despite the aforementioned strengths, there are several challenges facing this approach. As time progresses it may be difficult to integrate knowledge across studies without a community-agreed standardisation. For example, methods such as factor analysis (used to derive transdiagnostic symptom dimensions), are dependent on the self-report questionnaires used [3]. Furthermore, computational methods also require significant expertise, and non-experts may find the tools difficult to navigate.
Using transdiagnostic measurements in clinical trials focusing on multiple related diagnoses is an approach that may prove the utility of computational biomarkers. As treatments may differ between diagnoses, transdiagnostic trials could help clinicians to identify therapies which work across related diagnoses, and identify comorbid patients who could benefit from these therapies (Fig 1B). However, despite studies demonstrating that behaviours, on which diagnosis rests, can be characterised by computational parameters with specific neurobiological correlates [6, 7], patients with similar computational parameters may not all respond to a particular pharmaceutical agent. This group of individuals may also be heterogeneous and large experimental medicine studies will be required to establish the mapping between neurobiology, pharmacology and computation in order for the computational approach to realise its potential to improve person-specific treatments.
Implementing transdiagnostic tools in clinical practice may also prove challenging. However, these tools have the potential to complement current nosological approaches. Although symptom-based diagnoses overlap and may only provide a rough approximation of underlying pathology, symptoms are still the primary concern of the patient [2]. The complexity of individual psychiatric conditions will mean that clinical assessment will remain vital to providing holistic treatment and will also help to determine which particular investigative computational batteries e.g. questionnaires, tasks, imaging modality and specifications, would be most appropriate, in much the same way that a physician would run particular diagnostic biomarkers when there is suspicion of cancer. Because transdiagnostic factors are based on symptoms, they would augment diagnoses due to their demonstrable links to underlying psychopathology. As transdiagnostic factors and biomarkers develop, become standardised, and their predictive value is realised, greater weight can be placed on these investigative methods as diagnostic tools in psychiatry.
Finally, to realise the clinical potential of these tools, it will be imperative to make them both accessible and easy to operate in a low-resource environment. Factor analysis of self-report questionnaires may be one way to face this last challenge. Factor scores are robust and reliable if a consistent set of questions are used. If they are shown to be robustly related to specific computational descriptions of behaviour in cognitive tasks, the use of factors derived from self-report questionnaires could act as a proxy for computational biomarkers (Fig. 1B). This would preclude the need for time-consuming cognitive tasks and expensive imaging, making precision psychiatry more accessible. This is a particularly important consideration given the increased risk of mental illness in low socio-economic groups with limited access to well-funded mental health centres.
In summary, transdiagnostic computational phenotyping of mental illness may bridge the gap between biology and clinical practice, providing investigative and predictive tools which could precipitate a shift away from relatively crude group-average explanations that have had limited clinical utility in alleviating mental health symptomatology at the level of the individual.