Psychiatry still diagnoses by conversation. A Cambridge review maps the path to something better.

A cardiologist measures troponin. An oncologist sequences a tumor. A psychiatrist asks whether you have felt sad for two weeks. This is not a criticism of psychiatry. It is a description of the field's central diagnostic paradox, and a new review from the University of Cambridge confronts it head-on.

Published in Brain Medicine, the invited review by Dr. Jakub Tomasik, Jihan K. Zaki, and Professor Sabine Bahn at the Cambridge Centre for Neuropsychiatric Research synthesizes an enormous body of emerging research to map a translational pathway from where psychiatric diagnosis stands today toward where it might arrive. The portrait is honest, detailed, and deliberately cautious about timelines.

Why 250 symptom combinations do not make a diagnosis

The Diagnostic and Statistical Manual of Mental Disorders (DSM) and the International Classification of Diseases (ICD) standardized psychiatric language and improved diagnostic reliability. But they were built on expert consensus, not biological discovery. Major depressive disorder, the review notes, can be diagnosed through more than 250 possible symptom combinations. Two patients with identical diagnoses may present with entirely different clinical pictures. The thresholds separating illness from normal variation are, in the authors' assessment, largely arbitrary.

The consequences are practical: diagnostic labels often fail to predict which patients will respond to which treatments. Comorbidity is pervasive. And the subjective nature of symptom assessment introduces variability that would be unacceptable in most other branches of medicine.

Four frameworks that challenge the checklist

The review examines four conceptual alternatives. Network models treat psychiatric symptoms as interacting systems rather than passive reflections of a hidden disease. The Hierarchical Taxonomy of Psychopathology (HiTOP) organizes mental illness into data-driven dimensions. The Research Domain Criteria (RDoC) redefines disorders by neurobiological mechanisms. And clinical staging introduces temporal progression from early vulnerability through chronic disease.

None is ready for clinical deployment. Network structures often fail to replicate across samples. HiTOP is too complex for routine use. RDoC has been criticized for ignoring social context. Clinical staging is hampered by the fact that many mental disorders do not follow a predictable course. But together, they represent a shift from viewing disorders as fixed categories toward seeing them as dynamic, multidimensional systems.

The molecular signals assembling beneath the surface

Biomarker research has produced genuine signals. ENIGMA Consortium neuroimaging shows widespread cortical thinning in schizophrenia and more localized reductions in depression. GWAS from the Psychiatric Genomics Consortium have identified hundreds of risk loci converging on synaptic transmission and calcium signaling. Cross-disorder analyses reveal high genetic overlap between schizophrenia and bipolar disorder.

A few tools have reached the clinic. The VeriPsych proteomic panel, developed in Professor Bahn's laboratory, was validated and commercialized in the US for confirming recent-onset schizophrenia, though it was later withdrawn due to cost and limited uptake. The EDIT-B RNA-editing blood test has achieved European regulatory marking for distinguishing bipolar from unipolar depression. These are rare crossings from bench to bedside, and the review presents them as proof of concept rather than harbingers of imminent transformation.

When your phone becomes a clinical instrument

Digital phenotyping extends biological approaches by capturing what static biomarkers miss: change over time. Smartphone geolocation data can reveal reduced mobility associated with depression severity. Wearable sleep-wake data show that circadian rhythm shifts predict mood episodes in bipolar disorder. Speech recordings detect altered intonation and pause rates in depression. Even social media posts carry diagnostic signal.

The review argues that ecological momentary assessment, delivered through apps that prompt mood and energy reports multiple times daily, could eventually complement or partly replace the static snapshot of a clinical interview. But most digital markers have been derived from small cohorts, show modest effect sizes, and lack robust validation.

AI as translator, not oracle

Machine learning occupies the integrative layer of the translational pathway. Transformer architectures and multimodal models combining genomics, neuroimaging, and clinical text are being developed. But the review separates aspiration from achievement carefully. The largest publicly available psychiatric datasets include fewer than 1,000 samples. Clinical data remain rarely shared. And until AI reasoning can be made transparent and interpretable, deployment will face justified scrutiny.

The barriers that will not disappear on their own

The review's most valuable feature may be its refusal to oversell. It catalogs implementation challenges systematically: limited biomarker reproducibility, poor generalizability from research to real-world populations, regulatory uncertainty, algorithmic bias from homogeneous training data, fragmented data infrastructure, clinician resistance, and absent reimbursement pathways.

Even promising tools risk deepening health disparities if deployed mainly in well-resourced settings. Federated learning, which trains AI across decentralized datasets without sharing raw data, offers one path forward, but its implementation in mental health remains limited.

The review converges on a measured conclusion: psychiatry needs not a single leap but an accumulation of validated, interpretable, and accessible tools integrated into real-world health systems. The immediate promise lies in inflammatory markers that identify treatment-responsive subtypes, circadian data that flag impending mood episodes, and AI tools that reduce diagnostic delay. The longer vision involves diagnostic categories that evolve into empirically defined subtypes reflecting underlying mechanisms, much as biomarker classification transformed oncology.

The field has spent decades generating raw materials. The work that remains is integration, translation, and the unglamorous labor of implementation.