For most of psychiatry's modern history, mental illness has been understood as a problem of neurotransmitters. Depression is a deficit of serotonin. Schizophrenia is a dysregulation of dopamine. Treatment follows from that logic: adjust the chemistry, alleviate the symptoms.
That framework has produced real treatments for real people. It has also left a significant portion of patients without adequate options. Treatment-resistant depression affects roughly a third of people diagnosed with major depressive disorder. Bipolar disorder remains difficult to stabilise across its full course. The gap between diagnosis and effective intervention is not small. This is where metabolic psychiatry enters the picture, not as a replacement for existing models, but as an expansion of the biological territory under investigation.
The central hypothesis is straightforward, if still contested: disruptions in how the brain produces and uses energy may contribute to the development, progression, or expression of psychiatric symptoms. The brain consumes more than 20% of the body's energy despite representing only 2% of its volume, making it unusually sensitive to metabolic dysregulation. Recent research has begun exploring what happens when that metabolic state is deliberately altered.
The most studied intervention is the ketogenic diet, originally developed for epilepsy. By shifting the brain's primary energy source from glucose to ketone bodies, it changes the metabolic environment in ways that may influence neuronal efficiency, inflammation, and oxidative stress. A 2024 scoping review published in Frontiers in Psychiatry examined ketogenic metabolic therapy across several neuropsychiatric conditions and found signals of improvement in mood symptoms and metabolic health in a subset of patients (Campbell et al., 2024). A systematic review published in the Journal of Clinical Medicine in the same year reached broadly similar conclusions while emphasising that evidence remains preliminary and methodologically variable (Rogovik et al., 2024).
The limitation across almost all of this work is consistent: small sample sizes, open-label designs, and no large-scale randomised controlled trials. Researchers are cautious about what can be claimed. The question being pursued is not whether ketogenic diets treat mental illness in any general sense. It is narrower: whether metabolic state influences symptom expression in specific subpopulations, and if so, which ones.
This is where artificial intelligence has begun to make a practical difference. Machine learning models can integrate the kinds of heterogeneous data that metabolic psychiatry generates, clinical symptom profiles, metabolic markers, inflammatory indicators, neuroimaging, at a scale and resolution that previous analytical approaches could not manage. Recent work in computational psychiatry has identified latent patient subgroups that do not align neatly with existing diagnostic categories, suggesting that diagnostic boundaries in psychiatry may reflect the limits of available tools as much as the underlying biology (Venkatasubramanian et al., Nature Mental Health, 2024).
This matters particularly for metabolic psychiatry, where the underlying hypothesis is inherently multi-system. Rather than expecting a single biomarker to explain a psychiatric presentation, the field is beginning to model mental health conditions as networks of interacting biological processes: brain circuits, metabolic function, inflammation, and environment considered together. A 2025 review in Trends in Neurosciences examined inflammatory and metabolic correlates of cognitive dysfunction across diagnostic categories, finding that the same biological disruptions appeared across conditions otherwise separated by diagnosis (Wang et al., 2025).
The honest assessment is that both fields remain in an early translational phase. Machine learning models developed in one dataset frequently show reduced performance in independent populations, raising legitimate concerns about generalisation. The metabolic psychiatry evidence base is growing but not yet robust enough to support clinical protocols beyond carefully designed trials. These are not reasons to dismiss the research direction; they are reasons to take the methodology seriously.
For those building in digital mental health, the implications are worth sitting with. Tools designed around neurotransmitter models, whether for symptom tracking, intervention delivery, or outcome measurement, may be capturing only part of what is biologically relevant. If psychiatric conditions are increasingly understood as the result of interacting systems rather than single-pathway dysregulations, the data architectures, biomarker integrations, and personalisation approaches built into digital products may need to reflect that complexity. The convergence of metabolic research and computational psychiatry is not yet producing a unified new model of mental illness. It is, however, increasing the resolution of existing ones.
🪺