When two patients receive the same psychiatric or neurological diagnosis, we tend to assume they share a common underlying pathology. However, traditional diagnostic frameworks rely on behavioral umbrellas—aggregate symptom scores that group vastly different biological realities under a single label.
Consider two individuals diagnosed with Major Depressive Disorder. One might experience severe insomnia and systemic inflammation, while the other experiences hypersomnia and no inflammatory changes. Because they meet the same behavioral criteria, standard clinical practice treats them as a monolith. This clinical heterogeneity creates a fundamental problem: when we average data across a diagnostic group, the distinct biological signals of individual subgroups cancel each other out.
To solve this, a major research direction focuses on biological subtyping (or stratification). By applying unsupervised machine learning to multimodal biological and behavioral data, researchers are moving from top-down clinical labels to bottom-up, data-driven "biotypes" or "endotypes." This approach explains not just what symptoms a patient has, but why they respond to certain treatments.
Traditional Approach:
[Diverse Pathologies] ───> [Single Behavioral Umbrella] ───> [One-Size-Fits-All Treatment] (High Failure Rate)
Biological Subtyping:
[Multimodal Data] ───> [Unsupervised Clustering] ───> [Distinct Biotypes] ───> [Targeted Interventions]
The Failure of the Group Average
Why has it been so difficult to find reliable biomarkers for conditions like autism, depression, or schizophrenia? The answer lies in the mathematical consequence of averaging heterogeneous populations.
In a study on Autism Spectrum Disorder (ASD), Lefebvre et al. investigated whether alpha brain waves could serve as a diagnostic neuromarker. When comparing the average alpha frequency of the ASD group against neurotypical controls, they found no significant difference. However, the ASD group exhibited massive variability. The condition was not characterized by a uniform shift in brain activity, but by a diverse spectrum of neural configurations. When a diagnostic category contains multiple distinct physiological drivers, a single group average represents no one.
This phenomenon extends across medicine. For example, traumatic brain injury (TBI) has historically been classified using blunt behavioral scales like "mild, moderate, or severe." To parse this noise, Holst et al. applied unsupervised clustering to intensive care data, discovering six distinct biological endotypes that transcend traditional severity scores. Patients within the same behavioral tier belonged to entirely different biological subgroups, explaining why identical injuries often lead to wildly unpredictable outcomes.
Mapping the Latent Space
To uncover these hidden subgroups, researchers use machine learning to map noisy clinical observations into a low-dimensional latent space—a mathematical representation where individuals are grouped by shared underlying mechanisms rather than surface-level symptoms.
Instead of relying on raw, aggregate questionnaire scores, Chen et al. developed a deep learning framework that integrates multi-domain outcomes to model unobserved "latent mental states." By training models to find the mathematical configurations that best predict a patient's overall functioning, they can separate individuals who look identical on paper but possess fundamentally different pathological profiles.
[Symptom 1: Insomnia] ──┐
[Symptom 2: Anhedonia] ──┼─> [Deep Learning Model] ─> [Latent Space Position] ─> Biotype A (Inflammatory)
[Biomarker: Cytokines] ──┘ (Hidden Configuration) ─> Biotype B (Neurocircuit)
This latent structure exists even within standard diagnostic tools. Varoquaux et al. applied machine learning to the gold-standard schizophrenia assessment questionnaire (PANSS). They discovered that a tiny subset of just 11 questions could map patients into three distinct, mathematically stable subtypes that transcend traditional diagnostic boundaries. The behavioral data itself contained the mathematical signatures of biological diversity, waiting to be unmixed.
From Biotypes to Precision Treatment
The ultimate validation of biological subtyping is its clinical utility: a biotype is only meaningful if it can predict how a patient will respond to a specific therapy.
This was demonstrated directly in a study on bipolar depression by Lee et al.. The authors used machine learning to sort patients into three distinct biotypes based on peripheral inflammatory biomarkers. When testing the anti-inflammatory drug infliximab, a standard clinical trial analyzing the whole group would have concluded the drug failed. However, by breaking the population into biotypes, the authors revealed that "Biotype 3"—characterized by a specific baseline inflammatory profile—experienced significant and sustained relief from anhedonia. The drug worked, but only for the biological subtype it was mechanistically suited to treat.
Similarly, in neurodevelopmental research, Daniel Y.‐J. Yang et al. (paper) used functional MRI to measure brain activity in response to biological motion before children with autism began behavioral therapy. They found that baseline neural connectivity patterns could accurately forecast which children would show clinical improvement after a 16-week intervention.
To scale these insights from isolated studies to standard practice, large-scale international consortia are now required. As outlined by Frouin et al. regarding the EU-AIMS Longitudinal European Autism Project (LEAP), the field is moving toward integrating genomics, neuroimaging, eye-tracking, and deep phenotyping across thousands of individuals. The goal is to establish robust, reproducible "stratification biomarkers" that allow clinicians to abandon trial-and-error medicine in favor of targeted, biology-driven interventions.