Brain Biomarkers for Mental Health: Research Evidence and Limitations
Author: Ukrainian Psychological Hub · Published: October 12, 2026 · Editorial Policy
Brain biomarkers are measurable features of brain structure, function, chemistry, or physiology investigated as indicators of mental health processes, psychiatric symptoms, treatment response, or future outcomes. Their scientific importance is substantial: they allow researchers to test biological hypotheses, quantify variation, and investigate mechanisms that cannot be inferred from symptoms alone. Their clinical meaning depends on a different question: does a particular measurement improve an identifiable decision for an individual person, in the population and setting where that decision will be made?
The current evidence supports reproducible research findings for some group-level brain differences and a growing body of transdiagnostic research. It does not establish a general brain scan, EEG pattern, or algorithm that can independently diagnose depression, anxiety, bipolar disorder, PTSD, schizophrenia, ADHD, or another psychiatric condition in ordinary clinical practice. That distinction follows from the FDA–NIH BEST biomarker framework, the American Psychiatric Association's review of clinical neuroimaging, and large, multisite prediction studies such as ENIGMA-Anxiety.
What is a brain biomarker in mental health?
A biomarker is an objectively measurable characteristic indicating a biological process, a pathogenic process, or a response to an exposure or intervention. The FDA–NIH definition includes imaging and physiological characteristics. It also distinguishes biomarkers from clinical outcome assessments, which directly evaluate how a person feels, functions, or survives. A patient reporting distress, impairment, or symptom change is therefore providing clinically significant information of a different kind from an MRI-derived cortical thickness estimate or an EEG signal. Both can inform science; they answer different questions. See the BEST glossary.
In psychiatric neuroscience, a candidate brain biomarker may be regional anatomy measured by structural MRI, functional connectivity estimated from fMRI, an event-related EEG potential, a PET measure of molecular binding, or a reproducible combination of measurements. A composite biomarker may integrate several features by means of a prespecified statistical algorithm. The term candidate matters: identifying a statistically associated variable is a discovery result, whereas validating a clinical biomarker is a program of measurement, prediction, replication, and decision research.
A biomarker also has a defined context of use. A signal useful for investigating average differences between research groups may be unsuitable for recognizing a disorder in an individual. A measure useful for enriching a research trial may be unsuitable for deciding whether to prescribe treatment. Even a biomarker that predicts outcome in one clinical population may fail in another. The FDA's biomarker qualification guidance requires that claims be interpreted in relation to their specific proposed use.
Seven biomarker functions are different clinical questions
The FDA–NIH BEST framework identifies susceptibility/risk, diagnostic, monitoring, prognostic, predictive, pharmacodynamic/response, and safety biomarkers. A susceptibility marker estimates the likelihood of developing a condition; a diagnostic marker concerns whether a specified condition is present. Monitoring assesses biological status over time. Prognostic markers estimate future outcomes independently of which treatment is selected. Predictive markers identify differential benefit or harm from a particular intervention, while pharmacodynamic and safety markers concern biological responses and adverse effects. These categories are related but cannot be exchanged. A correlation with symptom severity does not make a measurement a diagnostic test, and a change after treatment does not make that change a validated treatment-selection rule.
A particularly important distinction is predictive versus prognostic validity. If individuals with one imaging pattern recover more often than individuals with another pattern regardless of treatment, the measurement may be prognostic. Demonstrating that one group benefits specifically from treatment A rather than treatment B requires evidence about the interaction between biomarker status and treatment, preferably in rigorously designed comparative trials. This difference is explained in the FDA–NIH BEST resource.
How did the field of psychiatric brain biomarkers develop?
Clinical psychiatry has historically organized many conditions around patterns of symptoms, duration, distress, and functional consequences. With the growth of structural MRI, functional MRI, PET, EEG, and quantitative neuroscience, investigators sought biological measurements that could clarify mechanisms and eventually support decisions. Early case-control reports sometimes attracted attention because they identified apparently distinctive regional changes or impressive machine-learning accuracies. Such results opened research programs, but also exposed challenges involving small samples, selective analysis, clinical heterogeneity, and performance outside the original cohort.
The National Institute of Mental Health's Research Domain Criteria (RDoC) advanced a complementary research framework centered on dimensions of functioning—including emotion, cognition, reward, and arousal—measured across molecular, circuit, behavioral, and experiential levels. RDoC is expressly a research framework; it is not a diagnostic guide. Its transdiagnostic logic is relevant because a measurable circuit pattern may relate to a dimension of experience occurring in several diagnostic categories.
Large collaborations and more stringent replication methods have substantially refined the field. The 2022 brain-wide association study by Marek and colleagues found that many correlations between common brain measures and mental health or cognitive phenotypes are smaller than earlier literature suggested; reproducible population-level associations often required samples numbering in the thousands. This result strengthened the case for better-powered research, careful generalization, and explicit separation of group discoveries from individual applications.
Which brain measurements are investigated?
Structural MRI: anatomy, volume, and cortical organization
Structural MRI uses magnetic signals and radiofrequency pulses to reconstruct anatomical images. Researchers estimate regional volumes, cortical thickness, surface area, and other structural properties using segmentation and computational models. Diffusion MRI derives information about water displacement that can inform models of white-matter organization; tractography estimates candidate pathways from diffusion data and does not directly visualize individual axons or prove the direction of signaling.
Structural measurements are useful for studying development, aging, illness-associated variation, and neuroanatomical patterns across samples. Their interpretation depends on preprocessing, head motion, scanner and site effects, age, sex-related variables, medication exposure, and other factors. Structural differences across groups cannot by themselves establish the cause of a psychiatric syndrome. The research role and inferential limits of these approaches are discussed in Poldrack and Farah's review. A separate English Hub overview explains what MRI, fMRI, PET, and EEG measure.
Functional MRI: blood-oxygenation signals and modeled connectivity
Functional MRI typically measures blood-oxygen-level-dependent, or BOLD, signals. These reflect complex relationships among local neural activity, oxygen metabolism, blood volume, and vascular responses. The BOLD signal is not a direct recording of action potentials, neurotransmitter concentrations, thoughts, or subjective experience. In task fMRI, investigators estimate signal changes associated with experimental conditions. In resting-state studies, correlations between time series from different regions are used to characterize functional connectivity; correlation alone does not specify a direct anatomical connection, a causal pathway, or a fixed psychological function.
Group-average task effects can be robust under appropriate conditions, while the stability of an individual's relative activation is a separate measurement property. In a meta-analysis of task-fMRI reliability, Elliott and colleagues found an average test–retest intraclass correlation of approximately 0.40 across the studies they examined, with substantial variability across tasks and settings. This limits the readiness of many standard task measures for person-level stratification; it does not invalidate all fMRI research or every possible imaging measure.
EEG and event-related potentials: electrical activity at the scalp
Electroencephalography detects voltage differences at scalp electrodes generated primarily by synchronized electrical currents in populations of neurons, shaped by tissue conduction and recording conditions. EEG has excellent millisecond-scale temporal resolution and is valuable for quantifying oscillatory activity, event-related responses, vigilance, and changes over time. Source localization is an inverse problem: distinct neural configurations can produce similar scalp measurements.
Mental health studies have investigated spectral power, asymmetry, event-related potentials, connectivity estimates, and machine-learned combinations. Performance can vary with sleep, arousal, medication, electrode placement, reference choice, artifacts, and preprocessing. An EEG pattern associated with a diagnostic group is not automatically sufficiently specific or stable to classify an individual. EEG has established medical roles in conditions such as epilepsy; those clinical applications must be distinguished from claims that an EEG printout diagnoses a psychiatric syndrome. For wider method comparison, see Brain Imaging Methods.
PET: radiotracer distributions and molecular targets
Positron emission tomography detects emissions associated with administered radiotracers. Depending on the radioligand, analytical model, and acquisition, PET can estimate aspects of metabolic activity or target binding, including some receptor or transporter systems. Estimates depend on tracer kinetics, reference regions, nonspecific binding, partial-volume effects, and medication-related occupancy. A PET-derived binding measure is not a direct readout of happiness, motivation, serotonin sufficiency, or dopamine 'levels' in any simple psychological sense.
PET has important established medical applications in selected neurological and other conditions. Psychiatric research uses molecular imaging to investigate hypotheses and medication effects, yet the existence of a receptor-related group difference does not automatically provide a practical, validated diagnosis. Radiation exposure, cost, specialized infrastructure, and tracer specificity also shape the feasibility and risk–benefit balance of possible uses.
MEG, spectroscopy, network measures, and multimodal panels
Magnetoencephalography measures extremely weak magnetic fields associated with neural currents; it offers high temporal resolution and faces source-reconstruction constraints. Magnetic resonance spectroscopy estimates selected metabolites within relatively large tissue volumes, subject to spectral overlap and modeling limits. Network and connectome features are derived computational descriptions, not independent biological substances. Multimodal panels may combine MRI, electrophysiology, behavior, and other measurements, but each added feature increases data-handling demands, potential leakage routes, and validation requirements.
An elaborate algorithm does not transform weak, unstable inputs into a reliable clinical tool by default. It must demonstrate reproducible added value against straightforward assessments and existing clinical practice in the population where it will be deployed.
What does the evidence show across mental disorders?
Group differences are scientifically meaningful but diagnostically overlapping
Research has documented average structural and functional differences associated with several psychiatric categories. These findings can identify statistical patterns, shape mechanistic hypotheses, and motivate further studies. Yet distributions commonly overlap: many people meeting a diagnosis do not show a proposed pattern, and some people without that diagnosis do. Group-level statistical significance and individual discrimination measure different properties.
Cross-disorder results make this limitation visible. In a 2026 systematic transdiagnostic meta-analysis, Qin and colleagues examined voxel-based morphometry findings spanning 221 studies and more than 22,000 participants across major depression, bipolar disorder, and schizophrenia. They reported both shared and disorder-related patterns of gray-matter differences. The synthesis helps map research associations; it does not convert those group maps into an independently validated clinical scan for one person.
The significance of transdiagnostic findings extends beyond whether each diagnosis can be represented by a unique anatomical signature. Similar circuit-associated variations may participate in overlapping processes, arise from shared exposures, or reflect consequences of symptoms, treatment, and other factors. Research must adjudicate these explanations with longitudinal, intervention, and converging-method evidence.
Large studies put individual classification into perspective
The 2024 ENIGMA-Anxiety machine-learning analysis assembled structural MRI data from 3,343 young people across 32 sites. Although investigators obtained some statistically meaningful classifications, reported area-under-the-curve values for selected anxiety-related comparisons were approximately 0.59–0.63. An AUC of 0.50 represents chance-level ranking for the specified comparison, while 1.00 represents perfect ranking. Thus, these results indicate modest discrimination under studied conditions, not a deployable stand-alone diagnostic test.
Such large, heterogeneous datasets are especially informative because methods can be assessed across differences in recruitment, age, treatment exposure, and scanning sites. A small study reporting much higher classification accuracy may have evaluated a narrower sample, experienced data leakage, or capitalized on its own idiosyncrasies. A method's performance on unusually clear cases versus carefully selected controls rarely answers how it will perform when a clinician must distinguish several plausible diagnoses and other causes of symptoms.
Depression 'biotypes' and the replication problem
An influential research direction sought to classify depression according to functional connectivity patterns. The scientific question was worthwhile: could reproducible neurobiological subgroups guide treatment or explain heterogeneity? A 2019 methodological replication by Dinga and colleagues examined a previously proposed connectivity-based biotype approach in a different, more heterogeneous sample. Although their procedures produced apparently high canonical correlations and clusters, formal testing did not support the claimed statistical significance or stability of those clusters in that replication.
This case illustrates why exploratory clustering and visually appealing separation are insufficient. Researchers must show that subtypes are stable across samples, biologically interpretable, related to meaningful outcomes, and prospectively useful. Even successful statistical replication would be an intermediate milestone before recommending subgroup-specific treatment decisions. Readers seeking a disorder-specific account can consult the English Hub article Depression and the Brain; the present article owns the cross-diagnostic biomarker-validity question.
What about anxiety, PTSD, bipolar disorder, psychosis, OCD, and neurodevelopmental conditions?
These areas contain important findings about threat learning, reward processes, large-scale networks, neurodevelopment, and altered group-level physiology. Clinical presentations, however, are diverse within each category and often overlapping across categories. A biomarker selected because it differs between an extreme case group and unusually healthy controls may perform poorly in a clinic where common alternatives include sleep problems, medication effects, substance exposure, neurological illness, and several co-occurring syndromes.
The English Hub maintains independent, condition-owned explanations of Anxiety and the Brain, Bipolar Disorder and the Brain, PTSD and the Brain, and OCD and the Brain. Those articles address the particular conditions, while this one addresses what evidence is required to validate a brain-derived marker across disorders.
Sensitivity, specificity, AUC, calibration, and clinical value
A diagnostic test's sensitivity is the proportion of people with the specified target condition whom it identifies as positive. Specificity is the proportion without that condition whom it identifies as negative. Positive predictive value is the proportion of positive results that truly represent the target condition in the population tested; negative predictive value gives the corresponding interpretation of a negative result. Predictive values depend on the underlying prevalence, even when sensitivity and specificity stay constant.
Consider an illustrative test with 80% sensitivity and 90% specificity used in 1,000 people, 10% of whom have the target condition. Of 100 people with the condition, approximately 80 test positive. Of 900 without it, approximately 90 also test positive. Consequently, 80 of 170 positive results are true positives: positive predictive value is about 47%. These figures are hypothetical, chosen to explain base rates, and are not performance estimates for any currently proposed psychiatric biomarker.
AUC summarizes ranking performance across different thresholds. It does not, by itself, specify a clinically acceptable threshold, the reliability of an individual's result, calibration of predicted probabilities, or benefits and harms of acting on them. Two models with similar AUC may give differently calibrated risk estimates. A high AUC in case-control data can still be misleading when the clinical base rate, case mix, and comparison conditions change.
For a clinically persuasive marker, researchers should report sensitivity and specificity with uncertainty intervals at prespecified operating thresholds; positive and negative predictive values in a representative population; calibration; and decision consequences. They should also test whether the marker adds information beyond symptoms, medical history, clinical interview, and simpler measures. A marginal improvement in statistical fit may carry no meaningful benefit if the test is costly, difficult to obtain, or prone to consequential errors.
From a candidate signal to a validated clinical biomarker
Analytical validation: can the method measure consistently?
Analytical validation addresses technical performance. For an MRI metric, this includes reproducible acquisition and processing across scanners, software versions, head coils, and sites. For EEG, it includes recording conditions, preprocessing choices, electrode configurations, and artifact handling. PET requires attention to radiotracer characteristics and quantification models. Repeatability and test–retest reliability must be evaluated on the intended scale and time interval.
High precision is necessary but insufficient. A cortical measurement may be technically repeatable yet clinically uninformative for diagnosing depression. Conversely, a statistically predictive signal that fluctuates dramatically on retesting may be unsuitable for decisions about an individual. The FDA–NIH validation framework distinguishes analytical validation from the next stages of evidence.
Clinical validation: does it detect the stated concept?
Clinical validation tests whether the proposed measure reliably identifies or predicts the intended clinical construct. This requires a clearly defined target, an appropriate reference standard, prespecified modeling procedures, clinically representative participants, and genuinely independent evaluation. Investigators should report discriminative performance as well as uncertainty, subgroup performance, calibration, and the consequences of important confounders.
External validation means testing on participants or settings independent of model development, ideally including sites, demographic groups, and acquisition circumstances not represented in the training process. Site-wise partitioning can be particularly important: if training and test sets share subtle characteristics of the same scanner, institution, or enrollment pathway, a model may exploit those shortcuts rather than learn transportable biological features. All preprocessing, feature selection, normalization, and hyperparameter tuning must be nested within the training procedure to prevent leakage into evaluation data.
Clinical utility: does its use improve decisions or outcomes?
A test may be technically reproducible and statistically predictive yet offer no clinical advantage. Clinical utility asks whether a specified decision informed by the test improves patient-relevant outcomes, reduces harm, enables more effective allocation of care, or delivers justified benefits relative to costs and alternatives. A proposed treatment-selection biomarker should demonstrate that biomarker-guided care produces better outcomes than a sensible standard-care approach, while accounting for access and equity.
The FDA–NIH BEST definitions distinguish validation, qualification for a particular regulatory context, and utility. A test validated as a research outcome marker is not thereby approved for general diagnostic marketing, and regulatory qualification in one context does not authorize unrelated clinical claims. The FDA's April 2026 biomarker-method validation guidance further underscores method-specific evaluation for measurements used in medical product development.
Why promising findings so often fail to translate
Small effects, sample size, and selection
Mental health phenotypes are multidetermined and noisy. Symptoms fluctuate with life circumstances, sleep, stress, medication, development, and environment. Brain-derived variables likewise vary with data acquisition and biology. Associations between these two complex levels may be real but small. The Marek and colleagues brain-wide association analysis demonstrates how small studies can yield inflated effect sizes and unstable associations, and why larger samples materially improve reproducibility for between-person associations.
Sample size alone is not enough. Participant selection must reflect the target use: clinic attendees with overlapping symptoms are harder to distinguish than selected cases and controls. Case-control sampling also affects prevalence-dependent metrics and may mask a shift in the kinds of patients who present for care.
Heterogeneity, comorbidity, and developmental context
A diagnostic category may contain several symptom profiles, trajectories, and environmental histories. Conversely, fatigue, concentration difficulty, altered arousal, or reward-related symptoms may arise across multiple categories. Age, pubertal development, chronic conditions, trauma exposure, and social context can change the relation between brain measurements and mental health outcomes. Averages may conceal different directions of change in individual subgroups.
Transdiagnostic research is useful when it asks precise questions about functioning and mechanisms. It becomes clinically meaningful only when its measurements are reproducible and its predictions useful for particular decisions. The NIMH RDoC framework intentionally integrates biological, behavioral, and self-report information rather than treating a single brain measure as an exhaustive account of mental health.
Confounding, reverse inference, and causation
Medication, substance use, sleep, movement, socioeconomic exposures, prior episodes, medical illness, and imaging site can be associated with both symptoms and brain measures. A cross-sectional difference does not establish whether a brain variation preceded the condition, emerged during its course, reflected treatment, or arose from a shared third factor. Longitudinal studies improve temporal inference but do not automatically establish causation.
Reverse inference occurs when a researcher treats activation in a region associated with some psychological process as proof that the process occurred. Regions and networks participate in multiple functions, and psychiatric symptoms recruit distributed, interacting systems. Brain abnormalities in a statistical model are rarely uniquely diagnostic of any one subjective experience. The Poldrack and Farah methodological review explains the distinction between observing a neural signal and inferring psychological mechanisms.
Technical shifts and algorithmic overfitting
Scanner hardware and acquisition sequences, EEG montages, preprocessing software, head-motion handling, and machine-learning choices influence measured features. A model may inadvertently learn institution-specific patterns, demographic proxies, or technical artifacts. Cross-validation conducted within one dataset can overestimate performance if repeated measurements from the same person, related participants, or preprocessing-derived information leak between folds.
A credible paper should make its analysis transparent, register confirmatory hypotheses where appropriate, make validation data genuinely independent, document missingness, and explain model failures. Researchers should ask whether a simpler model or routine clinical measurement performs as well. Algorithmic complexity and a colorful neural map are presentation characteristics, not substitutes for independent validation.
Brain biomarkers and treatment personalization
Treatment-response research explores whether an individual's network connectivity, EEG activity, brain anatomy, or molecular imaging pattern could help select medication, psychotherapy, neurostimulation, or another intervention. This is an important direction because treatment outcomes differ substantially across people and repeated trial-and-error can be burdensome. Yet improvement after an intervention, or association of a baseline signal with eventual improvement, does not prove that a biomarker-guided choice would outperform current care.
A valid predictive marker must be evaluated across meaningful treatment alternatives. If a baseline network feature predicts response equally well to several therapies, it may indicate general prognosis or other factors rather than treatment-specific benefit. Randomized comparisons and prospective biomarker-guided studies provide stronger evidence than retrospectively selected patterns. A marker associated with a neuromodulation target's physiology likewise does not establish its value as a stand-alone psychiatric diagnostic test.
Brain stimulation itself is a clinical intervention or research manipulation with indications, contraindications, and evidence specific to technique and condition. Therapeutic status must be evaluated separately from the promise of a proposed imaging or EEG selection rule. Readers should ask whether a published response predictor has been independently validated and whether using it demonstrably improves patient-relevant outcomes.
How can machine learning and AI help?
Machine learning can integrate measurements across many regions and time points, account for nonlinear associations, and derive probabilistic models from complex datasets. Artificial intelligence can help standardize extraction, identify candidate patterns, and estimate predictions. These capabilities expand research possibilities, especially in multisite projects. Their value is determined by valid targets, reliable inputs, sound data partitions, external calibration, clinical comparison, and prospective impact.
A training model may distinguish selected research participants while still lacking a mechanism-specific explanation. Feature importance maps can change across correlated predictors and algorithms; they do not prove which region causes a mental state. Neural decoding of a task or category also addresses a different outcome from psychiatric diagnosis. Predicting a label present in a training dataset does not by itself validate the biological meaning, clinical correctness, or ethical acceptability of the label.
Fairness and transportability require direct analysis. Models trained on unrepresentative age groups, languages, racialized social contexts, health systems, or scanning settings can generate systematically different error rates. Privacy and governance matter because neural and clinical data may be sensitive and reusable. Clinical use requires meaningful consent, security controls, explainable responsibilities, and robust monitoring after implementation.
Are psychiatric brain biomarkers used routinely in diagnosis?
At present, brain imaging and EEG have substantial established roles for specific neurological and medical indications, but ordinary psychiatric diagnoses are generally made through clinical assessment under established criteria rather than a universal neuroimaging or EEG signature. If a clinician orders MRI, EEG, or another neurological investigation during a mental health assessment, the test may help evaluate seizures, masses, neurodegenerative disease, or another medical explanation when indicated. The medical value of that investigation should not be confused with a scan that verifies depression, anxiety, PTSD, bipolar disorder, or ADHD.
The American Psychiatric Association's clinical neuroimaging resource reviews the distinction between scientifically informative findings and usable diagnostic or treatment-response instruments. Current research can update estimates and identify niche contexts, so evidence should be checked for the specific test, claim, population, and date. Marketing statements that promise precise personal diagnosis from a single colorful scan require documentation of validation and demonstrated clinical utility.
Mental health evaluation remains grounded in symptoms, functional consequences, context, longitudinal history, and appropriate differential assessment. Brain research can inform such care without being treated as the sole scale of explanation for relationships, meaning, culture, adversity, or life circumstances.
How should patients interpret a proposed brain-based test?
Ask what specific clinical decision the test is intended to change. Determine whether it is a research measure, an approved device for a narrowly defined indication, an adjunct in a differential medical investigation, or a validated diagnostic or treatment-selection tool for the stated psychiatric use. The manufacturer's terminology, media coverage, and a clinician's enthusiasm do not determine evidentiary status.
Ask for independent studies reporting performance on representative patients, including sensitivity, specificity, predictive values, uncertainty, and comparison with standard assessment. Check whether the proposed classification was evaluated prospectively and whether treatment decisions based on it improved outcomes. Discuss possible false positives, false negatives, incidental findings, financial costs, and what happens when a result conflicts with clinical history.
People who have undergone a brain scan or EEG should have its results interpreted within the ordering clinician's specific medical question. An unusual numerical value, asymmetry, activity pattern, or algorithmic score is not an individual psychiatric diagnosis without suitable validation. Established medical concerns, changes in neurological functioning, or severe symptoms warrant assessment based on their clinical presentation, independently of experimental biomarker claims.
Common misconceptions
A brain image that differs from the average proves mental illness
Statistical maps describe estimated differences or associations under particular analytic assumptions. Individual brains vary naturally, and overlapping distributions limit classification. Interpretation requires an appropriate population, measurement reliability, diagnostic reference, and evidence about alternative explanations.
A strong association proves causation
Associations can be influenced by treatment, chronicity, development, behavior, and third variables. Causal hypotheses require designs capable of distinguishing mechanisms, such as appropriately controlled interventions, longitudinal evidence, and converging experiments.
Artificial intelligence can convert any scan into a clinical diagnosis
Algorithms may learn site artifacts, selection patterns, or diagnostic labels whose biological relevance is uncertain. A high score on reused data is not external validation. Real-world usefulness requires independent testing, calibration, an explicit decision threshold, and demonstrated benefits.
Dopamine or serotonin imaging explains a person's entire emotional state
Neurotransmitter systems perform many functions and are regulated through multiple receptor populations and circuits. A PET binding estimate does not directly quantify subjective happiness, emotional worth, or a presumed universal chemical imbalance. Emotional and psychiatric phenomena also involve developmental, behavioral, social, and contextual levels of explanation.
Research Domain Criteria has replaced psychiatric diagnostic manuals
NIMH explicitly identifies RDoC as a research framework. It provides ways to investigate dimensions of function and biological systems; its constructs are not themselves a routine clinical diagnostic replacement.
Frequently asked questions
Can an MRI scan diagnose depression?
Structural or functional MRI can support neuroscience research on depression and may be ordered for other medical reasons when clinically indicated. No single MRI-derived signature has been established as a general stand-alone depression diagnostic test. An individual interpretation must be based on a test validated for that precise clinical purpose.
Can EEG diagnose anxiety or ADHD?
EEG records electrical activity and may yield research features associated with attention or arousal. This does not establish that any generic EEG pattern can independently diagnose anxiety or ADHD. The evidence and regulatory status of a specific device and claimed indication must be examined separately.
What is the difference between a brain biomarker and a brain scan?
A scan is a procedure producing measurements or images. A candidate biomarker is a specific feature or algorithm derived from measurements and proposed to indicate a defined biological or clinical concept. The existence of a scan does not automatically validate any proposed biomarker extracted from it.
Could brain biomarkers help choose antidepressants?
Treatment personalization is an active research goal. A useful predictor must show reproducible, treatment-specific differences and improve decisions beyond ordinary care. Baseline association with recovery is insufficient to establish that a particular medication is the best choice for an individual.
Why can a study find significant brain differences but fail to diagnose individuals?
Statistical significance addresses evidence against a specified population-level null hypothesis. Individual diagnosis depends on how much the measurements of people with and without the condition overlap, whether findings reproduce, and what errors occur at clinically relevant thresholds.
Is a 90% accurate AI model ready for clinical use?
The figure alone is inadequate. It may come from an imbalanced dataset, selected cases and controls, data leakage, or an internal test that does not reflect real practice. A clinical evaluation must specify the target population, comparator, thresholds, sensitivity, specificity, calibration, prospective utility, and independent validation.
Does a brain biomarker prove a disorder is purely biological?
Psychiatric conditions involve processes expressed through biological systems and lived in developmental, environmental, social, and cultural contexts. A brain association specifies one layer of measurement. It does not establish exclusive causation or eliminate the importance of psychological and social understanding.
Are there any genuine medical reasons to order brain imaging during psychiatric assessment?
Yes. Certain presentations justify investigation for neurological or other medical explanations. The indication, imaging modality, and interpretation are determined clinically. This is separate from using neuroimaging as a routine stand-alone test for common psychiatric categories.
Research priorities and future directions
The most productive next steps are large and diverse longitudinal samples, harmonized acquisition, systematic measurement-reliability studies, external site and population validation, transparent and reproducible modeling, and prospective trials of decision impact. Subtyping claims should establish stable boundaries and clinically meaningful consequences; individual risk models should present uncertainty and demonstrate calibration. More precise biological characterization may eventually support carefully bounded diagnostic, prognostic, or treatment-predictive applications.
Future progress will likely draw on multiple levels of evidence rather than one anatomical region or one chemical messenger. The essential standards remain consistent: a clear concept, a reliable measurement, independent clinical validation for a stated purpose, and evidence that acting on the result provides worthwhile benefit. Those standards are what turn intriguing neuroscience into responsible clinical practice.
Related Articles
References
American Psychiatric Association. (2018). Clinical applications of neuroimaging in psychiatric disorders. American Journal of Psychiatry. Official resource.
Bruin, W. B., Zhutovsky, P., van Wingen, G. A., et al. (2024). Brain-based classification of youth with anxiety disorders: Transdiagnostic examinations within the ENIGMA-Anxiety database using machine learning. Nature Mental Health, 2, 104–118. https://doi.org/10.1038/s44220-023-00173-2.
Dinga, R., Schmaal, L., Penninx, B. W. J. H., et al. (2019). Evaluating the evidence for biotypes of depression: Methodological replication and extension of Drysdale et al. (2017). NeuroImage: Clinical, 22, 101796. https://doi.org/10.1016/j.nicl.2019.101796.
Elliott, M. L., Knodt, A. R., Ireland, D., et al. (2020). What is the test–retest reliability of common task-functional MRI measures? New empirical evidence and a meta-analysis. Psychological Science, 31, 792–806. https://doi.org/10.1177/0956797620916786.
FDA–NIH Biomarker Working Group. (2016–present; glossary updated 2025). BEST (Biomarkers, EndpointS, and other Tools) Resource. U.S. Food and Drug Administration and National Institutes of Health. https://www.ncbi.nlm.nih.gov/books/NBK338448/.
FDA–NIH Biomarker Working Group. (2020 update). Validation. BEST (Biomarkers, EndpointS, and other Tools) Resource. https://www.ncbi.nlm.nih.gov/books/NBK464453/.
Marek, S., Tervo-Clemmens, B., Calabro, F. J., et al. (2022). Reproducible brain-wide association studies require thousands of individuals. Nature, 603, 654–660. https://doi.org/10.1038/s41586-022-04492-9.
National Institute of Mental Health. (2014). Biomarker development and validation: Establishing standards of evidence for their context of use in clinical trials. Official resource.
National Institute of Mental Health. (Current framework). About Research Domain Criteria (RDoC). Official resource.
Poldrack, R. A., & Farah, M. J. (2015). Progress and challenges in probing the human brain. Nature, 526, 371–379. https://doi.org/10.1038/nature15692.
Qin, K., Zhu, P., Chen, X., et al. (2026). Transdiagnostic mapping of brain structural abnormalities and network-constrained gray matter atrophy patterns across affective and psychotic disorders. Molecular Psychiatry. https://doi.org/10.1038/s41380-026-03701-5.
U.S. Food and Drug Administration. (2026). Bioanalytical method validation for biomarkers: Final guidance. Official resource.
U.S. Food and Drug Administration. (Current guidance). About biomarkers and qualification. Official resource.