Skip to content

Biomarkers and digital phenotyping

TL;DR — Bipolar disorder has many group-level biological correlates but, as of a live literature re-search on 2026-08-30, no blood, imaging, electrophysiologic, genomic or digital marker has independent validation and demonstrated utility as a stand-alone diagnostic test or treatment selector. The strongest reproducible architecture is polygenic: a 2019 GWAS of 20,352 cases and 31,358 controls, with follow-up in 9,412 cases and 137,760 controls, identified 30 genome-wide significant loci, each with small effects (Stahl 2019, PMID 31043756). Portable-device studies are proliferating, but a 2026 synthesis of seven reviews covering 111 studies and 19,945 participants concluded that none of the technologies was robust enough to replace clinical outcomes (Astill Wright 2026, PMID 42228842). Smartphone mood forecasting and multimodal classification can work within research cohorts, yet performance falls for the clinically important bipolar-versus-unipolar distinction: one 12-week study reported AUC 0.62 (Langholm 2023, PMID 38129571). The near-term use case is longitudinal measurement and relapse-warning support, not diagnosis by app.

What counts as a biomarker?

A useful clinical biomarker needs a defined context of use, reproducible measurement, external validation, calibration in the intended population, and evidence that acting on it improves decisions. Statistical separation between groups is insufficient. Reviews of staging, blood, imaging and digital candidates consistently identify heterogeneity, cross-sectional designs, medication confounding and absent external validation as the gap between association and clinical utility (Roda 2015, PMID 25860561; Hu 2023, PMID 36347076; Saccaro 2021, PMID 34488086).

Candidate domains

Domain Representative evidence What it can support now Principal limitation
Common-variant genetics 30 genome-wide significant loci in a large case-control GWAS (Stahl 2019, PMID 31043756) Etiologic architecture and cross-disorder biology Individual effects are small; ancestry representation and treatment utility remain limited
Inflammation / neurotrophins Systematic reviews catalog state and treatment associations (Ruiz-Sastre 2024, PMID 38879067) Mechanistic and stratification hypotheses Assay, state, treatment and comorbidity confounding
Oxidative stress Malondialdehyde meta-analysis (Capuzzi 2022, PMID 34740710) Group-level research signal Nonspecific to bipolar disorder
S100B Meta-analysis in mania (da Rosa 2016, PMID 27475892) State-related hypothesis Small, heterogeneous clinical samples
Metabolomics Systematic review across major depression and bipolar disorder (MacDonald 2019, PMID 30411484) Multivariate discovery Platforms and pipelines differ; replication sparse
MicroRNA Bipolar “miRNome” review (Fries 2018, PMID 28969861) Regulatory-mechanism hypotheses Tissue and medication effects; no clinical cut-point
Gut microbiota Systematic review of gut-brain-axis abnormalities (Obi-Azuike 2023, PMID 37127945) Hypothesis generation Diet, geography, medication and sequencing-method confounding
Structural MRI 21-study cortical-thickness review; 11-study meta-analysis, 649 cases/818 controls (Zhu 2022, PMID 34971699) Group-level anatomy Entire meta-analyzed evidence was cross-sectional
P300 Systematic review and meta-analysis (Wada 2019, PMID 31200163) Electrophysiologic research phenotype Paradigm and clinical-state heterogeneity
Retina OCT cytoarchitecture meta-analysis across schizophrenia and bipolar disorder (Lizano 2020, PMID 31112601); ERG systematic review of 32 mood-disorder studies (n=1,334; 5 meta-analyzed) (de Deus 2026, PMID 42669383) Accessible CNS-correlate hypothesis OCT analyses combine diagnoses; ERG quantitative pooling was MDD-only, and bipolar findings remain qualitative/state-dependent
Actigraphy 13 studies, 821 subjects in bipolar-specific meta-analysis (De Crescenzo 2017, PMID 28185811) Objective sleep/activity description Devices, epochs and mood states differ
Smartphone sensing 62-study portable-technology review (Saccaro 2021, PMID 34488086) Dense longitudinal measurement Small samples, overfitting, nonstandard performance reporting
Geolocation 18 publications/16 studies; 11 bipolar studies (Fraccaro 2019, PMID 31260049) Mobility/routine research features Privacy burden; 7 studies had <10 patients and 11 lasted <12 weeks
Speech / audiovisual 11 of 21 differential-diagnosis studies in a 2025 review (Zhong 2025, PMID 40408762) Multimodal classifier research Feature instability, language and recording-context dependence

Genomics: biology without a diagnostic test

The 2019 Psychiatric Genomics Consortium analysis used 20,352 bipolar cases and 31,358 controls of European descent, then followed 822 variants in another 9,412 cases and 137,760 controls. Thirty loci were genome-wide significant, including 20 newly identified loci; enriched biology included ion channels, synaptic components, insulin-secretion regulation and endocannabinoid signaling (Stahl 2019, PMID 31043756). Genetic correlations differed by subtype: bipolar I correlated more strongly with schizophrenia, driven by psychosis, whereas bipolar II correlated more strongly with major depression (Stahl 2019, PMID 31043756).

These are etiologic findings, not a clinical assay. Polygenic scores distribute risk continuously, depend on ancestry-matched discovery data, and do not establish episode state or treatment response. A diagnostic table should therefore separate “genome-wide significant” from “clinically discriminative.”

Circulating and molecular candidates

Reviews have grouped candidate signals into neurotrophic, inflammatory, oxidative-stress, endocrine/metabolic and epigenetic classes (Roda 2015, PMID 25860561; Hu 2023, PMID 36347076). The problem is specificity: infection, obesity, smoking, sleep loss, episode polarity and medication can change the same pathways. A treatment-responsive biomarker may reflect exposure rather than disease biology (Ruiz-Sastre 2024, PMID 38879067).

Candidate Evidence type Interpretation ceiling
BDNF Systematic review/meta-analysis of cognition across the schizophrenia-bipolar spectrum (Dombi 2022, PMID 35686181) Possible cross-diagnostic cognition correlate, not bipolar diagnosis
Folate Serum-level systematic review/meta-analysis (Hsieh 2019, PMID 31640634) Nutritional/metabolic association; causality unresolved
Leptin Bipolar systematic review/meta-analysis (Fernandes 2016, PMID 27065008) Metabolic-state signal with adiposity confounding
S100B Mania-focused systematic review/meta-analysis (da Rosa 2016, PMID 27475892) Candidate state marker; not validated for decisions
Malondialdehyde Meta-analysis (Capuzzi 2022, PMID 34740710) Oxidative-stress association, biologically nonspecific
Metabolite panels Systematic review (MacDonald 2019, PMID 30411484) Multivariate discovery requiring locked external validation
MicroRNAs Literature synthesis (Fries 2018, PMID 28969861) Regulatory candidates; tissue origin and normalization matter

Neuroimaging and electrophysiology

The cortical-thickness meta-analysis included 649 bipolar participants and 818 healthy controls and found thinning in frontoinsular, frontal and anterior-cingulate regions. However, all contributing studies were cross-sectional, and medication and mood-state effects could not be tested adequately (Zhu 2022, PMID 34971699). Findings therefore characterize group distributions, not a scan-based diagnosis.

P300 and retinal studies widen the measurement space but face the same translation test: incremental prediction beyond clinical history, reproducibility across sites, and a prospective decision consequence (Wada 2019, PMID 31200163; Lizano 2020, PMID 31112601). A 2026 electroretinography review included 32 mood-disorder studies (n=1,334), of which only five entered the quantitative synthesis, and reported generally reduced amplitudes and prolonged implicit times; the only significant pooled parameter was a small MDD mixed-scotopic b-wave implicit-time delay (g=0.23, 95% CI 0.00–0.46, its lower bound at the null), while bipolar and seasonal-affective findings were described as mood-state-dependent rather than meta-analyzed as a diagnostic test (de Deus 2026, PMID 42669383). Adolescent imaging comparisons also underline diagnostic overlap rather than a clean anatomical boundary (Long 2023, PMID 36669567).

Actigraphy

The bipolar-specific actigraphy review found 13 quality-eligible studies comprising 821 subjects, with lower mean activity and altered sleep patterns across manic, depressed and euthymic phases (De Crescenzo 2017, PMID 28185811). A broader mood-disorder meta-analysis included 38 studies and 3,758 participants. In euthymic/remitted patients versus controls, total sleep time was longer (SMD -0.33, 95% CI -0.55 to -0.11), sleep latency longer (SMD -0.22, 95% CI -0.42 to -0.02), and wake after sleep onset longer (SMD -0.22, 95% CI -0.39 to -0.04); device and severity heterogeneity constrained interpretation (Tazawa 2019, PMID 31060012).

Actigraphy is therefore useful as an objective longitudinal layer, especially when anchored to a person's own baseline. It does not by itself determine whether reduced activity reflects bipolar depression, medication sedation, medical illness or behavior.

Smartphone and wearable phenotyping

Evidence map

The 2021 portable-technology review included 62 studies: 27 smartphone studies, 15 wearable-sensor studies, 17 audiovisual studies and 3 multimodal studies. The combined samples were 2,325 bipolar participants and 724 healthy controls; two-thirds used artificial intelligence. Reviewers found fair-to-excellent reported classification but warned about small samples, heterogeneity, overfitting and nonstandard metrics (Saccaro 2021, PMID 34488086).

The geolocation review found that 12 of 14 publications assessing clinical concepts reported an association, especially with mood. Yet 7 studies had fewer than 10 participants and 11 lasted under 12 weeks, making apparent accuracy fragile (Fraccaro 2019, PMID 31260049). Later reviews reached the same boundary: electronic monitoring is promising for episode prediction, but methods and validation remain insufficient for clinical deployment (Antosik-Wójcińska 2020, PMID 32305023; Ortiz 2021, PMID 34706433; Maatoug 2022, PMID 35958638).

Forecasting rather than diagnosis

Busk and colleagues modeled 15,975 daily smartphone self-assessments from 84 participants. A hierarchical Bayesian model using four days of history predicted next-day mood with R² 0.51 and RMSE 0.32 on a -3 to +3 scale; error increased as the horizon extended toward seven days (Busk 2020, PMID 32234702). This is a within-cohort forecasting result, not proof that acting on alerts prevents episodes.

Differential diagnosis

In a 12-week feasibility study, passive geolocation, accelerometer and screen-state data plus ecological momentary assessment separated control from mood-disorder participants with AUC 0.91, but bipolar I/II from major depression with AUC 0.62 (Langholm 2023, PMID 38129571). A 2025 systematic review found 21 differential-diagnosis studies: 11 directly compared unipolar and bipolar depression; 6 used smartphone apps, 3 wearables, 11 audiovisual data and 1 multimodal technology (Zhong 2025, PMID 40408762).

Activity timing and speech features are candidate signals, but the 2025 review did not establish a single stable feature set. Performance depends on diagnosis quality, validation design, missingness and whether train/test data from the same participant leak across folds (Zhong 2025, PMID 40408762).

A minimum validation ladder

Step Required evidence Common failure
1. Analytical validity Stable measurement across devices, laboratories or software versions Sensor and assay drift
2. Internal clinical validity Locked model and participant-level cross-validation Repeated observations leak between train and test sets
3. External validity New sites, languages, devices, ancestry groups and care settings Single-site convenience sample
4. Prospective calibration Sensitivity, specificity, positive predictive value and alert burden in real time Retrospective AUC only
5. Clinical utility Randomized evidence that biomarker-guided action improves outcomes Prediction without an effective response pathway
6. Equity and governance Consent, data minimization, security, access and bias assessment Surveillance burden shifted to patients

A 2026 meta-analysis identified eight monitoring RCTs with 1,230 participants. In the bipolar subgroup (n=873), monitoring did not significantly improve manic or depressive symptoms, and the review found no evidence of benefit for relapse, readmission or quality of life (Astill Wright 2026, PMID 41499681). SmartBipolar then randomized 201 people with progressed bipolar disorder to monitoring, monitoring plus clinical feedback, or treatment as usual and found no six-month difference in mood instability or secondary outcomes (Faurholt-Jepsen 2026, PMID 41865316).

That is not evidence that all feedback interventions are inert. A 2026 multicenter double-blind RCT randomized 93 people with major depressive disorder or bipolar disorder to a circadian-rhythm feedback app or sham; among 80 in the modified intention-to-treat analysis, recurrent episodes were more frequent with sham (incidence-rate ratio 3.39, 95% CI 1.86–6.17) (Yeom 2026, PMID 42337416). Because diagnoses were pooled, the result supports a targeted circadian intervention but does not establish bipolar-specific efficacy or validate passive diagnosis. Reviews also emphasize privacy, data security and legal concerns (Faurholt-Jepsen 2018, PMID 29510813; Tatham 2022, PMID 35319470). Patient preference and burden therefore belong in the accuracy assessment, not after it.

Reporting checklist for future studies

Item Minimum report
Cohort Diagnostic method, episode state, medications, recruitment setting and attrition
Signal Device/assay version, sampling frequency, missingness and preprocessing
Model Prespecified outcome, feature-selection boundary and locked hyperparameters
Validation Participant-separated folds and an external cohort
Performance Sensitivity, specificity, calibration, confidence intervals and decision threshold—not accuracy alone
Governance Consent, retention, access, secondary use, withdrawal and response to high-risk alerts

Without these elements, seemingly high classification performance cannot be compared across studies, a central problem identified by the portable-technology review (Saccaro 2021, PMID 34488086).

State markers are not diagnostic classifiers

Among 51 studies (4,547 bipolar participants), only 18 entered meta-analysis; circulating BDNF correlated weakly with depression severity (SMD −0.22, 95% CI −0.38 to −0.05), while inflammatory findings were inconsistent (Vega-Núñez 2022, PMID 35740389). A separate 53-study synthesis (2,467 cases; 2,360 controls) found that no single peripheral marker distinguished mood phase, although combinations involving hsCRP/IL-6, BDNF/TNF-α and soluble TNF receptor 1 generated candidate signatures (Rowland 2018, PMID 30113291). Post-hoc multi-marker combinations still require locked thresholds and prospective external validation.

Large-scale imaging complicates simple inherited-marker models. ENIGMA combined 11 samples: 408 unaffected first-degree bipolar relatives, 542 schizophrenia relatives, 841 controls, 255 bipolar cases and 464 schizophrenia cases. Bipolar relatives showed higher fractional anisotropy only in the posterior limb of the internal capsule, while affected groups showed lower values across several tracts; the authors inferred that patient white-matter deficits were not straightforward familial-risk signatures (Barendse 2026, PMID 42168580).

Polygenic scores measure liability, not diagnosis. In twin data, each SD higher combined psychosis PRS was associated with case/concordance odds of 2.12 (95% CI 1.23–3.87) in monozygotic twins and 2.74 (1.56–5.30) in dizygotic twins; twin heritability was 0.73 (0.30–1.00) (Song 2024, PMID 39196586). Wide intervals, combined schizophrenia/bipolar phenotype and population dependence preclude individual clinical interpretation.

Digital relapse models: internal performance versus transport

A 2026 review included 52 passive-sensing studies (4,814 participants), 27% in bipolar disorder. Reported AUCs ranged 0.70–0.88 for relapse one to four weeks ahead, but 75% of studies were high risk of bias and most used internal validation; sleep and activity appeared in 83% of models, GPS mobility in 75% and communication frequency in 65% (Fang 2026, PMID 42106724). The unresolved test is a prospective closed-loop trial in which alerts change care and improve outcomes without unacceptable false alarms.

Open questions

  • Can passive data predict a prospectively adjudicated manic or depressive episode early enough to change its course, with acceptable false-alert rates (Ortiz 2021, PMID 34706433)?
  • Does combining within-person change with population-level models outperform either alone?
  • Can biomarker panels distinguish bipolar from unipolar depression beyond structured clinical history in independent health systems (Zhong 2025, PMID 40408762)?
  • Which signals remain robust across phone operating systems, languages, work patterns and socioeconomic groups?
  • Can intervention trials show that biomarker-triggered care reduces hospitalization without increasing coercive surveillance?
  • How much apparent molecular “staging” is explained by cumulative medication, smoking, adiposity and episode burden (Gama 2013, PMID 23567604; Roda 2015, PMID 25860561)?

References

  1. Stahl EA, et al. Genome-wide association study identifies 30 loci associated with bipolar disorder. Nat Genet. 2019;51:793-803. PMID 31043756
  2. Roda Â, et al. Biomarkers and staging of bipolar disorder: a systematic review. Trends Psychiatry Psychother. 2015;37:3-11. PMID 25860561
  3. Hu X, et al. Biomarkers and detection methods of bipolar disorder. Biosens Bioelectron. 2023;220:114842. PMID 36347076
  4. Ruiz-Sastre P, et al. Pharmacotherapeutic value of inflammatory and neurotrophic biomarkers in bipolar disorder: A systematic review. Prog Neuropsychopharmacol Biol Psychiatry. 2024;134:111056. PMID 38879067
  5. Dombi ZB, et al. Brain Derived Neurotrophic Factor and Cognitive Dysfunction in the Schizophrenia-Bipolar Spectrum: A Systematic Review and Meta-Analysis. Front Psychiatry. 2022;13:827322. PMID 35686181
  6. Capuzzi E, et al. Malondialdehyde as a candidate biomarker for bipolar disorder: A meta-analysis. Prog Neuropsychopharmacol Biol Psychiatry. 2022;113:110469. PMID 34740710
  7. da Rosa MI, et al. Serum S100B in manic bipolar disorder patients: Systematic review and meta-analysis. J Affect Disord. 2016;206:210-215. PMID 27475892
  8. Hsieh YC, et al. Serum folate levels in bipolar disorder: a systematic review and meta-analysis. BMC Psychiatry. 2019;19:305. PMID 31640634
  9. Fernandes BS, et al. Leptin in bipolar disorder: A systematic review and meta-analysis. Eur Psychiatry. 2016;35:1-7. PMID 27065008
  10. MacDonald K, et al. Biomarkers for major depressive and bipolar disorders using metabolomics: A systematic review. Am J Med Genet B. 2019;180:122-137. PMID 30411484
  11. Fries GR, et al. The miRNome of bipolar disorder. J Affect Disord. 2018;233:110-116. PMID 28969861
  12. Obi-Azuike C, et al. A systematic review on gut-brain axis aberrations in bipolar disorder. Brain Behav. 2023;13:e3037. PMID 37127945
  13. Zhu Z, et al. Cortical thickness abnormalities in patients with bipolar disorder: A systematic review and meta-analysis. J Affect Disord. 2022;300:209-218. PMID 34971699
  14. Wada M, et al. The P300 event-related potential in bipolar disorder: A systematic review and meta-analysis. J Affect Disord. 2019;256:234-249. PMID 31200163
  15. Lizano P, et al. A Meta-analysis of Retinal Cytoarchitectural Abnormalities in Schizophrenia and Bipolar Disorder. Schizophr Bull. 2020;46:43-53. PMID 31112601
  16. Long X, et al. Gray matter alterations in adolescent major depressive disorder and adolescent bipolar disorder. J Affect Disord. 2023;325:550-563. PMID 36669567
  17. Gama CS, et al. Staging and neuroprogression in bipolar disorder: a systematic review. Rev Bras Psiquiatr. 2013;35:70-74. PMID 23567604
  18. De Crescenzo F, et al. Actigraphic features of bipolar disorder: A systematic review and meta-analysis. Sleep Med Rev. 2017;33:58-69. PMID 28185811
  19. Tazawa Y, et al. Actigraphy for evaluation of mood disorders: A systematic review and meta-analysis. J Affect Disord. 2019;253:257-269. PMID 31060012
  20. Saccaro LF, et al. Portable technologies for digital phenotyping of bipolar disorder: A systematic review. J Affect Disord. 2021;295:323-338. PMID 34488086
  21. Fraccaro P, et al. Digital biomarkers from geolocation data in bipolar disorder and schizophrenia: a systematic review. J Am Med Inform Assoc. 2019;26:1412-1420. PMID 31260049
  22. Antosik-Wójcińska AZ, et al. Smartphone as a monitoring tool for bipolar disorder: a systematic review. Int J Med Inform. 2020;138:104131. PMID 32305023
  23. Ortiz A, et al. Apps and gaps in bipolar disorder: electronic monitoring for episode prediction. J Affect Disord. 2021;295:1190-1200. PMID 34706433
  24. Maatoug R, et al. Digital phenotype of mood disorders: A conceptual and critical review. Front Psychiatry. 2022;13:895860. PMID 35958638
  25. Zhong R, et al. Using Digital Phenotyping to Discriminate Unipolar Depression and Bipolar Disorder: Systematic Review. J Med Internet Res. 2025;27:e72229. PMID 40408762
  26. Busk J, et al. Forecasting Mood in Bipolar Disorder From Smartphone Self-assessments. JMIR Mhealth Uhealth. 2020;8:e15028. PMID 32234702
  27. Langholm C, et al. Classifying and clustering mood disorder patients using smartphone data. NPJ Digit Med. 2023;6:238. PMID 38129571
  28. Faurholt-Jepsen M. Electronic monitoring in bipolar disorder. Dan Med J. 2018;65:B5460. PMID 29510813
  29. Tatham I, et al. Process and Outcome Evaluations of Smartphone Apps for Bipolar Disorder: Scoping Review. J Med Internet Res. 2022;24:e29114. PMID 35319470
  30. Astill Wright L, et al. Mood Monitoring, Mood Tracking, and Ambulatory Assessment Interventions in Depression and Bipolar Disorder: Systematic Review and Meta-Analysis of Randomized Controlled Trials. JMIR Ment Health. 2026;13:e84020. PMID 41499681
  31. Faurholt-Jepsen M, et al. The effect of smartphone-based monitoring and treatment including clinical feedback in progressed bipolar disorder: the SmartBipolar trial randomised controlled parallel-group trial. Int J Bipolar Disord. 2026;14:17. PMID 41865316
  32. Yeom JW, et al. Circadian Rhythm Stabilization App to Prevent Mood Episode Recurrence in Patients With Mood Disorders: A Multicenter, Double-Blind, Sham-Controlled, Randomized Clinical Trial. Am J Psychiatry. 2026. PMID 42337416
  33. Astill Wright L, et al. Recommendations for Research and Clinical Implementation of Ambulatory Assessment, Mood Monitoring, Digital Phenotyping, and Remote Measurement Technology in Mood Disorders. JMIR Ment Health. 2026;13:e79501. PMID 42228842
  34. de Deus M, et al. Retinal electrophysiology in mood disorders: Toward clinically relevant biomarkers. A systematic review and meta-analysis. J Affect Disord. 2026. PMID 42669383
  35. Vega-Núñez A, et al. Clinical value of inflammatory and neurotrophic biomarkers in bipolar disorder: a systematic review and meta-analysis. Biomedicines. 2022;10:1368. PMID 35740389
  36. Rowland T, et al. Neurotrophins, cytokines, oxidative stress mediators and mood state in bipolar disorder: systematic review and meta-analyses. Br J Psychiatry. 2018;213:514–525. PMID 30113291
  37. Barendse MEA, et al. White matter microstructure in relatives of people with schizophrenia or bipolar disorder: an ENIGMA meta-analysis. Mol Psychiatry. 2026;31:5423–5430. PMID 42168580
  38. Song J, et al. Polygenic risk scores and twin concordance for schizophrenia and bipolar disorder. JAMA Psychiatry. 2024;81:1246–1252. PMID 39196586
  39. Fang SS, et al. Digital phenotyping for predicting relapse in psychiatric disorders: a systematic review of passive sensing approaches. BMC Psychiatry. 2026;26:507. PMID 42106724