01 PERSONAL SITE
/
← RESEARCHER

Interscalar medicine / long-form research note / January 2025

The Future of Multiscale Medicine

Digital medicine has become good at recording fragments: a laboratory value, an image, a prescription, a step count, a visit, a claim. The harder task is to represent the relations among those fragments. A person does not recover at one scale. Molecular signalling changes tissue; tissue changes organ capacity; organ capacity constrains movement; movement interacts with pain, motivation and the environment; access to care changes what is actually done; and all of these processes unfold at different speeds.

Multiscale medicine is the attempt to make some of these cross-scale relations explicit, computable and testable. Its goal is not to collect every possible datum. Its goal is to build models that can answer bounded questions such as: Which mechanism is most likely limiting this patient’s recovery? What may happen if the dose, exercise load or timing of follow-up changes? Which observation would reduce uncertainty enough to change a decision?

Four terms that should not be confused

A multiscale model links two or more spatial, temporal or organizational levels—for example, ion-channel kinetics, myocardial conduction and whole-heart rhythm. The Physiome tradition describes computational physiology across scales, while systems-biology literature includes continuous, discrete and hybrid approaches for coupling levels of organization (Hunter, 2004, DOI 10.1016/j.pbiomolbio.2004.02.006, PMID 15142761; Southern et al., 2008, DOI 10.1016/j.pbiomolbio.2007.07.019, PMID 17888502).

A multiphysics model couples different physical processes, such as blood flow, tissue deformation and electrical conduction. A model may be multiphysics without being personalized, and personalized without spanning many scales.

A patient-specific model adjusts anatomy, parameters or boundary conditions using data from one person. A static three-dimensional reconstruction is useful, but it is not automatically a digital twin.

A digital twin, in the stricter definition adopted by the US National Academies, is a virtual representation that is dynamically updated with data from its physical counterpart, has predictive capability, informs decisions and includes bidirectional interaction (NASEM, 2024, DOI 10.17226/26894). A dashboard, one-off simulation or risk score should therefore not be called a twin merely because it is personalized.

The boundary is becoming more precise, but it is not completely settled. ISO/IEC 30173:2023 supplies cross-sector concepts and terminology for digital twins, not a clinical certification test. A 2026 review of circulatory-system modelling distinguishes a digital model, a data-fed digital shadow and a twin with dynamic two-way coupling; it also identifies continuous coupling across scales as an unresolved engineering problem (Wu et al., 2026, DOI 10.1038/s44222-026-00427-5, PMID 42293032). The taxonomy is useful for intellectual honesty: increasing anatomical detail does not compensate for stale data or the absence of feedback.

This article uses interscalar medicine as a proposed extension, not as an established scientific classification. It connects biological scales with function, behaviour, care delivery and economics. This extension has a legitimate conceptual precedent in the WHO’s International Classification of Functioning, Disability and Health, which treats body functions, activities, participation and environmental factors as interacting components. But a useful interscalar model must specify how those components are measured and causally connected; placing them in one database is not enough.

From a record to an executable hypothesis

An electronic record mainly describes what was observed and documented. An interscalar model should additionally express a hypothesis about how states change and how an intervention may alter that change. A rehabilitation example makes the distinction concrete:

inflammatory state → muscle force and fatigue → balance and gait → confidence and activity → attendance and training dose → functional outcome → need for assistance and cost

This chain is not a universal law. It is a map of candidate mechanisms. Different links require different models and data. Biomarkers may inform inflammation; dynamometry may estimate force; wearables may describe activity; patient-reported outcomes may capture pain and confidence; scheduling data may reveal that the prescribed dose was never delivered. The model is valuable only if it can distinguish among competing explanations and expose its uncertainty.

That requirement matters because prediction is not the same as explanation. A model may accurately predict missed rehabilitation sessions from prior attendance without showing what intervention would prevent the next absence. Counterfactual questions—“what would happen if transport were provided?”—require causal assumptions, suitable comparison data or a designed experiment. Mechanistic equations do not become true merely because they are interpretable, and machine-learning correlations do not become causal merely because they are accurate.

A plausible platform architecture

A clinically useful platform would be modular rather than a single total model of the person. A 2025 Health Policy paper offers a useful external check by naming five indispensable components: the patient, a data connection, a patient-in-silico, an interface and twin synchronization (Sadée et al., 2025, DOI 10.1016/j.landig.2025.02.004, PMID 40518342). The layers below divide the computational workflow differently, but none of those five components may disappear: a model without a usable interface cannot support a decision, and a simulation without synchronization is not a live twin.

  1. Event and provenance layer. Measurements, symptoms, actions, decisions and outcomes are stored with time, source, units, quality flags and consent status. HL7 FHIR R5 is a standard for health-data exchange; its Provenance resource records who or what created or changed a resource and supports assessments of authenticity and trust. The newer ISO/TS 6201:2025 frames personalized digital health around multi-level interoperability, dynamic consent and knowledge sharing. It is a 15-page technical specification, not proof that two implementations share clinical meaning or that either is a validated twin.

  2. State-estimation layer. Sparse and noisy observations are converted into estimates of latent states such as hydration, fatigue, disease activity or functional capacity. Missingness must itself be modelled: absence of wearable data can mean device failure, non-use, hospitalization or recovery—not zero activity.

  3. Model layer. Mechanistic models encode conserved quantities and known physiology; statistical models learn patterns that are difficult to specify; causal models state assumptions about interventions; and human knowledge sets constraints. Hybrid models can combine these strengths, but every component adds failure modes. Standards such as SBML Level 3 (DOI 10.15252/msb.20199110, PMID 32845085) and CellML 2.0 (DOI 10.1515/jib-2020-0021, PMID 32759406) make mathematical models more exchangeable and reusable; they do not by themselves make a model clinically valid.

  4. Scenario layer. The system compares explicit alternatives—continue, intensify, delay, substitute or stop—and reports a distribution of possible outcomes, not a single authoritative number. It should also state which variables drive the result and where the model is extrapolating beyond its validation domain.

  5. Decision and learning layer. A clinician and patient decide whether the information is actionable. Subsequent observations are used to recalibrate the model, detect drift and test whether the earlier prediction was useful. The platform should preserve the original prediction so that it cannot be silently rewritten after the outcome is known.

Across all five layers, identity, time and version are safety properties. Every output should be bound to the correct person, the exact model and parameter version, its calibration window, source devices, last synchronization time and intended use. A clinically plausible answer generated from another patient’s data, a retired sensor or an unapproved model revision is still a wrong answer. Generated or imputed observations must remain distinguishable from measured history; otherwise the system can validate itself against data it created.

What has already worked

The strongest evidence concerns bounded models with a defined context of use—not universal whole-person twins.

Drug development and dosing. Physiologically based pharmacokinetic and pharmacodynamic models connect drug properties with organ volumes, blood flows, enzymes and patient characteristics. Model-informed drug development is already part of regulatory practice. The final ICH M15 guideline, adopted in January 2026, requires a question of interest, context of use, assessment of model influence and consequences of a wrong decision, model evaluation and documented evidence. FDA published its final implementation guidance in June 2026 (FDA M15 page). This is an important precedent: credibility is proportional to the decision, not to the visual sophistication of the simulation.

Diabetes technology. The UVA/Padova type 1 diabetes simulator created virtual populations for preclinical testing of closed-loop insulin algorithms. FDA accepted the 2008 version as a substitute for certain animal experiments; an updated simulator was checked against clinical glucose traces (Visentin et al., 2014, DOI 10.1089/dia.2013.0377, PMID 24571584). This did not eliminate human trials. It narrowed and accelerated a particular development step.

Cardiac electrophysiology. Personalized heart models have been used to estimate arrhythmia risk after myocardial infarction (Arevalo et al., 2016, DOI 10.1038/ncomms11437, PMID 27164184) and to identify potential ablation targets. The latter study included retrospective animal and human cohorts and a prospective feasibility series of five patients (Prakosa et al., 2018, DOI 10.1038/s41551-018-0282-2, PMID 30847259). These are compelling translational examples, but the small prospective sample and specialized workflow do not justify broad claims of proven outcome benefit. The field’s own roadmap emphasizes the synergy of mechanistic and statistical models and the unresolved problem of clinical validation (Corral-Acero et al., 2020, DOI 10.1093/eurheartj/ehaa159, PMID 32128588).

The wider evidence base remains early. A systematic review published in 2026, with searches through May 2025, covered 26 studies across several clinical domains but reported that real-world integration was still uncommon and that validation, infrastructure detail, interoperability and equity remained recurring gaps (Calcaterra et al., 2026, DOI 10.1177/20552076261415934, PMID 42370013). This finding tempers the fast-growing publication count: a prototype, retrospective cohort or framework paper is not evidence of durable clinical benefit.

These cases support a restrained conclusion: in silico models can reduce experiments, refine trials and improve selected decisions when the question, inputs and validation target are narrow. They do not show that a continuously updated, whole-body, behavioural and economic twin is ready for routine care.

Practical scenarios for interscalar medicine

Adaptive rehabilitation. A bounded model could combine diagnosis, tissue-healing constraints, pain, strength, gait, home activity, sleep, adherence and access barriers. Its immediate role would be to identify why the delivered rehabilitation dose differs from the prescribed dose and to compare safe scheduling or load options. The first useful output may be a request for a missing measurement or a transport intervention—not an automated exercise prescription.

Complex treatment planning. Cardiac, oncological or surgical teams could compare a small set of interventions using patient-specific anatomy and physiology, with explicit uncertainty and multidisciplinary review. The model should be advisory and should identify conditions under which its ranking of options changes.

Dose individualization and interaction risk. PBPK and quantitative systems pharmacology can link molecular mechanisms, organ function, co-medications and population variability. The practical benefit is not a perfect biochemical replica but a justified estimate for a defined population or patient state.

Continuity of care. An event model can reveal that apparent biological deterioration followed a medication gap, delayed authorization, equipment failure or missed follow-up. This is where the interscalar extension may add more value than a more detailed organ model: it can connect physiology to the care actually received.

In silico trials and service design. Virtual cohorts may explore sensitivity to eligibility criteria, adherence, measurement schedules and workflow constraints before a prospective trial. They can prioritize experiments, not replace empirical evidence by default. Operational or economic objectives must remain separate from the patient’s clinical objective; otherwise a system may quietly optimize bed turnover or short-term cost at the expense of function and equity.

Counterarguments and limits

More scales can make a model less identifiable. Many combinations of unobserved parameters can reproduce the same observations. This “equifinality” creates plausible narratives with different treatment implications. Adding variables can increase uncertainty faster than information.

The patient is not the data stream. Sensors drift, records are copied, codes reflect billing, and people selectively report symptoms. Frequent measurements can still omit meaning, values, informal care and changing goals.

A model can be right for the wrong reason. Retrospective accuracy may exploit site-specific workflow or treatment patterns. When the model changes care, the data-generating process changes too. External validation, prospective evaluation and drift monitoring are therefore necessary.

Mechanism and causality are not interchangeable. A mechanistic model contains causal structure only to the extent that its equations and boundary conditions are correct for the intervention. A data-driven model needs explicit causal design before it can answer “what if” questions.

Real time has a cost. Continuous calibration increases computation, infrastructure, cybersecurity exposure and clinical workload. A slower model used at a decision point may be safer and more valuable than a permanently synchronized twin.

Human factors can erase technical gains. Unclear uncertainty displays, alert fatigue and automation bias can degrade decisions. Early-stage evaluation should examine the performance of the human–AI team, workflow and failure recovery, not only model discrimination; DECIDE-AI provides a relevant reporting framework (DOI 10.1038/s41591-022-01772-9, PMID 35585198).

Privacy is not the only ethical issue. Continuous multimodal models can enable surveillance, insurance segmentation, employment discrimination or behavioural pressure. They can also shift authority from patients and clinicians to vendors that control the model. Ethical analyses have warned that digital twins may either reduce or amplify inequality (Bruynseels et al., 2018, DOI 10.3389/fgene.2018.00031, PMID 29487613). Consent must be revisable, secondary uses visible, access logged, and a patient must be able to contest or refuse model-based recommendations without losing ordinary care.

Keeping data local is not the same as making computation private. Federated learning can train across institutions without pooling raw records, but shared model updates can leak information and limited trust between participants creates additional attack surfaces. Primary attack research has reconstructed high-resolution inputs from parameter gradients, including gradients from trained networks (Geiping et al., 2020). A healthcare-focused review maps the broader threats and mitigations such as secure aggregation, differential privacy and trusted execution (Pati et al., 2024, DOI 10.1016/j.patter.2024.100974, PMID 39081567). Each mitigation has assumptions and costs: differential privacy trades information against a measurable privacy budget, encryption does not fix biased local data, and federation does not harmonize outcomes or consent. The relevant question is therefore not “centralized or federated?” but which actors can learn what, from which messages, under which failure model.

Public acceptance is therefore conditional rather than automatic. A 2025 consensus project combining expert work, a patient-consumer focus group and a nationally representative Swiss survey found interest in using medical twins among 62% of respondents, while 87% opposed mandatory use and 75% assigned responsibility for infrastructure to the state (Iqbal et al., 2025, DOI 10.1038/s41746-025-01897-4, PMID 40721854). The percentages should not be generalized beyond that setting, but the tension is instructive: curiosity about personal benefit can coexist with strong resistance to compulsion and private control.

Cost saving is a hypothesis. Building, integrating, validating and maintaining models is expensive. An economic claim must compare alternatives, time horizons and consequences transparently. CHEERS 2022 is a reporting standard for such evaluations (DOI 10.1136/bmj-2021-067975, PMID 35017145); it is not a substitute for a well-designed comparative study.

What evidence should be required?

A credible clinical model should pass a staged test.

  1. Define the question of interest, intended user, decision, population, setting and unacceptable failure.
  2. Verify that equations and code are solved correctly; document versions, units, provenance and reproducibility.
  3. Validate relevant outputs against independent data at the scales used by the decision.
  4. Quantify parameter, measurement, structural and population uncertainty; test sensitivity and out-of-domain behaviour.
  5. Run a prospective “silent” study in the target workflow before recommendations affect care.
  6. Evaluate usability, equity, failure recovery and clinician–patient interaction.
  7. Conduct an impact study against current care using patient-relevant outcomes and harms, not only technical accuracy.
  8. Monitor calibration, drift, incidents, subgroup performance and model changes after deployment.
  9. Predefine which updates are permitted, how each revision will be verified and validated, when human authorization is required and how rollback will work.

This logic is consistent with ASME V&V 40-2018, the FDA’s final 2023 guidance on credibility of computational models in medical-device submissions, and ISO 14971:2019 for medical-device risk management. Verification asks whether the model was solved correctly; validation asks whether it is sufficiently faithful for a stated use; uncertainty quantification asks how much confidence a prediction deserves. None is a one-time certificate for all future uses. A review of in silico regulatory evidence makes the same point: credibility belongs to a specific context of use (Viceconti et al., 2021, DOI 10.1016/j.ymeth.2020.01.011, PMID 31991193).

For components that meet the definition of AI-enabled medical-device software, lifecycle governance is now more explicit. IMDRF/AIML WG/N88 FINAL:2025 sets international good-machine-learning-practice principles, while the FDA’s current final guidance on a Predetermined Change Control Plan (August 2025; docket FDA-2022-D-2628) describes planned modifications, their development and validation methods, and impact assessment. These documents do not regulate every research simulation, nor do they authorize unconstrained self-learning. They support a narrower conclusion: an adaptive clinical twin must have an auditable update envelope rather than a claim of permanent validation.

Reporting must also match the claim. If a twin is evaluated as an AI diagnostic test, the corrected STARD-AI guideline asks authors to report dataset practices, the index test, its evaluation, bias, fairness, applicability and generalizability (main article DOI 10.1038/s41591-025-03953-8, PMID 40954311; 2026 author correction DOI 10.1038/s41591-026-04570-9, PMID 42443516). DECIDE-AI is more relevant to early live clinical use; ASME V&V 40 is relevant to model credibility; CHEERS addresses economic evaluation. None is a universal “digital-twin checklist,” and complete reporting does not itself demonstrate low bias or improved outcomes.

Patents: a map of commercial direction, not clinical proof

Patent filings show that organizations are trying to protect implementations that combine anatomy, sensors, behaviour and prediction. Examples include Philips’ “Digital twin of a person” (US 12,367,985 B2, granted 2025), a patient model selected and updated from imaging and wearable data; Southwest Research Institute’s “Digital Twin For Predicting Performance Outcomes” (US 11,721,437 B2, granted 2023), combining neuromusculoskeletal, imaging, metabolic and wearable inputs; and Twin Health’s “Precision treatment platform enabled by whole body digital twin technology” (US 11,957,484 B2, granted 2024), focused on metabolic state and personalized recommendations.

Two pending applications expose less visible implementation problems. Philips’ “Verifying a digital twin of a subject” (US 2024/0290502 A1, published 2024) compares simulated and actual responses to test whether a twin is associated with the stated subject. IBM’s “Synthetic digital twin for a patient” (US 2025/0006367 A1, published 2025) describes generating disease-specific time series from episodic measurements. The first highlights identity binding; the second highlights the attraction—and provenance risk—of filling temporal gaps with synthetic data. Both were pending applications when checked, not granted patents or evidence of novelty or clinical effectiveness.

A rehabilitation-specific grant makes the gap between implementability and efficacy especially visible. Innovision’s “Digital twin” (US 12,046,374 B2, granted 2024) claims a platform linking prescribed scoliosis exercises, supervised baseline capture, a patient-facing avatar, home performance capture and twin updates. This is a concrete workflow close to the scenario above, but the grant reports a protected implementation, not comparative clinical outcomes, adherence benefit or safe autonomy.

These records are relevant to freedom-to-operate and market analysis. They do not establish that the claimed system improves health outcomes, that every claim is valid in every jurisdiction, or that “whole body” is a scientifically validated scope. Patent status should be rechecked before publication or legal use.

A realistic development path

The near future is likely to consist of bounded twins: a heart for ablation planning, a glucose–insulin system for controller testing, a pharmacology model for dosing, or a rehabilitation model for one functional goal. The next step is composable models with explicit interfaces, shared semantics and versioned provenance. Only then does it become reasonable to connect biology with behaviour, service delivery and cost.

The decisive advance will not be a larger database or a more human-looking avatar. It will be an auditable loop:

observation → state estimate → competing explanations → intervention scenarios → shared decision → outcome → model revision

The promise of interscalar medicine is to explain recovery without reducing the patient to one organ or one score. Its discipline must be the opposite of technological totalism: model only what the decision requires, reveal what is not known, validate where harm can occur, and preserve the patient’s right to choose. The future platform should not replace clinical judgment. It should make the assumptions behind judgment visible, testable and corrigible.

References / Список литературы

The register below is a standalone bibliography for both language versions. Article titles, publication records, persistent identifiers and official-document status were checked against PubMed/PMC, publishers, standards organizations, regulators and patent databases through 2026-07-29.

Foundational and peer-reviewed publications

  1. Hunter PJ. The IUPS Physiome Project: a framework for computational physiology. Prog Biophys Mol Biol. 2004;85(2-3):551-569. PubMed. DOI: 10.1016/j.pbiomolbio.2004.02.006. PMID: 15142761.
  2. Southern J et al. Multi-scale computational modelling in biology and physiology. Prog Biophys Mol Biol. 2008;96(1-3):60-89. PubMed. DOI: 10.1016/j.pbiomolbio.2007.07.019. PMID: 17888502.
  3. Winslow RL et al. Computational Medicine: Translating Models to Clinical Care. Sci Transl Med. 2012;4(158):158rv11. Open full text. DOI: 10.1126/scitranslmed.3003528. PMID: 23115356.
  4. Björnsson B et al. Digital twins to personalize medicine. Genome Med. 2020;12:4 (published 2019-12-31). PubMed. DOI: 10.1186/s13073-019-0701-3. PMID: 31892363.
  5. Corral-Acero J et al. The “Digital Twin” to enable the vision of precision cardiology. Eur Heart J. 2020;41(48):4556-4564. PubMed. DOI: 10.1093/eurheartj/ehaa159. PMID: 32128588.
  6. Arevalo HJ et al. Arrhythmia risk stratification of patients after myocardial infarction using personalized heart models. Nat Commun. 2016;7:11437. PubMed. DOI: 10.1038/ncomms11437. PMID: 27164184.
  7. Prakosa A et al. Personalized virtual-heart technology for guiding the ablation of infarct-related ventricular tachycardia. Nat Biomed Eng. 2018;2(10):732-740. PubMed. DOI: 10.1038/s41551-018-0282-2. PMID: 30847259.
  8. Visentin R et al. The University of Virginia/Padova Type 1 Diabetes Simulator Matches the Glucose Traces of a Clinical Trial. Diabetes Technol Ther. 2014;16(7):428-434. PubMed. DOI: 10.1089/dia.2013.0377. PMID: 24571584.
  9. Viceconti M et al. In silico trials: Verification, validation and uncertainty quantification of predictive models used in the regulatory evaluation of biomedical products. Methods. 2021;185:120-127. PubMed. DOI: 10.1016/j.ymeth.2020.01.011. PMID: 31991193.
  10. Bruynseels K et al. Digital Twins in Health Care: Ethical Implications of an Emerging Engineering Paradigm. Front Genet. 2018;9:31. PubMed. DOI: 10.3389/fgene.2018.00031. PMID: 29487613.
  11. Vasey B et al. Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. Nat Med. 2022;28:924-933. PubMed. DOI: 10.1038/s41591-022-01772-9. PMID: 35585198.
  12. Husereau D et al. Consolidated Health Economic Evaluation Reporting Standards 2022 (CHEERS 2022). BMJ. 2022;376:e067975. PubMed. DOI: 10.1136/bmj-2021-067975. PMID: 35017145.
  13. Clerx M et al. CellML 2.0. J Integr Bioinform. 2020;17(2-3):20200021. PubMed. DOI: 10.1515/jib-2020-0021. PMID: 32759406.
  14. Keating SM et al. SBML Level 3: an extensible format for the exchange and reuse of biological models. Mol Syst Biol. 2020;16:e9110. PubMed. DOI: 10.15252/msb.20199110. PMID: 32845085.
  15. Burr C et al. Realising the digital twin: a thematic review and analysis of the ethical, legal, and social issues for digital twins in healthcare. AI & Society. 2026;41:5243-5267. PubMed. DOI: 10.1007/s00146-025-02833-6. PMID: 42254856. (Useful recent thematic review; not used as the sole support for any clinical claim.)
  16. Iqbal JD et al. A consensus statement on the use of digital twins in medicine. npj Digit Med. 2025;8(1):484. PubMed. DOI: 10.1038/s41746-025-01897-4. PMID: 40721854.
  17. Calcaterra V et al. Digital twins in healthcare: A systematic review of current applications, frameworks, and future directions. Digit Health. 2026;12:20552076261415934. PubMed. DOI: 10.1177/20552076261415934. PMID: 42370013.
  18. Wu R et al. Digital twins and digital models of the human circulatory system. Nat Rev Bioeng. Published online 2026-04-30. PubMed. DOI: 10.1038/s44222-026-00427-5. PMID: 42293032.
  19. Sadée C et al. Medical digital twins: enabling precision medicine and medical artificial intelligence. Lancet Digit Health. 2025;7(7):100864. PubMed. DOI: 10.1016/j.landig.2025.02.004. PMID: 40518342.
  20. Pati S et al. Privacy preservation for federated learning in health care. Patterns (N Y). 2024;5(7):100974. PubMed. DOI: 10.1016/j.patter.2024.100974. PMID: 39081567.
  21. Sounderajah V et al.; STARD-AI Steering Committee. The STARD-AI reporting guideline for diagnostic accuracy studies using artificial intelligence. Nat Med. 2025;31(10):3283-3289. PubMed. DOI: 10.1038/s41591-025-03953-8. PMID: 40954311. Author correction, 2026: DOI 10.1038/s41591-026-04570-9; PMID 42443516.
  22. Geiping J, Bauermeister H, Dröge H, Moeller M. Inverting Gradients—How easy is it to break privacy in federated learning? Adv Neural Inf Process Syst. 2020;33:16937-16947. Official NeurIPS proceedings. No DOI or PMID assigned in the official proceedings record.

Official reports, standards and regulatory guidance

  1. National Academies of Sciences, Engineering, and Medicine. Foundational Research Gaps and Future Directions for Digital Twins. 2024. Official report. DOI: 10.17226/26894.
  2. WHO. International Classification of Functioning, Disability and Health (ICF). Official classification page.
  3. HL7. FHIR Release 5, version 5.0.0. Official specification and Provenance resource.
  4. ICH. M15: General Principles for Model-Informed Drug Development. Final version adopted 2026-01-29. Official PDF. FDA final-guidance page, June 2026, docket FDA-2024-D-5580.
  5. ASME. V&V 40-2018: Assessing Credibility of Computational Modeling through Verification and Validation—Application to Medical Devices. Official standard page.
  6. FDA. Assessing the Credibility of Computational Modeling and Simulation in Medical Device Submissions. Final guidance, November 2023. Official guidance, docket FDA-2021-D-0980.
  7. ISO. ISO 14971:2019 Medical devices—Application of risk management to medical devices. Confirmed current in 2025. Official standard page.
  8. ISO/IEC. ISO/IEC 30173:2023 Digital twin—Concepts and terminology. Published November 2023. Official standard page.
  9. IMDRF. Good machine learning practice for medical device development: Guiding principles. IMDRF/AIML WG/N88 FINAL:2025, published 2025-01-29. Official document page and official PDF.
  10. FDA. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions. Final guidance, August 2025. Official guidance, docket FDA-2022-D-2628.
  11. ISO. ISO/TS 6201:2025 Health informatics—Personalized digital health framework. Published February 2025. Official technical-specification page.

Patent documents

  1. Koninklijke Philips N.V. Digital twin of a person. US 12,367,985 B2. Priority 2018-12-19; granted 2025-07-22. Google Patents listed it active when checked 2026-07-28.
  2. Southwest Research Institute. Digital Twin For Predicting Performance Outcomes. US 11,721,437 B2 (published application: US 2019/0371466 A1). Priority 2018-06-04; granted 2023-08-08. Google Patents listed it active when checked 2026-07-28.
  3. Twin Health, Inc. Precision treatment platform enabled by whole body digital twin technology. US 11,957,484 B2. Priority 2019-08-13; granted 2024-04-16. Google Patents listed it active when checked 2026-07-28.
  4. Koninklijke Philips N.V. Verifying a digital twin of a subject. US 2024/0290502 A1. Priority 2021-06-24; published 2024-08-29. Google Patents listed the application as pending when checked 2026-07-29.
  5. International Business Machines Corporation. Synthetic digital twin for a patient. US 2025/0006367 A1. Filed 2023-06-30; published 2025-01-02. Google Patents listed the application as pending when checked 2026-07-29.
  6. Innovision LLC. Digital twin. US 12,046,374 B2 (published application: US 2022/0375621 A1). Priority 2021-05-23; granted 2024-07-23. Google Patents listed it active when checked 2026-07-29.

Patent databases explicitly warn that aggregated legal-status fields are not legal opinions. For freedom-to-operate work, verify status and family members in USPTO Patent Center, WIPO PATENTSCOPE and the relevant national registers.