01 PERSONAL SITE
/
← RESEARCHER

Interscalar medicine / long-form research note / January 2025

Digital Patient Twin in Rehabilitation: A Model of Recovery, Not a Digital Clone

A digital patient twin is not simply a richer electronic health record, a three-dimensional avatar, or an algorithm that predicts an outcome. In the strict sense, it is an individualized computational model that is repeatedly updated from observations of a real person and can simulate clinically relevant future states under alternative actions. The model’s predictions return to care as decision support; subsequent outcomes are then used to test and update the model.

This distinction matters. A record describes what has been documented. A dashboard visualizes it. A prediction model estimates an outcome. A digital twin adds a synchronized model, explicit uncertainty, and a feedback loop between observation, simulation, decision, action, and reassessment. The general terminology is formalized in ISO/IEC 30173:2023, while a healthcare scoping review proposes three indispensable components: a physical entity, a virtual representation, and a connection between them (DOI 10.1038/s41746-024-01073-0; PMID 38519626).

The name is nevertheless used inconsistently. For example, patent publication US20250006367A1 calls cohort-conditioned synthetic time-series generated by a generative adversarial network a “synthetic digital twin,” although such a construct need not remain synchronized with one physical patient. Patent language is not a scientific definition, but it reveals a practical problem: papers and products should disclose the physical counterpart, update cadence, simulated question, feedback authority and retirement rule instead of relying on the label alone.

Why rehabilitation is a demanding test case

Rehabilitation does not optimize one biomarker. It attempts to improve functioning in a changing person living in a particular environment. Two patients with a similar lesion may follow different trajectories because of pain, fatigue, cognition, motor learning, medication, sleep, fear of falling, family support, housing, access to transport, therapy intensity, and the timing of care.

That is why a useful rehabilitation twin must be interscalar. It should connect at least six layers:

  1. Biology and pathology: lesion, inflammation, healing, comorbidities, medication and physiological reserve.
  2. Body functions and structures: strength, range of motion, tone, balance, cardiorespiratory response, pain and cognition.
  3. Activity and participation: walking, dressing, communicating, working and participating in family or community life.
  4. Behavior and learning: adherence, confidence, fatigue management, preferences, motor practice and response to feedback.
  5. Physical and social environment: home layout, assistive products, caregivers, accessibility, work demands and social support.
  6. Care delivery and economics: appointments, therapist capacity, waiting time, transport, device availability, reimbursement and the patient’s time.

This structure is compatible with the biopsychosocial logic of the WHO’s International Classification of Functioning, Disability and Health (ICF), which treats functioning as contextual and includes environmental factors. The twin should therefore optimize a patient-valued functional goal—not merely a sensor metric. Step count, joint angle, or movement symmetry can be useful intermediate measurements, but the target may be “walk safely to the local shop,” “use the affected hand to prepare a meal,” or “return to work without intolerable fatigue.”

The operational loop

A clinically meaningful twin can be described as a sequence:

goal → observations and events → current-state estimate → alternative-action simulations → uncertainty-aware recommendation → shared decision → intervention → measured outcome → model update

The state estimate may combine mechanistic models (for example, anatomy, tissue loading, muscle dynamics or cardiovascular response) with statistical models learned from cohorts. Precision cardiology has described these as complementary deductive and inductive pillars rather than competing approaches (DOI 10.1093/eurheartj/ehaa159). The same principle applies to rehabilitation: a biomechanical model may estimate what load an exercise produces, while a statistical model may estimate the probability that the person will tolerate and complete it.

Every output should include its time horizon, intended use, confidence or credible interval, important missing inputs, population on which the model was validated, and conditions under which the recommendation should not be used. “The patient will recover” is not an acceptable output. “Given the current measurements and stated assumptions, option A is estimated to have a higher probability of achieving the agreed walking goal by week 8, but uncertainty is high because home activity and pain data are incomplete” is closer to a clinically responsible one.

Practical scenarios

1. Adaptive motor rehabilitation after stroke

Wearable inertial sensors, force measurements, motion capture, therapist assessments and patient-reported fatigue can update a model of gait or upper-limb performance. The twin could compare therapy doses or device settings, detect an unexpected plateau, and recommend a reassessment rather than automatically increasing intensity.

This is plausible but not mature. A 2025 scoping review found 16 empirical digital-twin studies in stroke rehabilitation. One retrospective study included 1,216 patients; the other 15 studies included only 54 patients in total, with a median sample size of one. Clinical integration, cost-effectiveness and several other desirable properties had low or very low adoption (DOI 10.1080/10749357.2025.2584031; PMID 41200887). The finding supports continued research, not claims of established clinical benefit.

Two broader 2026 reviews reinforce that caution. A healthcare-wide systematic review found that simulation-oriented designs dominate while real-world clinical integration and validation remain uncommon (DOI 10.1177/20552076261415934; PMID 42370013). A review focused on sports and rehabilitation analyzed 32 papers: 20 concerned rehabilitation, but most evaluations involved fewer than 20 participants, only two studies reported user participation during design, and descriptions of AI were often thin (DOI 10.1080/0144929X.2026.2660222). Publication counts therefore measure activity, not therapeutic effect; heterogeneous systems should not be pooled as if they were one intervention.

2. Safer human–robot interaction

A neuromusculoskeletal twin may estimate joint load, muscle contribution and cardiovascular demand while an exoskeleton or rehabilitation robot is operating. Prototype studies have demonstrated real-time virtual–physical interaction and patient-specific gait control, including a self-balancing lower-limb exoskeleton (DOI 10.3233/THC-220087; PMID 35754239) and an ankle rehabilitation robot (DOI 10.1177/09287329251337237; PMID 40370054). These are engineering and feasibility results; they do not yet establish better long-term functional outcomes.

3. Musculoskeletal and postoperative recovery

Imaging-informed anatomy, range of motion, load, pain, patient-reported outcomes, activity and treatment events could be combined to compare rehabilitation pathways after a fracture, joint replacement or tendon injury. A proposed framework for precision neuromusculoskeletal care maps digital-twin concepts to Achilles tendon mechanobiology and neurorehabilitation interfaces, while explicitly calling for accepted measurement and modeling standards (DOI 10.1123/jab.2023-0114; PMID 37567581).

A 2025 knee study used quantitative MRI and machine learning in Osteoarthritis Initiative data to identify interpretable imaging features associated with incident osteoarthritis and knee replacement. Its authors accurately describe the pipeline as a foundation for a future knee-joint twin; longitudinal state updating, treatment-response simulation and prospective clinical utility still require validation (DOI 10.1038/s41746-025-01507-3; PMID 39984725). This is a useful boundary case: a strong patient representation can be an enabling component without yet being a complete twin.

More directly clinical evidence now exists, but it also exposes definitional ambiguity. In a single-site open-label randomized trial of 201 people with knee osteoarthritis considering total knee arthroplasty, an AI-enabled decision aid combining education, preference elicitation and individualized benefit–risk predictions improved decision quality, decision conflict, treatment concordance, later regret and knee-specific health compared with education alone; several other outcomes, including shared-decision-making scores, satisfaction, visit duration and arthroplasty rates, were similar (DOI 10.1016/j.eclinm.2025.103545; PMID 41112505; NCT04805554). This supports the whole decision-support package, not the claim that a continuously synchronized rehabilitation twin caused the benefit: the intervention had multiple inseparable components and used “digital twin” more broadly than the strict definition adopted here.

4. Rehabilitation between visits

A home-based twin could combine exercise performance, symptoms, sleep, medication, activity and contextual events. Its value would not be constant surveillance. It would be selective detection of clinically meaningful divergence: worsening pain after a dose change, declining activity after a transport disruption, or repeated failure of an exercise because the home environment makes it impractical. The proper response may be a call, a new assessment, an environmental modification or a simpler plan—not an automated warning to “improve adherence.”

Evidence from outside rehabilitation illustrates a feasible operating pattern without establishing rehabilitation effectiveness. A 2026 single-center randomized proof-of-concept in type 2 diabetes used an iteratively updated predictive model to generate daily messages, all reviewed by a human interventionist. The ancillary sample was only 19 participants, so the reported weight difference is exploratory, and the authors explicitly distinguish their predictive approximation from a fully realized, mechanistic, bidirectionally coupled twin (DOI 10.1038/s44401-026-00118-8). The transferable lesson is the combination of frequent updating, narrow outputs and human review—not the clinical result itself.

5. Resource-aware care planning

The clinically best intervention is not effective if it is unavailable or impossible for the patient to attend. A pathway twin could simulate not only physiological response but also waiting time, therapist capacity, travel burden and cost. This makes organizational and economic constraints part of the model rather than unexplained noise. However, a scarcity-aware model must not normalize poor access or quietly optimize lower-quality care for disadvantaged groups.

Arguments—and the counterarguments they require

More data should produce a better twin. Not necessarily. More variables can increase missingness, measurement error, spurious correlations and operational burden. Data should be collected because they alter a specified decision, improve calibration, or reveal a safety condition.

Continuous sensors are more objective than clinical assessment. Sensors are consistent only within their validated conditions. Placement, battery state, firmware, skin tone, assistive-device use, atypical movement and changes in the home can alter measurements. A precise signal can still be an invalid proxy for functioning.

An individualized model is inherently more accurate. Personalization can overfit sparse observations. A safe twin must show how much of an estimate comes from the individual, from a reference population, and from assumptions. It must be tested across sites and subgroups.

A closed feedback loop automatically improves care. A feedback loop can amplify a biased measurement or unsafe recommendation. High-risk changes require human authorization, conservative limits, rollback, and an independent way to observe harm.

Accurate prediction identifies the best intervention. Prognosis and treatment effect are different questions. A model that predicts poor recovery may be excellent at identifying risk but unable to say whether more therapy, a different device, or a social intervention will change that risk. Counterfactual recommendations require randomized evidence or defensible causal assumptions; high predictive accuracy alone does not establish clinical utility (DOI 10.1186/s12911-024-02513-3).

One comprehensive whole-person twin is the destination. A modular federation may be safer and more achievable: a validated gait model, a medication model, a scheduling model and a shared event/provenance layer. Each module can then have a defined purpose, evidence boundary, owner and retirement rule.

Data and interoperability

Interoperability is more than moving fields between systems. The receiver must know what was measured, by which device or person, under what protocol, at what time, with what unit and quality, and whether the value was observed, inferred, corrected or simulated.

A practical implementation can reuse existing standards:

  • HL7 FHIR R5 for exchange and resources such as Observation, Goal, CarePlan, Device, Consent, Provenance and AuditEvent;
  • DICOM / ISO 12052 for medical images and related information;
  • ISO/IEEE 11073-20601:2022 for interoperable personal health-device communication;
  • WHO ICF for functioning and environmental context, and WHO ICHI for interventions.

These standards do not create a twin by themselves. They make its inputs, goals, interventions and provenance less ambiguous. The model version, code, parameters, training data lineage, device firmware, overrides and generated recommendation should be traceable. Without provenance, retrospective audit and reproducibility are impossible.

Validation, regulation and governance

Validation should proceed in gates:

  1. verify sensor and data-pipeline accuracy;
  2. verify the software and mathematical implementation;
  3. assess calibration, discrimination and uncertainty;
  4. perform temporal, geographic and subgroup validation;
  5. run in silent mode without influencing care;
  6. evaluate usability, workflow, automation bias and failure recovery;
  7. compare patient outcomes and harms against current care;
  8. evaluate workload, access, cost-effectiveness and opportunity cost;
  9. monitor drift, incidents and inequitable performance after deployment.

When software provides information used for diagnosis or treatment, its intended purpose may bring it within software-as-a-medical-device regulation. The IMDRF clinical-evaluation framework N41 separates valid clinical association, analytical/technical validation and clinical validation. AI components should follow total-product-lifecycle principles such as the 2025 IMDRF Good Machine Learning Practice. Early live evaluation can be reported with DECIDE-AI, predictive models with TRIPOD+AI, and randomized trials with CONSORT-AI.

No single medical-device standard certifies a digital twin as safe. If its intended purpose makes it a medical device, however, established lifecycle controls are directly relevant: ISO 14971:2019 for risk management, IEC 62304:2006 with Amendment 1:2015 for medical-device software lifecycle processes, and IEC 81001-5-1:2021 for secure health-software lifecycle activities. Conformity with a process standard is not evidence of clinical benefit; it is part of the assurance case that must accompany clinical evidence.

Governance must also address consent over time, secondary uses, cybersecurity, access control, patient correction of erroneous data, data portability, model retirement, responsibility for overrides and liability when model and clinician disagree. Because rehabilitation often involves cognitively or physically vulnerable people and long periods of home monitoring, refusal must remain meaningful: declining a twin should not automatically mean losing access to ordinary care.

A 2026 thematic review of ethical, legal and social issues found a persistent gap between proposed value and empirical validation and mapped adoption barriers to the NASSS framework (DOI 10.1007/s00146-025-02833-6). Its implication is organizational as much as ethical: trustworthy deployment requires named capabilities and responsibilities for patients, clinicians, developers, providers and regulators, not a generic ethics statement appended after development.

The central risk: mistaking the model for the person

The twin will always be incomplete. Pain, motivation, dignity, relationships and the meaning of recovery cannot be reduced without loss. Treating the model as the patient invites automation bias, coercive adherence monitoring and unjust resource allocation. Ethical analysis has warned that digital twins can both broaden access and intensify inequality, segmentation and discrimination (DOI 10.3389/fgene.2018.00031).

The correct design principle is therefore epistemic humility: show what the model knows, how it knows it, what it does not know, and who remains accountable for acting on it.

Conclusion

The most valuable digital patient twin in rehabilitation will not be the most visually realistic or the one with the largest data lake. It will be the smallest model that can answer a defined clinical question, update when the patient changes, quantify uncertainty, expose its evidence and assumptions, and improve a functional outcome that matters to the patient.

The purpose of the twin is not to replace clinical judgment or to manufacture certainty. It is to make recovery trajectories more observable, alternative actions more testable, and decisions more explicit—while keeping the real person, not the simulation, as the source of goals and the final measure of success.


References / Список литературы

All sources below were verified against publisher, bibliographic-database, standards-body, regulator or patent records. They are new relative to the current short article, which has no citations.

Peer-reviewed publications and methodological guidance

  1. Katsoulakis E, et al. “Digital twins for health: a scoping review.” npj Digital Medicine 2024;7:77. DOI 10.1038/s41746-024-01073-0. PMID 38519626.
  2. Corral-Acero J, et al. “The ‘Digital Twin’ to enable the vision of precision cardiology.” European Heart Journal 2020;41:4556–4564. DOI 10.1093/eurheartj/ehaa159.
  3. Niederer SA, Sacks MS, Girolami M, Willcox K. “Scaling digital twins from the artisanal to the industrial.” Nature Computational Science 2021;1:313–320. DOI 10.1038/s43588-021-00072-5.
  4. Laubenbacher R, Mehrad B, Shmulevich I, Trayanova N. “Digital twins in medicine.” Nature Computational Science 2024;4:184–191. DOI 10.1038/s43588-024-00607-6.
  5. García-Rudolph A, et al. “Digital twins in stroke rehabilitation: a scoping review of objectives, data sources, mechanisms, outcomes, and desirable properties.” Topics in Stroke Rehabilitation 2025. DOI 10.1080/10749357.2025.2584031. PMID 41200887.
  6. Saxby DJ, Pizzolato C, Diamond LE. “A Digital Twin Framework for Precision Neuromusculoskeletal Health Care: Extension Upon Industrial Standards.” Journal of Applied Biomechanics 2023;39:347–354. DOI 10.1123/jab.2023-0114. PMID 37567581.
  7. Wang W, et al. “Digital twin rehabilitation system based on self-balancing lower limb exoskeleton.” Technology and Health Care 2023;31:103–115. DOI 10.3233/THC-220087. PMID 35754239.
  8. Xie S, Zhan M, Li Y, Xi F. “The virtual-real interaction system design and interaction characteristics research of an ankle rehabilitation robot based on digital twin.” Technology and Health Care 2025;33:2279–2304. DOI 10.1177/09287329251337237. PMID 40370054.
  9. Bruynseels K, Santoni de Sio F, van den Hoven J. “Digital Twins in Health Care: Ethical Implications of an Emerging Engineering Paradigm.” Frontiers in Genetics 2018;9:31. DOI 10.3389/fgene.2018.00031.
  10. Iqbal JD, et al. “A consensus statement on the use of digital twins in medicine.” npj Digital Medicine 2025;8:484. DOI 10.1038/s41746-025-01897-4.
  11. van Amsterdam WAC, de Jong PA, Verhoeff JJC, Leiner T, Ranganath R. “From algorithms to action: improving patient care requires causality.” BMC Medical Informatics and Decision Making 2024;24:111. DOI 10.1186/s12911-024-02513-3. PMCID PMC11046962.
  12. Vasey B, et al. “Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI.” Nature Medicine 2022;28:924–933. DOI 10.1038/s41591-022-01772-9. PMID 35585198.
  13. Collins GS, et al. “TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods.” BMJ 2024;385:e078378. DOI 10.1136/bmj-2023-078378. PMID 38636956.
  14. Liu X, et al. “Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension.” Nature Medicine 2020;26:1364–1374. DOI 10.1038/s41591-020-1034-x.
  15. Calcaterra V, Guardamagna L, Gatti A, et al. “Digital twins in healthcare: A systematic review of current applications, frameworks, and future directions.” Digital Health 2026;12:20552076261415934. DOI 10.1177/20552076261415934. PMID 42370013; PMCID PMC13305746.
  16. Barricelli BR, Cerutti F, Morzenti S. “Human digital twins in sports and rehabilitation: a systematic review.” Behaviour & Information Technology 2026. DOI 10.1080/0144929X.2026.2660222.
  17. Hoyer G, Gao KT, Gassert FG, et al. “Foundations of a knee joint digital twin from qMRI biomarkers for osteoarthritis and knee replacement.” npj Digital Medicine 2025;8:118. DOI 10.1038/s41746-025-01507-3. PMID 39984725; PMCID PMC11845592.
  18. Jayakumar P, Rathouz PJ, Lin E, et al. “Shared decision making using digital twins in knee osteoarthritis care: a randomized clinical trial of an AI-enabled decision aid versus education alone on decision quality, physical function, and user experience.” EClinicalMedicine 2025;89:103545. DOI 10.1016/j.eclinm.2025.103545. PMID 41112505; PMCID PMC12528923; trial registration NCT04805554.
  19. Wang J, Faruqui SHA, Alaeddini A, et al. “Human-in-the-loop AI predictive digital twin to extend virtual precision diabetes care between visits.” npj Health Systems 2026;3:59. DOI 10.1038/s44401-026-00118-8.
  20. Burr CD, Qian S, Winter P, 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. DOI 10.1007/s00146-025-02833-6.

Official standards and regulatory documents

  1. ISO/IEC. ISO/IEC 30173:2023 — Digital twin — Concepts and terminology.
  2. WHO. International Classification of Functioning, Disability and Health (ICF), endorsed by WHA resolution 54.21.
  3. WHO. International Classification of Health Interventions (ICHI).
  4. HL7. FHIR Release 5 specification, including Observation, Goal, CarePlan and Provenance.
  5. DICOM Standards Committee. DICOM overview; recognized as ISO 12052.
  6. ISO/IEEE. ISO/IEEE 11073-20601:2022 — Personal health device communication.
  7. IMDRF. IMDRF/SaMD WG/N41FINAL:2017 — Software as a Medical Device: Clinical Evaluation.
  8. IMDRF/FDA. Good Machine Learning Practice for Medical Device Development: Guiding Principles, final IMDRF principles published in 2025.
  9. ISO. ISO 14971:2019 — Medical devices — Application of risk management to medical devices; confirmed current in 2025.
  10. IEC. IEC 62304:2006 — Medical device software — Software life cycle processes, with Amendment 1:2015; confirmed current in 2021 and under systematic review at the research date.
  11. IEC. IEC 81001-5-1:2021 — Health software and health IT systems safety, effectiveness and security — Part 5-1: Security — Activities in the product life cycle; published standard scheduled for revision at the research date.

Patent publications: evidence of the design space, not clinical efficacy

  1. US20190005200A1 / WO2019005184A1 — “Methods and systems for generating a patient digital twin”. Combines records, imaging, genetics and history for query, simulation and rule-based recommendations.
  2. WO2020243781A1 — “Biospine: a digital twin neurorehabilitation system”. Describes biosensor-informed control of electrical stimulation and motor assistance with load and cardiovascular constraints.
  3. US20220375621A1 — “Digital twin”. Describes a physical-therapy platform that updates a patient twin from captured exercise performance, with a scoliosis-specific embodiment.
  4. US20230064408A1 — “Digital twin systems, devices, and methods for treatment of the musculoskeletal system”. Covers a musculoskeletal treatment journey, PROMs/PREMs, wearables, treatment scenarios and rehabilitation monitoring.
  5. CN115670870A — “A lower limb rehabilitation exoskeleton system based on digital twin”. Describes individualized gait fitting, wearable sensing, safety limits and adaptive adjustment.
  6. EP3992978A1 — “Knee arthroplasty functional digital twin”. Describes a patient-specific knee model built from imaging and sensor data for simulated motion and preoperative testing of surgical changes; published application, pending status shown in the linked record at the research date.
  7. US20250006367A1 — “Synthetic digital twin for a patient”. Describes GAN-generated disease-specific synthetic time-series conditioned on episodic patient metadata; published application, pending status shown in the linked record at the research date.

Patent note: publication or grant status does not demonstrate novelty in every jurisdiction, freedom to operate, technical feasibility, safety, regulatory clearance, or clinical benefit. Legal status must be checked in the relevant patent office before any product or intellectual-property conclusion is drawn.