Human Digital Twins: A Comprehensive Overview

• Human Digital Twins (HDTs) are digital representations that continuously synchronize with human data to mirror physiology, behavior, and health states.
• They integrate heterogeneous data from sensors, EHRs, and social inputs using multi-scale modeling and advanced computation for real-time simulation.
• HDTs are applied in healthcare, industrial safety, and human-agent interaction, emphasizing ethical governance, data security, and personalized support.

Human Digital Twins (HDTs) are digital representations of humans, or of specific human aspects, maintained through ongoing synchronization with the physical person and used to generate feedback, predictions, and decision support. In healthcare-oriented formulations, an HDT “mirrors the physical entity of a human body through a dynamic virtual model that continuously reflects changes in molecular, physiological, emotional, and lifestyle factors,” while broader reviews treat the represented target as the whole body, an organ, a flow process, cells, or human behavior (Mokhtari, 15 Mar 2025, Pascual et al., 2023). Cross-domain surveys therefore describe HDTs not as static records or avatars, but as bi-directionally linked, data-driven, feedback-producing systems whose scope depends on the application and level of abstraction (Lauer-Schmaltz et al., 2024).

1. Definition and conceptual scope

The defining feature of an HDT is the coupling of representation, synchronization, and feedback. A cross-domain survey defines an HDT as “a digital representation of a human (or specific aspect of a human); which uses a bi-directional data thread for on-going synchronization with the human’s state; and provision of feedback based on data aggregation and predictions to directly (or indirectly) influence the human entity, user(s) and/or their environment” (Lauer-Schmaltz et al., 2024). In personalized healthcare, this idea is sharpened into the notion of a living digital mirror that is continuously updated from heterogeneous data feeds and used to support monitoring, diagnosis, prescription, surgery, rehabilitation, and preventive care (Mokhtari, 15 Mar 2025).

The concept is intentionally broader than full-body replication. Reviews of implementations note that HDTs may represent “a complete human body, some parts of it -as can be organs, flows, cells-, or even human behaviors,” and that most practical systems remain partial rather than whole-body because full-body twinning is resource-intensive and technically difficult (Pascual et al., 2023). This multi-level scope has led to taxonomies that distinguish human-in-the-loop, physiological, mechanistic, and cognitive HDTs, while allowing composite systems that combine several of these categories in one application (Lauer-Schmaltz et al., 2024).

A holistic specification literature further treats HDTs as cumulative functional systems rather than single-purpose models. In that framing, functionality progresses through six levels: store, analyze, personalize, predict, control, and optimize, with higher levels including lower ones (Mandischer et al., 20 Jul 2025). This suggests that an HDT can range from a secure, individualized data lake to a runtime control or optimization layer for cyber-physical systems, depending on how much modeling, inference, and intervention capability is required.

2. Relation to conventional digital twins

HDTs inherit core ideas from the broader digital twin lineage but alter them substantially by changing the physical referent from an engineered artifact to a person. Surveys trace the digital twin idea from NASA’s Apollo-era twin logic and Grieves’ product lifecycle management formulation toward increasingly human-centered applications, culminating in explicit HDT research after 2018 (Lin et al., 2022). Conventional DTs and HDTs remain similar in that both are virtual representations of physical entities, both rely on model-centric design, both depend on continuous data exchange, and both use analytics and simulation to support decisions (Mokhtari, 15 Mar 2025).

The decisive difference is that the HDT must represent a living, mobile, psychologically and socially situated subject. Reviews repeatedly emphasize that HDTs must account for emotions, mental states, genetic variation, subjective responses, ethical concerns, and social dynamics; that they typically cannot rely on embedded sensors in the way industrial assets can; and that trust, privacy, ownership, and regulation become central rather than peripheral (Mokhtari, 15 Mar 2025, Lauer-Schmaltz et al., 2024). In this sense, HDTs do not merely extend DTs to a new domain; they change the sensing problem, the modeling problem, and the governance problem.

This distinction also clarifies a common misconception. Not every human-related digital artifact is an HDT. The survey literature differentiates a digital model, which is not connected to the physical entity; a digital shadow, which has one-way data flow from physical to digital; and a digital twin, which is bi-directionally linked and capable of influencing the physical entity or its environment (Lauer-Schmaltz et al., 2024). By that criterion, many systems loosely described as HDTs are more accurately digital shadows or narrowly scoped patient models.

Another recurrent misconception is that an HDT must already be a faithful whole-body replica. Systematic reviews of implementations show the opposite: most current systems twin one organ, one physiological subsystem, one behavioral function, or one work-related role rather than the entire person (Pascual et al., 2023). The field is therefore better understood as modular and hierarchical, with whole-human twinning as a long-term ambition rather than a current baseline.

3. Architectures and enabling infrastructure

Healthcare-oriented HDT literature often converges on a common operational stack. One overview organizes HDT operation into five interdependent components—data acquisition, digital modeling and virtualization, communication and computation, data management, and data analysis and decision-making—and maps them to a five-layer pipeline of data acquisition, communication, data management, computational processing, and data analysis/decision-making, with feedback to the physical twin (Mokhtari, 15 Mar 2025). A closely related networking survey presents the same logic as an end-to-end architecture for personalized healthcare (Chen et al., 2023).

At the acquisition layer, HDTs aggregate heterogeneous signals from electronic health records, diagnostic images, biometric test results, wearable sensors, implantable devices, and “soft sensors” derived from social media content. The goal is a holistic representation of the person that includes clinical, physiological, psychological, and behavioral cues. The sensing substrate spans head-mounted devices such as smart glasses and contact lenses, torso-worn sensors such as smart garments and patches, limb-worn accessories such as armbands and shoes, implantable sensors for internal measurements, and EHR-based historical grounding (Mokhtari, 15 Mar 2025).

Communication is typically split into on-body and beyond-body tiers. On-body communication is commonly realized through wireless body area networks with star or hierarchical topologies and protocols such as Bluetooth Low Energy, ZigBee, NB-IoT, 6LoWPAN, and molecular communication for in-body scenarios. Beyond-body communication links local hubs to remote servers or clouds hosting the virtual twin and must support ultra-reliable, low-latency transport of multimodal data, including text, images, video, haptics, and 3D content (Chen et al., 2023). In this context, the literature highlights the Tactile Internet for real-time haptic exchange and semantic communication for transmitting meaning-bearing content rather than raw data (Mokhtari, 15 Mar 2025).

Latency and reliability requirements are correspondingly severe. One healthcare overview explicitly defines xURLLC requirements as “speeds of at least 100. Gbps, almost perfect reliability (99.99999%), and delays of 1 millisecond or less,” particularly for tele-surgery and other time-critical interactions (Mokhtari, 15 Mar 2025). Immersive communication work on edge-computing-empowered Tactile Internet extends this point by showing that HDT interaction requires pervasive connectivity, real-time feedback, high-fidelity virtual modeling, ultra-high reliability, and synchronous multimodal interaction; its physical-therapy testbed reports that edge-computing-enhanced deployment improves delay, jitter, and subjective synchronization relative to a TI-only baseline (Xiang et al., 2023).

Computation is increasingly distributed across edge and cloud. Because many HDT tasks are too latency-sensitive for cloud-only processing, the dominant pattern is Multi-Access Edge Computing with edge-cloud collaboration, so that time-critical tasks remain near the user while heavier simulations are offloaded to the cloud (Mokhtari, 15 Mar 2025). Data management, meanwhile, includes noise and outlier cleaning, missing-value imputation, feature selection or extraction, multimodal fusion, storage in systems such as HDFS, HBase, or OpenStack Swift, and security mechanisms including intrusion detection, cryptography, anonymization, federated learning, differential privacy, and blockchain-style immutable logging (Chen et al., 2023). Generative AI surveys add that GANs, VAEs, Transformers, diffusion models, normalizing flows, and score-based generative models can assist every stage of this pipeline by synthesizing data, filling gaps, denoising signals, reconstructing missing modalities, and generating useful representations under scarcity, bias, noise, or privacy constraints (Chen et al., 2024).

4. Modeling scales and computational paradigms

HDT modeling is intrinsically multi-scale. Systematic reviews describe implementations at the level of the whole body, organs, physiological flows, cells, and behavior, with organ-level systems currently dominating because they are technically feasible while still clinically useful (Pascual et al., 2023). A more specialized survey of organ digital twins formalizes this landscape as a two-stage process of anatomical twinning and functional twinning: the first builds patient-specific geometry from imaging, and the second simulates physiology on that geometry using calibrated or inferred patient-specific parameters (Lyu et al., 16 Jan 2026).

This organ-centered literature further classifies anatomy into structural-static, motion-dominant, and fluid-conduit paradigms. Structural-static organs prioritize geometric fidelity; motion-dominant organs require 4D spatiotemporal reconstruction; fluid-conduit organs require topological correctness and connectivity of branching trees because flow simulation is sensitive to missing or disconnected vessels (Lyu et al., 16 Jan 2026). Functional twinning is then organized around three dominant technical paradigms: electrophysiology and signal propagation, fluid or hemodynamic simulation, and solid biomechanics, often coupled into multi-physics systems. Representative formulations include the incompressible Navier–Stokes equation

$$\rho\left(\frac{\partial \mathbf{v}}{\partial t} + \mathbf{v}\cdot \nabla \mathbf{v}\right) = -\nabla p + \mu \nabla^2 \mathbf{v} + \mathbf{f},$$

reaction-diffusion dynamics

$$\frac{\partial \mathbf{u}}{\partial t} = \nabla \cdot (D\nabla \mathbf{u}) + R(\mathbf{u}, t),$$

and physics-informed learning objectives of the form

$$\mathcal{L} = \mathcal{L}_{data} + \lambda \mathcal{L}_{physics}.$$

These equations are not universal HDT laws, but they illustrate how patient-specific twinning often embeds established PDE families directly into the computational twin (Lyu et al., 16 Jan 2026).

Alongside mechanistic modeling, the field has developed explicitly data-driven and hybrid paradigms. Digital Twin Generators model a patient’s future clinical trajectory as a patient-specific, longitudinal, multivariate, probabilistic simulator conditioned on baseline and prior observations; one general-purpose Neural Boltzmann Machine architecture was trained across 13 indications by changing the training set and tuning hyperparameters rather than redesigning the architecture for each disease (Alam et al., 2024). More recent health-twin frameworks also reject monolithic modeling in favor of modular composition. OmniBioTwin, for example, treats the health twin as a System-of-Twinned-Systems with seven coordinated layers—Data, Twin, Coupling, Synchronization, Decision, Interaction, and Audit—and formalizes each twin as an autonomous subsystem with its own state, update law, and uncertainty representation (Wang et al., 9 Jun 2026).

A further variant is the bio-inspired HDT. One framework for the human-centric Metaverse replaces heterogeneous wearable stacks with Brain-Computer Interface sensing, uses Spiking Neural Networks for event-driven neuromorphic learning, and applies Federated Learning so that raw brain data remain local. In a case study on the EEG Motor Movement/Imagery Dataset, the proposed FL4BCI-SNN achieved 71.05% identification accuracy at 0.5 μJ, compared with 54.92% and 3.13 μJ for FL4BCI-CNN and 76.62% and 5.08 μJ for FL4BCI-LSTM, yielding a Weighted System Performance improvement of 2.12× and 2.11× over the two baselines (Shang et al., 2024). This indicates that HDT modeling is diversifying not only across biological scales but also across sensing and computation paradigms.

5. Application domains and domain-specific reinterpretations

Healthcare remains the most developed HDT domain. Overviews consistently emphasize remote monitoring, diagnosis, prescription and treatment planning, surgery, rehabilitation, preventive care, vaccine development, clinical trials, and hospital operations as primary use cases. The practical logic is to continuously observe the patient, detect subtle changes before overt symptoms emerge, simulate interventions on the virtual twin, and personalize therapy while reducing adverse effects (Mokhtari, 15 Mar 2025). Reviews of implementations report that the strongest concentration of real deployments is in healthcare, with 15 relevant implementations identified in the surveyed literature, 9 in healthcare and 6 in manufacturing (Pascual et al., 2023).

Cardiology is one of the fastest-growing subfields. Reviews in this area describe heart twins as patient-specific computational representations that combine anatomy, physiology, imaging, electrophysiology, hemodynamics, and treatment response, with automatic segmentation and three-dimensional reconstruction from MRI or CT as key bottlenecks. They also stress the role of XR—VR, AR, and MR—as the human-facing layer for visualizing and operationalizing these twins in diagnostics, interventional planning, rehabilitation, education, and remote care (Rudnicka et al., 2024).

Industrial HDTs use the same twin logic but center on workers rather than patients. Job-shop studies define HDTs as real-time digital representations of workers’ physical and cognitive states for safety, productivity, process adaptability, and task allocation. In a Fuzzy AHP-TOPSIS prioritization framework, the dominant decision criteria were Technological Maturity $= 0.352$ and Safety Impact $= 0.343$, and the ranked use-cases were Posture Monitoring $0.639$, Fatigue Prediction $0.628$, PPE Compliance Tracking $0.466$, Health-Based Task Assignment $0.398$, and Skill Training Simulation $0.379$ (Sardar et al., 10 Nov 2025). This literature treats HDT adoption as both a technical and organizational problem, especially in semi-digital environments.

Another branch of HDT research targets cognition, conversation, and human-agent interaction rather than physiology alone. In cybersecurity, Cybonto proposes HDTs as interactive, self-updating cognitive agents capable of modeling beliefs, goals, perceptions, evaluations, and other psychological constructs; its ontology contains 108 constructs and thousands of cognitive-related paths based on 20 psychology theories, and network analysis identifies Behavior, Arousal, Goals, Perception, Self-efficacy, Circumstances, Evaluating, Behavior-Controllability, Knowledge, and Intentional Modality as the top 10 constructs (Nguyen, 2021). In Human-Autonomy Teaming, an HDT architecture integrates speech recognition, visual input capture, context construction, ChatGPT-4o, emotion modeling, lip-syncing, and multimodal rendering so that the HDT can act as a visually and behaviorally realistic teammate across training, deployment, and after-action review; in a gun-assembly task, the prototype achieved 100% success in component identification, instructional guidance, recommendations, and emotional support, while total interaction pipeline time was $$\frac{\partial \mathbf{u}}{\partial t} = \nabla \cdot (D\nabla \mathbf{u}) + R(\mathbf{u}, t),$$0 seconds (Mohammed et al., 4 Apr 2025).

Conversational HDTs extend this trend from task support to personal identity modeling. One “Digital Me” architecture combines GPT-4o with context-aware memory retrieval, neural plasticity-inspired consolidation, and adaptive learning so that the system can mirror an individual’s conversational style, memories, opinions, and interlocutor-dependent behavior. Its retrieval score combines normalized recency, importance, and relevance, with creation-time and access-time decay terms and a GPT-4o-derived importance score on a 0–10 scale (Coll et al., 30 Jun 2025). Related HAT-trust work treats HDTs as LLM-powered computational surrogates for studying trust development in human-AI teams and argues that valid HDT trust models must capture propensity to trust, emergent trust patterns, and time-varying calibration rather than only static self-report scores (Nguyen et al., 2024).

6. Challenges, ethics, and research frontiers

The technical barriers to HDTs are widely acknowledged. Personalized-healthcare surveys identify data quality as foundational, since bad or incomplete data make the twin inaccurate; communication must satisfy ultra-reliable, low-latency constraints; storage and computing are major concerns because HDTs may generate gigabytes of data per day; AI interpretability limits clinician trust and regulatory approval; mobility complicates synchronization; and interoperability across bodily subsystems remains difficult because human physiology is deeply interconnected (Mokhtari, 15 Mar 2025). Cross-domain surveys add that HDT design must confront fidelity and accuracy, sensitivity and robustness, interconnectivity, adaptability, identifiability and traceability, credibility and explainability, embodiment, simplicity and intuitiveness, data security and privacy, safe interaction, and equally distributed access (Lauer-Schmaltz et al., 2024).

Ethical and legal issues are not secondary. The literature warns that HDTs aggregate highly sensitive, often lifelong personal data and therefore raise risks of misuse by insurers, cyberattacks, unauthorized surveillance, bias from unrepresentative datasets, discrimination through genetic profiling, and widening healthcare disparities if access is expensive or poorly insured (Mokhtari, 15 Mar 2025). Regulatory reviews argue that HDTs blur the boundaries among medical device, simulation engine, decision-support system, and data platform, thereby placing unusual pressure on consent, auditability, provenance, traceability, version history, fairness, privacy, and accountability (Pan et al., 18 Aug 2025). A recurrent position in the healthcare literature is that HDTs should be support tools rather than replacements for human judgment (Mokhtari, 15 Mar 2025).

The future research agenda is correspondingly broad. Priorities repeatedly cited include overcoming data scarcity through federated sharing, standardization, synthetic data generation, and data minimization; supporting mobility through adaptive migration and handover protocols; improving semantic interoperability across physiological subsystems; designing user-centric interfaces; integrating lightweight and scalable blockchain mechanisms; strengthening explainable AI and human-in-the-loop decision support; developing more generalized and transferable AI; and creating green, energy-efficient HDTs (Mokhtari, 15 Mar 2025). At the architectural level, recent work points toward interconnected multi-organ and multi-scale systems rather than isolated monoliths. OmniBioTwin explicitly frames future health twins as modular assemblies coupled through explicit interaction operators, with human-in-the-loop decision support and audit layers to support reproducibility, explainability, and eventual regulatory review (Wang et al., 9 Jun 2026).

A plausible implication is that HDT research is converging on two simultaneous demands: richer representation of human complexity and stricter governance of how that representation is built, updated, and used. The field’s central tension is therefore not whether a human can be digitized in some narrow sense, but how far synchronized modeling, simulation, and intervention can be extended without losing fidelity, interpretability, privacy, or human agency.