PaperlayerView exact revision on arXiv

Source-linked AI summary

Generative AI-Driven Human Digital Twin in IoT-Healthcare: A Comprehensive Survey

Jiayuan Chen, You Shi, Changyan Yi, Hongyang Du, Jiawen Kang, Dusit Niyato

arXiv:2401.13699v2cs.HCcs.AIcs.LG

TL;DR

Integrating GAI-driven HDT into healthcare IT ecosystems requires addressing interoperability, including efficient data exchange and effective data integration. The survey reviews GAI-driven HDT implementation across data acquisition, communication, data management, and digital modeling, alongside related applications. The survey concludes that GAI can assist HDT data acquisition, communication, data management, and digital modeling.

  • Problem

    Integrating GAI-driven HDT into healthcare IT ecosystems requires addressing interoperability, including efficient data exchange and effective data integration.

    Source: D. Seamless Integration of GAI-driven… · p. 21

  • Method

    The survey reviews GAI-driven HDT implementation across data acquisition, communication, data management, and digital modeling, alongside related applications.

    Source: §VI · p. 21

  • Results

    The survey concludes that GAI can assist HDT data acquisition, communication, data management, and digital modeling.

    Source: §VI · p. 21

  • Takeaways & Limitations

    The reviewed applications include personalized health monitoring and diagnosis, prescription, and rehabilitation.

    Source: §VI · p. 21

Abstract

from arXiv · show

The Internet of things (IoT) can significantly enhance the quality of human life, specifically in healthcare, attracting extensive attentions to IoT-healthcare services. Meanwhile, the human digital twin (HDT) is proposed as an innovative paradigm that can comprehensively characterize the replication of the individual human body in the digital world and reflect its physical status in real time. Naturally, HDT is envisioned to empower IoT-healthcare beyond the application of healthcare monitoring by acting as a versatile and vivid human digital testbed, simulating the outcomes and guiding the practical treatments. However, successfully establishing HDT requires high-fidelity virtual modeling and strong information interactions but possibly with scarce, biased and noisy data. Fortunately, a recent popular technology called generative artificial intelligence (GAI) may be a promising solution because it can leverage advanced AI algorithms to automatically create, manipulate, and modify valuable while diverse data. This survey particularly focuses on the implementation of GAI-driven HDT in IoT-healthcare. We start by introducing the background of IoT-healthcare and the potential of GAI-driven HDT. Then, we delve into the fundamental techniques and present the overall framework of GAI-driven HDT. After that, we explore the realization of GAI-driven HDT in detail, including GAI-enabled data acquisition, communication, data management, digital modeling, and data analysis. Besides, we discuss typical IoT-healthcare applications that can be revolutionized by GAI-driven HDT, namely personalized health monitoring and diagnosis, personalized prescription, and personalized rehabilitation. Finally, we conclude this survey by highlighting some future research directions.

I. INTRODUCTION

IoT-healthcare provides real-time health-data collection and monitoring, while HDT extends this toward individualized digital replicas and treatment simulations. The survey examines how GAI can support HDT despite data scarcity, bias, noise, and modeling challenges, using a systematic review process.

A. Background

  • IoT-healthcare collects and transmits health data in real time for analysis, alerts, and proactive interventions, supporting monitoring of glucose, cardiac, respiration, and blood pressure.HDT is envisioned as a digital testbed for simulating treatments and developing personalized plans.Source: A. Background · p. 1
  • HDT digitally replicates an individual and reflects physical status in real time, but requires high-fidelity modeling and comprehensive individual-level data.Data scarcity, bias, noise, and intricate modeling make these requirements difficult to meet.Source: A. Background · pp. 1–2
  • The survey focuses on how GAI models can generate and modify diverse data and support GAI-driven HDT in IoT-healthcare.The named models include GANs, VAEs, transformers, and diffusion models.Source: A. Background · p. 2

B. Review Methodology

  • The review follows PRISMA’s identification, screening, eligibility, and inclusion phases to examine GAI techniques for HDT in IoT-healthcare.It narrowed 1,736 papers after identification to 369 after title-and-abstract screening, then included 172 after full-text screening.Source: B. Review Methodology · p. 2
  • The search used Google Scholar, Scopus, and Web of Science, included only English-language papers, and initially identified 2,416, 900, and 1,679 papers, respectively.Duplicates and papers judged not highly relevant were removed during identification.Source: B. Review Methodology · p. 2

C. Related Work and Contributions

Earlier surveys separately addressed IoT-healthcare, HDT, or GAI, leaving the role of GAI in HDT for IoT-healthcare unexamined. This survey brings these topics together, reviews implementation and applications, and identifies open issues and future directions.

Related-work gap

  • Earlier IoT-healthcare surveys emphasized monitoring, while HDT surveys covered personalized care and GAI surveys covered healthcare applications; none addressed GAI’s role in HDT for IoT-healthcare.The survey presents itself as the first in-depth review of this combined topic.Source: C. Related Work and Contributions · pp. 2–3

Survey contributions

  • The survey reviews HDT and GAI fundamentals, differences between HDT and conventional DT, and popular GAI models.It also explains how GAI supports HDT framework components, including acquisition, modeling, communication, and data management.Source: C. Related Work and Contributions · p. 3
  • The survey covers GAI-driven HDT applications in personalized health monitoring and diagnosis, prescription, and rehabilitation.It also outlines open issues and future directions for the field.Source: C. Related Work and Contributions · p. 4

II. FUNDAMENTALS OF HDT AND GAI

HDT represents an individual human rather than a non-living entity and must account for human-specific physiology, behavior, evaluation needs, data complexity, and mobility. Its framework comprises five components, each of which the survey considers for GAI support.

A. Human Digital Twin

  • HDT represents human beings, whereas conventional DT focuses on non-living physical entities such as machines and networks.HDT also differs in its human-centered evaluation metrics, which include usability and user experience rather than only production and economic performance.Source: A. Human Digital Twin · p. 4
  • HDT must represent physiological and psychological attributes and more uncertain, complex human behaviors, rather than predominantly predetermined machine behaviors.Relevant attributes include physiological signals, perceptual abilities, emotional states, and personality characteristics.Source: A. Human Digital Twin · p. 4
  • High-fidelity HDT modeling requires diverse, complex data from multiple sources, including physiological, environmental, and social-media data.Human positional and postural mobility can also affect radio propagation and signal strength.Source: A. Human Digital Twin · pp. 4–5
  • The HDT framework comprises data acquisition, communication, data management, digital modeling, and data analysis.The paper calls the individual in the physical world the physical twin (PT) and its digital counterpart the virtual twin (VT).Source: A. Human Digital Twin · p. 5

B. GAI Techniques

The survey describes GANs, VAEs, transformers, and diffusion models as GAI techniques used in HDT for IoT-healthcare. Their reported applications include synthesizing and reconstructing human or health data, modeling biological processes, and supporting prediction and monitoring.

Generative adversarial network

  • GANs use adversarially trained generator and discriminator networks to produce samples resembling training data, and chained GANs can model tissue pathology.The cited chained approach uses shape generation, heuristic-based initial simulation, and detail refinement.Source: B. GAI Techniques · p. 5

Variational autoencoder

  • VAEs encode data into a low-dimensional latent space and decode samples to reconstruct inputs; applications include completing motion and predicting clinical trajectories,,.The cited stroke trajectory model was trained on data from 1,216 patients.Source: B. GAI Techniques · pp. 5–6

Transformer

  • Transformers use encoder-decoder architectures and self-attention to process contextual and long-distance dependencies, with applications to patient inquiries and gene-cell modeling,.The survey also describes their use with sequential health data such as EHRs and sensor information.Source: B. GAI Techniques · p. 6

Diffusion model

  • Diffusion models add noise through sequential transformations and reverse the process to generate samples; HDT applications include EEG generation, PET denoising, and 3D human generation [57],, [59].The cited EEG approach uses electrode-frequency maps labeled with emotions, while the PET approach uses an assistive modality embedding as prior information.Source: B. GAI Techniques · p. 6

Other GAI models

C. Framework of GAI-driven HDT

The GAI-driven HDT framework combines data acquisition, communication, data management, digital modeling, and data analysis. These components support IoT-healthcare applications including personalized monitoring and diagnosis, prescription, and rehabilitation.

Framework components

  • GAI enriches datasets with generated health data when traditional acquisition from IoT sensors and medical institutions is inefficient.Acquisition commonly includes EEG, ECG, and MRI data.Source: C. Framework of GAI-driven HDT · p. 6
  • GAI-enabled semantic and cross-modal communication can support HDT transmissions by generating transmitted data at the receiver.These approaches address large-scale, multimodal, time-sensitive communication needs.Source: C. Framework of GAI-driven HDT · pp. 6–7
  • GAI supports data management through imputation and denoising, and can generate deidentified data while preserving original data patterns.Data management also requires preprocessing and security and privacy schemes for large-scale, complex, leakage-sensitive data.Source: C. Framework of GAI-driven HDT · p. 7
  • GAI can model cells, tissues, organs, and other human structures, and analyze data through classification, segmentation, anomaly detection, and prediction.The framework uses these components to support personalized monitoring and diagnosis, prescription, and rehabilitation.Source: C. Framework of GAI-driven HDT · p. 7, Figure 2 · p. 7

III. GAI-DRIVEN HDT IMPLEMENTATION

GAI-based data acquisition addresses the cost, time, intrusiveness, and scalability limits of collecting extensive individual-level HDT data. The survey reviews synthetic physiological, medical imaging, and motion data, including methods and reported findings for each.

A. GAI-enabled Data Acquisition

  • GAI can generate diverse, realistic synthetic data to address time-consuming, costly, intrusive, and poorly scalable acquisition of individual-level data.Examples include physiological, medical imaging, and motion data.Source: A. GAI-enabled Data Acquisition · pp. 7, 9
  • SynSigGAN generates PPG signals using a bidirectional grid LSTM generator and CNN discriminator, while SSSD-ECG generates synthetic 12-lead ECG data.SynSigGAN was trained on BIDMC PPG and respiration datasets; SSSD-ECG used the PTB-XL dataset of 21,837 records from 18,885 patients.Source: A. GAI-enabled Data Acquisition · pp. 7–8, Figure 3 · p. 8
  • The diffusion-based histopathology approach outperformed a GAN-based approach in generating high-quality brain-cancer images, while RealNVP was reported as an attractive alternative to GAN and diffusion approaches.The diffusion method uses color normalization and perception-prioritized weighting; RealNVP was trained on chest X-ray and skin cancer datasets.Source: A. GAI-enabled Data Acquisition · p. 8, Figure 4 · p. 9
  • VAE-based motion generation can reconstruct full-body movement from head and hand signals, and AGRoL achieved SOTA performance with real-time inference speed.AGRoL uses block-wise timestep embedding and generates realistic, smooth motions without extra losses.Source: A. GAI-enabled Data Acquisition · p. 9, Figure 5 · p. 9

B. GAI-enabled Communication

HDT communication must synchronize physical and virtual twins in real time while carrying large, complex, multimodal information. GAI-enabled semantic and cross-modal communication offers ways to reduce transmission demands and reconstruct information at the receiver.

B. GAI-enabled Communication

  • A GAN-aided semantic communication framework reduced communication cost by up to 93.45% while maintaining original information.It extracts essential image or video semantics with a lightweight model and reconstructs and denoises semantic frames at the receiver.Source: B. GAI-enabled Communication · p. 10
  • Semantic diffusion can reconstruct semantically consistent images from compressed semantic images sent over noisy channels, supporting lower-bandwidth rehabilitation monitoring.The described rehabilitation workflow extracts motion features, transmits compressed semantics, and regenerates images or videos for virtual-twin synchronization, [87].Source: B. GAI-enabled Communication · p. 10
  • GAI-enabled cross-modal methods reconstruct missing haptic signals from visual signals, including unidirectional image-to-haptic and bidirectional visual-haptic mappings.The proposed VAE and conditional-flow approach maps between visual and haptic latent spaces; the text gives virtual acupuncture training as an example.Source: B. GAI-enabled Communication · pp. 10–11

C. GAI-enabled Data Management

HDT data management must handle heterogeneous, noisy, incomplete, and sensitive individual-level data. The survey describes GAI methods for imputation, modality translation, denoising, and privacy-preserving data synthesis.

C. GAI-enabled Data Management

  • GAIN outperformed MICE and missForest on NRMSE and PFC for clinical-data imputation at 20% and 50% missingness.Its generator imputes missing components conditioned on observed values, while a discriminator distinguishes observed from imputed components using a hint vector.Source: C. GAI-enabled Data Management · p. 11
  • For MRI-to-CT translation, the conditional diffusion method achieved SSIM = 0.86 and PSNR = 27.31, exceeding GAN (SSIM = 0.83, PSNR = 26.84) and CNN-based methods (SSIM = 0.84, PSNR = 26.98).UMM-CSGM also generated missing-modality images with higher fidelity and structural information than GAN-based methods –.Source: C. GAI-enabled Data Management · p. 12
  • PET-DDPM achieved SSIM = 0.91 and PSNR = 33. 84, outperforming U-Net-based denoising networks with SSIM = 0.87 and PSNR = 33.32.DDM2 suppressed MRI noise and restored anatomical details; compared with N2S, it showed average improvements of 0.95 in SNR and 0.93 in Contrast-to-Noise Ratio.Source: C. GAI-enabled Data Management · p. 12
  • At the same generated-sample quality, ADS-GAN reduced identifiability by 41% compared with WGAN-GP.At the same identifiability, ADS-GAN was 27% and 67% better than PATE-GAN in JSD and Wasserstein distance, respectively.Source: C. GAI-enabled Data Management · pp. 12–13

D. GAI-enabled Digital Modeling

GAI-based digital modeling is presented as an alternative to mechanistic modeling, which requires extensive accurate parameters and domain knowledge. The survey reviews generative models of cells, tissues, and organs, with applications including cellular response prediction and simulated pathology.

D. GAI-enabled Digital Modeling

  • GAI can model from learned data distributions and temporal relations, addressing mechanistic modeling’s demands for accurate simulation parameters and extensive prior domain knowledge.The survey concludes that data-driven GAI modeling can overcome these shortcomings and is promising for HDT.Source: D. GAI-enabled Digital Modeling · pp. 13, 15
  • VAE-based scGen models cellular responses to drug perturbations across cell types, while CPA predicts responses to unseen drugs, combinations, and dosages.The survey also describes conditional β-VAE cell-image generation and scGPT, whose attention captures gene-to-gene interactions and adds interpretability.Source: D. GAI-enabled Digital Modeling · pp. 13–14, Figure 7 · p. 14
  • PathologyGANs generate high-fidelity cancer tissue images that pathologists could not reliably distinguish from real images, while ScarGAN simulates scar tissue on healthy myocardium.ScarGAN uses a mask generator, a domain-specific heuristic, and a refining generator; experienced radiologists could not distinguish its simulated scars from real ones.Source: D. GAI-enabled Digital Modeling · p. 14, Figure 8 · p. 14
  • GAN-based vertebra modeling simulated fracture risk, while a hybrid UNet-GAN generated myocardial velocity maps that quantitatively assessed left-ventricle motion in three orthogonal directions.The vertebra digital twin used synthetic trabecular microstructures, patient CT scans, and finite-element modeling; its reported fracture-risk evaluation concerned a cancer patient with spinal metastasis.Source: D. GAI-enabled Digital Modeling · pp. 14–15

E. GAI-enabled Data Analysis

The survey presents GAI as a means of analyzing HDT data for classification, segmentation, anomaly detection, and prediction, areas where traditional methods may depend on extensive labeled data or struggle with high-dimensional data.

Classification

  • CNN-DVIT combines CNN features with deformable attention and spatial pyramid pooling for varied-length ECGs, and outperformed recent transformer-based ECG classification methods.The survey also describes ViT strategies for Alzheimer’s classification from 3D brain MRI.Source: E. GAI-enabled Data Analysis · p. 15, Figure 9 · p. 15

Segmentation

  • CIMD produces multiple plausible segmentation masks from one image, while DARL’s diffusion and generation modules support vessel-mask generation and synthetic angiograms.CIMD was validated on CT, ultrasound, and MRI and outperformed existing state-of-the-art ambiguous segmentation networks in accuracy while preserving natural variation; DARL outperformed existing unsupervised and self-supervised vessel-segmentation methods.Source: E. GAI-enabled Data Analysis · pp. 15–16, Figure 10 · p. 16

Anomaly Detection

  • CVAD combines multi-scale latent representations with reconstruction error and out-of-distribution probability to detect anomalies; DDIM-based translation preserves unaffected image details while representing diseased regions.The survey notes that VAE- and GAN-based anomaly detectors can be difficult to train or may fail to preserve fine image details.Source: E. GAI-enabled Data Analysis · pp. 16–17

Prediction

  • A β-VAE model forecast clinical-measurement trajectories for patients who later experienced ischemic stroke, while four BERT-based models outperformed ophthalmologist clinical predictions for glaucoma progression.The β-VAE was trained using trajectories from 1,216 patients with ischemic stroke in the MIMIC-IV database.Source: E. GAI-enabled Data Analysis · p. 17

F. Promoting the GAI-driven HDT

The survey recommends involving patients and healthcare professionals in developing and adopting GAI-driven HDT, and describes applications spanning monitoring and diagnosis, prescription, and rehabilitation. For monitoring, ECG analysis can detect abnormalities and classify arrhythmias, with reported performance advantages over traditional methods.

F. Promoting the GAI-driven HDT

  • Healthcare-professional training should address GAI-driven HDT’s benefits, limitations, privacy concerns, and possible adverse effects so professionals can use it and communicate about it effectively.Source: F. Promoting the GAI-driven HDT · p. 17
  • Collaborative development should align applications with healthcare workflows and incorporate patients’ demands and concerns to improve usefulness and user-friendliness.Source: F. Promoting the GAI-driven HDT · p. 17
  • The survey calls for transparency, consent, bias testing, accountability, and privacy safeguards, including options such as federated learning and differential privacy.Source: F. Promoting the GAI-driven HDT · p. 17

IV. GAI-DRIVEN HDT IN IOT-HEALTHCARE APPLICATIONS

  • GAI-driven HDT applications discussed include personalized monitoring and diagnosis, prescription, and rehabilitation.Source: §IV · p. 17

A. Personalized Health Monitoring and Diagnosis

  • A patient’s virtual twin can analyze incoming health data for abnormalities and trigger alerts to the patient or providers; the cited ECG approach detects abnormal signals, then classifies arrhythmia types.Its first level uses MadeGAN for anomaly detection, and its second level combines transfer learning with multi-branch classification to address data scarcity and imbalance.Source: A. Personalized Health Monitoring and… · p. 18, Figure 11 · p. 18
  • MadeGAN achieved AUROC 0.954 and AUPRC 0.936, exceeding AE (0.906 and 0.886) and BeatGAN (0.941 and 0.922) on the respective metrics.The survey also reports that Paige products reduced cancer-detection errors by 70% and time to diagnosis by 65.5%.Source: A. Personalized Health Monitoring and… · p. 18

B. Personalized Prescription

The survey describes GAI-driven HDT as a setting for predicting and simulating treatment outcomes to support personalized prescriptions. It also presents a pose-guided diffusion model for enriching rehabilitation training data.

B. Personalized Prescription

  • SNF-CVAE integrates drug-similarity matrices and drug-disease interactions to predict novel interactions, with outputs used to select a patient’s optimal drug.The survey frames virtual testing of candidate drugs against a patient’s disease as a way for doctors to derive personalized prescription insights.Source: B. Personalized Prescription · p. 19
  • AUC-PR = 0.970 for SNF-CVAE versus AUC-PR = 0.225 for Pareto dominance and collaborative filtering-based drug-disease prediction.SNF-CVAE combines drug-similarity measures, similarity selection, similarity network fusion, and a collective VAE to predict drug-disease interactions.Source: B. Personalized Prescription · p. 19
  • For rehabilitation, a pose-guided conditional diffusion model enriches training data, while the physical twin’s motions are collected to derive personalized insights.The model combines latent diffusion with ControlNet; the physical twin also mimics the virtual twin’s motions.Source: Figure 13 · p. 20
  • INS018 055 reached Phase 2a trials in under four years from discovery, compared with the typical 10-to-15-year drug-development timeline described by the survey.The cited example was nominated as a preclinical candidate 18 months after target discovery; the survey dates the Phase 2a status to March 2024.Source: B. Personalized Prescription · pp. 19–20
  • The survey says more diverse and accurate treatment-outcome predictions and simulations could support more versatile personalized prescriptions and optimize treatment outcomes.Source: B. Personalized Prescription · p. 20

C. Personalized Rehabilitation

Personalized rehabilitation uses a virtual twin (VT) to simulate patient-specific exercises, monitor movement, and provide coaching, with GAI-generated data supporting monitoring and evaluation. The section also identifies energy use and human-centric evaluation as implementation challenges.

C. Personalized Rehabilitation

  • Therapists set the VT’s target posture to simulate patient-specific training; IoT motion data returns to the VT, which coaches posture correction and progress.For balance training, the therapist can create an unstable state tailored to the patient’s profile, which the patient attempts to mimic.Source: C. Personalized Rehabilitation · p. 20
  • IS = 162.94, FID = 299.06, SSMI = 0.545, and DS-person = 0.789 were reported for Mennella et al.’s synthetic rehabilitation images.Their latent diffusion and ControlNet framework was trained on a pre-labeled dataset of six rehabilitation exercises.Source: C. Personalized Rehabilitation · p. 20
  • Synthetic datasets can target postural correction, joint mobility, balance, and functional movement, and can be customized to patient profiles for personalized training.The survey identifies GAI-based data augmentation as a way to enhance VT monitoring and evaluation.Source: C. Personalized Rehabilitation · p. 20

A. Deriving Energy-Efficient Scheme

  • High model complexity and large datasets demand substantial computation, communication, and storage; energy-saving techniques may sacrifice accuracy, requiring a balance with sustainability.The section cites 10,000 GPUs used to run an HDT brain as an example of the resource burden.Source: A. Deriving Energy-Efficient Scheme · p. 21

B. Designing Human-centric Evaluation Metrics

  • Human-centric evaluation should measure generated content against patients’ real health data, including VT behavior, responses, and predictive capabilities.The survey proposes combining GAI metrics such as inception score and Frechet inception distance with healthcare knowledge, then continually evaluating and refining the system.Source: B. Designing Human-centric Evaluation… · p. 21

C. Accelerating Real-time Responsiveness

The survey identifies slow GAI generation as a barrier to responsive HDT interactions and interventions, and discusses computational approaches to reduce delays. It also reviews GAI-driven HDT components, healthcare-system integration, applications, and future directions.

C. Accelerating Real-time Responsiveness

  • Slow GAI generation can delay immersive VT interactions and urgent diagnostics or treatment recommendations; algorithm optimization and parallel or distributed computing are proposed to reduce latency.The passage frames timely responsiveness as important for healthcare outcomes.Source: C. Accelerating Real-time Responsiveness · p. 21

D. Seamless Integration of GAI-driven HDT with Existing Healthcare IT Ecosystems

  • Integration with existing healthcare IT requires data exchange and interoperability; HL7 and FHIR offer uniform formats, but their complexity and data mapping pose challenges.The survey also identifies standardized APIs and web services as practical solutions to integration issues.Source: D. Seamless Integration of GAI-driven… · p. 21

VI. CONCLUSIONS

  • The survey identifies GAI-generated physiological, medical-imaging, and motion data as support for HDT data acquisition.Source: §VI · p. 21
  • GAI-aided semantic and cross-modal communication schemes are among the reviewed HDT capabilities.Source: §VI · p. 21
  • For HDT data management, GAI can impute missing data, denoise data, and synthesize authentic data while preserving distributions and supporting privacy and security.Source: §VI · p. 21
  • The survey describes GAI as supporting HDT digital modeling of human cells, tissues, and organs.Source: §VI · p. 21
  • GAI can facilitate HDT data analysis through classification, segmentation, anomaly detection, and prediction.Source: §VI · p. 21
  • The survey covers personalized monitoring and diagnosis, prescription, and rehabilitation, and outlines future research directions.Source: §VI · p. 21
Loading 2401.13699v2…