Source-linked AI summary
A Revolution of Personalized Healthcare: Enabling Human Digital Twin with Mobile AIGC
Jiayuan Chen, Changyan Yi, Hongyang Du, Dusit Niyato, Jiawen Kang, Jun Cai, Xuemin, Shen
TL;DR
Personalized healthcare HDTs require large amounts of individual-level, multimodal data, while conventional collection can be difficult to personalize and scale. The paper proposes a mobile AIGC-driven HDT architecture and examines its design challenges and applications. In a virtual physical-therapy testbed, the proposed CODI approach outperforms SAC and PPO in total users’ QoE.
Problem
HDT for personalized healthcare depends on extensive individual-level data, which traditional collection methods can struggle to personalize and scale.
Source: 1) Data Collection · p. 2Method
The paper proposes a mobile AIGC-driven HDT architecture and lifecycle, analyzes its design challenges, illustrates two use cases, and studies a conditional diffusion approach in a physical-therapy testbed.
Source: B. Related Work and Contributions · p. 2, B. Solution and Result Analysis · p. 7Results
In the virtual physical-therapy testbed, CODI achieves higher total users’ QoE than PPO and SAC as the number of users varies.
Source: B. Solution and Result Analysis · p. 7Takeaways & Limitations
The article presents mobile AIGC-driven HDT as a framework for personalized healthcare and reports a testbed case study demonstrating its effectiveness.
Source: §V · p. 7The virtual physical-therapy platform study is described as a preliminary exploration.
Source: B. Related Work and Contributions · p. 2
Abstract
from arXiv · showhide
Mobile Artificial Intelligence-Generated Content (AIGC) technology refers to the adoption of AI algorithms deployed at mobile edge networks to automate the information creation process while fulfilling the requirements of end users. Mobile AIGC has recently attracted phenomenal attentions and can be a key enabling technology for an emerging application, called human digital twin (HDT). HDT empowered by the mobile AIGC is expected to revolutionize the personalized healthcare by generating rare disease data, modeling high-fidelity digital twin, building versatile testbeds, and providing 24/7 customized medical services. To promote the development of this new breed of paradigm, in this article, we propose a system architecture of mobile AIGC-driven HDT and highlight the corresponding design requirements and challenges. Moreover, we illustrate two use cases, i.e., mobile AIGC-driven HDT in customized surgery planning and personalized medication. In addition, we conduct an experimental study to prove the effectiveness of the proposed mobile AIGC-driven HDT solution, which shows a particular application in a virtual physical therapy teaching platform. Finally, we conclude this article by briefly discussing several open issues and future directions.
I. INTRODUCTION
HDT represents an individual’s physical and psychological status in virtual space, but personalized healthcare requires large amounts of sensitive, multimodal data that traditional collection methods struggle to personalize and scale. The article presents mobile AIGC as a way to support HDT and outlines its architecture, challenges, use cases, and preliminary testbed study.
A. Mobile AIGC-Driven Human Digital Twin
- HDT mirrors an individual’s psychological and physiological status in virtual space, but its implementation depends on extensive multimodal data that are difficult to collect at the individual level.Traditional collection can be hard to personalize and scale, particularly because medical records are highly private.Source: A. Mobile AIGC-Driven Human Digital Twin · p. 1
- Mobile AIGC can generate rare-disease data, balance data-based and model-based HDT modeling, support healthcare testbeds, and provide customized services around the clock.The article discusses examples including synthetic medical data, candidate treatment testing, generated haptic responses, drug-interaction simulation, and doctor HDTs.Source: A. Mobile AIGC-Driven Human Digital Twin · pp. 1–2
B. Related Work and Contributions
- Prior work surveyed HDT, AIGC, and medical generative models, but the article identifies no existing work addressing AIGC-enabled HDT for personalized healthcare.It responds with a mobile AIGC-driven HDT framework and discussion of its design requirements and challenges.Source: B. Related Work and Contributions · p. 2
- The article proposes a holistic architecture and lifecycle, analyzes design challenges, illustrates two use cases, presents a preliminary virtual physical-therapy platform case study, and discusses future research directions.The lifecycle covers data collection, management, pre-training, fine-tuning, inference, and product management.Source: B. Related Work and Contributions · p. 2
A. System Architecture and Key Techniques
The mobile AIGC-driven HDT architecture distributes its lifecycle across mobile-device, edge, and cloud layers. Its initial stages collect personalized data and manage diverse, sensitive datasets for model training and sharing.
A. System Architecture and Key Techniques
- The architecture spans mobile-device, edge, and cloud layers, with data collection, management, pre-training, fine-tuning, and inference circulating across core and edge networks.The architecture figure also presents a timeline of major technical developments.Source: A. System Architecture and Key… · p. 2, Figure 1 · p. 3
1) Data Collection:
- Mobile devices collect public and individual data for model training, including personal signals gathered through wearable and implantable biomedical devices.Hospital EHRs may lack sufficient rare-disease samples; wider administrative claims datasets can be an alternative for rare-disease diagnosis models. Social networks may supply psychology-related data.Source: 1) Data Collection · pp. 2–3
2) Data Management:
- Data management preprocesses and robustly stores large, complex datasets from physical and virtual spaces, while addressing the ethical risks of sensitive-data leakage.Diffusion models and GANs have been applied to medical-data preprocessing tasks including translation, reconstruction, registration, classification, segmentation, and denoising.Source: 2) Data Management · p. 3
3) AIGC Pre-training:
The paper describes a mobile AIGC lifecycle in which cloud resources pre-train models, cloud or edge resources fine-tune them, and edge servers or mobile devices can run inference. Generated medical content and responses can support HDT tasks and services.
3) AIGC Pre-training:
- Cloud servers pre-train generative models to learn features from collected data, with technology choice depending on task requirements, data amount, desired output, and available resources.Diffusion models learn an inverse diffusion process to generate medical images and are used for their mode coverage and content quality.Source: 3) AIGC Pre-training · p. 3
4) AIGC Fine-tuning:
- Fine-tuning adjusts a pre-trained AIGC model to a new task using modest amounts of data, and can be performed at the cloud or edge using processed mobile-device data.An example is generating synthetic MRI data for rare diseases.Source: 4) AIGC Fine-tuning · p. 3
5) AIGC Inference:
- Because inference generally needs less computing power than pre-training, it can run on edge servers or mobile devices to provide low-latency, private services and generate content for HDT tasks.Generated medical images and 3D models can support HDT modeling and evolution; generated responses can support testbeds and ubiquitous medical services.Source: 5) AIGC Inference · p. 4
B. Key Design Requirements and Challenges
HDT self-evolution must be personalized to each physical–virtual twin pair and adapt to available resources as physical status changes. The paper identifies resource orchestration, timely data integration, and pervasive connectivity as challenges for that process.
1) Personalized and Adaptive Self-evolution of Mobile AIGC-driven HDT:
- HDT self-evolution must synchronize each physical–virtual twin pair as physical status changes, while adapting to differences in individual characteristics and available resources.Relevant resources include computation, communication, and data accessibility.Source: 1) Personalized and Adaptive… · p. 4
- Different task QoS needs, increasingly complex models, and user mobility make unified allocation of computing and networking resources challenging.The passage gives network slicing-based allocation as one potential multi-objective solution.Source: 1) Personalized and Adaptive… · p. 4
- Dynamic fine-tuning requires timely collection and integration of new bodily data, while fine-tuned models can generate rare-disease data to expand training datasets.The passage notes that this process may require 6G communications and Time-Sensitive Networking.Source: 1) Personalized and Adaptive… · p. 4
- Positional mobility can change radio propagation or trigger service migration, while postural mobility can cause body shadowing, interruptions, or packet loss.These mobility effects can impact reliability, stability, security, and cost of model self-evolution; the passage calls for pervasive connectivity.Source: 1) Personalized and Adaptive… · p. 4
2) Customized and Multi-modal Intelligent-interaction of Mobile AIGC-driven HDT:
Mobile AIGC-driven HDT combines virtual twins and multimodal interaction to support personalized healthcare services. Its design must address synchronized data delivery, privacy and security, and evaluation that captures both technical performance and user experience.
2) Customized and Multi-modal Intelligent-interaction of Mobile AIGC-driven HDT:
- VTs can provide customized feedback and immersive healthcare interactions through audio, video, and touch.The passage gives virtual surgery as an example of an interactive scene.Source: 2) Customized and Multi-modal… · p. 4
- Multimodal feedback must synchronize auditory, visual, and haptic streams; surgery simulation requires <1 ms end-to-end latency and >99.999999% reliability.For visual and tactile movement, the perceived interval should not exceed 1 ms; virtual outpatient service has different transmission requirements.Source: 2) Customized and Multi-modal… · p. 4
- Exchanging large volumes of data between PTs and VTs raises privacy, security, integrity, availability, and ethical concerns.The passage identifies differential privacy, secure multiparty computation, and homomorphic encryption as techniques to design and implement.Source: 2) Customized and Multi-modal… · pp. 4–5
- Evaluation must combine technical measures such as latency and reliability with users’ experience and professional assessment of generated healthcare content.The passage notes that generated content can sound reasonable yet be factually unreliable, and that integrated evaluation is difficult to design.Source: 2) Customized and Multi-modal… · p. 5
III. APPLICATION OF MOBILE AIGC-DRIVEN HDT IN PERSONALIZED HEALTHCARE
Mobile AIGC-driven HDT is presented for customized surgery planning and personalized medication. The applications include improving medical imaging data and modeling patient-specific drug responses and rare-disease care.
A. Mobile AIGC-driven HDT in Customized Surgery Planning
- High-fidelity 3D heart models can help physicians assess lesions and develop individualized surgery plans, but limited labeled imaging data constrains segmentation-model accuracy and generalization.The passage attributes limited dataset scale to medical-data privacy and annotation difficulty.Source: A. Mobile AIGC-driven HDT in… · p. 5
- AIGC can generate labeled TOF-MRA patches without manual annotation, while score-based reconstruction and flow-based augmentation have been applied to surgery-planning images and data.The passage says the reconstruction and augmentation approaches substantially improved customized-surgery performance.Source: A. Mobile AIGC-driven HDT in… · p. 5
B. Mobile AIGC-driven HDT in Personalized Medication
- Virtual medication testing simulates patient-specific drug metabolism, efficacy, and toxic side effects across doses so doctors can optimize treatment plans.The passage notes that optimization can account for slower metabolism or excessive reactions associated with genetic factors.Source: B. Mobile AIGC-driven HDT in… · pp. 5–6
- The medication use case also encompasses drug-disease interaction prediction, rare-disease diagnosis, and anomaly detection, with the article describing potential 24/7 access to customized care.Examples include RareBERT for XLH diagnosis and a diffusion model for brain-tumor MRI anomaly detection.Source: B. Mobile AIGC-driven HDT in… · p. 6
IV. A CASE STUDY OF MOBILE AIGC-DRIVEN HDT IN PERSONALIZED HEALTHCARE
The case study builds an edge-deployed virtual physical therapy teaching platform for trainer and trainee interaction with patient virtual twins. It uses semantic video communication and optimizes video resolution and diffusion steps against user QoE and resource constraints.
A. Experimental Setup and System Design
- The edge-server platform lets trainers and trainees use virtual twins to massage a patient virtual twin, with haptic signals and VR video streams providing immersive feedback.The platform uses tactile Internet to transmit haptic signals to users’ tactile devices.Source: A. Experimental Setup and System Design · p. 6, Figure 3 · p. 7
- Semantic extraction and encoding reduce VR-video data size and communication overhead, while an edge-side AIGC model generates a resembling video from received semantic information and skeleton data.The approach is intended to address low-latency interaction and video corruption from collection errors or packet loss.Source: A. Experimental Setup and System Design · p. 6
- The optimization targets total users’ QoE under bandwidth, computation, and QoE-threshold constraints by choosing resolution ratio and diffusion step.The passage defines QoE as a linear combination of VR-video bitrate and similarity to the actual video; resolution ratio and diffusion step are proportional to bitrate and similarity, respectively.Source: A. Experimental Setup and System Design · pp. 6–7
B. Solution and Result Analysis
In the virtual physical therapy testbed, CODI achieves higher total users’ QoE than PPO and SAC as user count varies. The authors attribute its stronger decisions to iterative denoising, which reduces the effect of uncertainty and noise.
B. Solution and Result Analysis
- CODI is superior to PPO and SAC in total users’ QoE across the compared user amounts.Fig. 4(b) compares total users’ QoE against amount of users.Source: Figure 4 · p. 7, B. Solution and Result Analysis · p. 7
- During training, PPO takes more steps to converge, while SAC stabilizes faster at a higher reward than PPO but finishes below CODI.Fig. 4(a) compares obtained reward with training steps.Source: Figure 4 · p. 7, B. Solution and Result Analysis · p. 7
- CODI’s iterative denoising gradually adjusts decisions, reducing uncertainty and noise compared with SAC and PPO’s one-step environment-to-output mapping.The decisions are users’ resolution ratios and diffusion steps.Source: B. Solution and Result Analysis · p. 7
V. CONCLUSIONS AND FUTURE RESEARCH DIRECTIONS
The article presents mobile AIGC-driven HDT as a framework for personalized healthcare and identifies data sharing, cross-modal generation, and sustainable deployment as open challenges.
V. CONCLUSIONS AND FUTURE RESEARCH DIRECTIONS
- The proposed framework includes a system architecture, design requirements, two potential applications, and a testbed case study demonstrating its effectiveness.The two applications are customized surgery planning and personalized medication, as stated in the paper abstract.Source: §V · p. 7
- Medical-data silos reflect fragmented institutional systems and privacy concerns, which contribute to data heterogeneity and bias.The authors call for unified, secure data management, potentially integrating blockchain and federated learning, to train a full-featured HDT.Source: §V · pp. 7–8
- Cross-modal generation is challenged by complex modality matching and nonstandard data formats across institutions.The authors suggest cross-domain collaboration using meta-learning and open-set learning to characterize modality relationships and improve robustness and generalization.Source: §V · p. 8
- Asynchronous transfers of high-dimensional model parameters consume substantial communication and computation energy, making efficient, sustainable training and deployment an open issue.Suggested approaches include green computing and communication, pruning, quantization, and knowledge distillation.Source: §V · p. 8
BIOGRAPHY
The biographies list the contributors’ affiliations and research interests across edge computing, digital twins, generative AI, communications, and networking.
BIOGRAPHY
- Chen’s interests include edge computing and digital twins; Yi’s include edge computing, industrial IoT, digital twins, and 5G and beyond.Chen is pursuing a Ph.D. at Nanjing University of Aeronautics and Astronautics; Yi is a professor there.Source: BIOGRAPHY · p. 8
- Du’s interests include semantic communications, generative AI, and the Metaverse; Niyato’s include edge intelligence, machine learning, and incentive mechanism design.Both are affiliated with Nanyang Technological University’s School of Computer Science and Engineering.Source: BIOGRAPHY · p. 8
- Kang’s interests include blockchain, IoT, and the Metaverse; Cai’s include edge/fog computing and eHealth.Kang is a professor at Guangdong University of Technology, while Cai is a professor and PER-FORM Centre Research Chair at Concordia University.Source: BIOGRAPHY · p. 8
- Shen’s research focuses on network resource management, wireless network security, social networks, and vehicular ad hoc networks.He is a University Professor at the University of Waterloo.Source: BIOGRAPHY · p. 8