Edge-intelligent AIoT digital twins for scalable and privacy-aware healthcare monitoring

1 Introduction

Continuous and personalized healthcare monitoring has been increasingly enabled by advances in wearable sensors, Internet of Things (IoT) devices, and artificial intelligence (AI) (Ji et al., 2024). These technologies allow physiological signals, movement patterns, and daily activities to be observed beyond clinical environments, supporting applications such as remote patient monitoring, rehabilitation, and chronic disease management. Such systems are particularly relevant for remote elderly care, assisted living facilities, and post-discharge rehabilitation monitoring, where continuous mobility assessment and fall-risk detection are essential for patient safety and recovery.

However, most existing healthcare AI systems remain cloud-centric and monitoring-oriented, relying on continuous data streaming and centralized processing (Karakra et al., 2018; Liu et al., 2019; Elayan et al., 2021; Sharma et al., 2024). Such designs introduce significant challenges in power consumption, communication bandwidth, privacy exposure, and scalability, particularly when deployed across large patient populations over extended periods.

Digital twins have recently emerged as a promising paradigm for modeling complex and evolving systems (Huang et al., 2024). In healthcare, a digital twin represents a dynamic virtual counterpart of a patient that integrates multimodal data over time to track health states and trajectories (see Figure 1). While digital twins offer the potential for personalized care, current implementations often inherit the limitations of conventional healthcare IoT systems, including continuous raw data transmission, high resource demands, and increased privacy risk (Katsoulakis et al., 2024).

To address these challenges, this paper proposes an edge-intelligent Artificial Intelligence of Things (AIoT) digital twin system that integrates event-driven intelligence with digital twin modeling (see Figure 2). Rather than continuously transmitting raw sensor data, the proposed system performs sensing and inference locally at the edge and generates semantic digital twin updates only when clinically meaningful changes are detected. This event-driven design reduces unnecessary computation and communication while preserving continuous health state tracking. By shifting intelligence closer to the data source, the system also minimizes raw data exposure, supporting privacy-aware operation.

The main contributions of this paper are threefold. First, we design an edge-intelligent digital twin architecture for healthcare that combines local AI inference to enable continuous yet efficient monitoring. Second, we introduce an event-driven edge update mechanism that significantly reduces on-device computation and communication bandwidth by transmitting semantic health state changes instead of raw data streams. Third, we evaluate the proposed system through computational analysis and simulations that model power consumption and communication behavior using realistic sensor traces, and we conduct a scalability analysis to assess system performance as the number of monitored patients increases. This work focuses on architectural modeling and analytical evaluation to study the system-level implications of event-driven edge intelligence for digital twin healthcare systems.

The remainder of this paper is organized as follows. Section 2 provides a concise literature review and background on AI and digital twin technologies for healthcare systems, motivating the proposed approach. Section 3 presents the proposed system architecture and event-driven digital twin design. Section 4 describes the computational evaluation methodology and experimental setup. Section 5 reports experimental results on power consumption, communication bandwidth, and scalability. Section 6 discusses implications, limitations, and future extensions, followed by conclusions in Section 7.

2 Literature review

2.1 Technical literature on digital twin framework for healthcare

The technical literature on digital twin frameworks for healthcare includes representative studies such as (Karakra et al., 2018; Liu et al., 2019; Elayan et al., 2021; Sharma et al., 2024) (see Table 1), which explore the use of virtual replicas to improve healthcare planning, monitoring, and decision support. Study Karakra et al. (2018) introduces a digital twin framework for hospital services based on Discrete Event Simulation (DES), integrated with hospital information systems and IoT devices, enabling administrators to evaluate resource utilization, patient flow, and operational changes without disrupting daily hospital activities. Study Liu et al. (2019) proposes Cloud Digital Twin Healthcare (CloudDTH), a cloud-based framework that combines big data, cloud computing, and IoT technologies to support personalized health monitoring and lifecycle-oriented care, particularly for elderly populations, and demonstrates feasibility through application scenarios and case studies. Study Elayan et al. (2021) presents a context-aware digital twin healthcare system focused on continuous patient monitoring, where machine-learning–based ECG classification is used to diagnose cardiac conditions and illustrate how digital twins can integrate physiological data with AI-driven analysis. Similarly, study Sharma et al. (2024) develops a digital twin–based intelligent healthcare system that incorporates neural-network–based ECG analysis to support cardiac diagnosis and early abnormality detection, highlighting the potential of digital twins to enable predictive and personalized healthcare services.

TABLE 1. Overview of prior digital twin and AIoT healthcare technical literature.

Taken together, these studies demonstrate that digital twins are effective for healthcare modeling, monitoring, and decision support; however, they predominantly adopt cloud-centric, application-specific architectures and treat sensing and IoT components mainly as data sources, lacking edge-intelligent AIoT integration and resource-aware design required for scalable, privacy-preserving, and long-term healthcare monitoring systems.

2.2 Technical literature on AIoT in healthcare

The technical literature on AIoT in healthcare includes representative studies such as Soltani Zarrin et al. (2020), Wei et al. (2022), Chang et al. (2022), Saleh et al. (2023), and Tseng et al. (2025) (see Table 1), which investigate how embedded intelligence, sensing, and communication technologies can support real-time health monitoring and diagnosis. Study Soltani Zarrin et al. (2020) develops an energy- and computation-efficient spiking neural network (SNN) for early epileptic seizure detection, targeting deployment on neuromorphic edge devices to enable low-power, local analysis of sensitive neural signals without cloud dependence. Study Wei et al. (2022) presents an IoT-based intelligent medical system for traditional Chinese medicine, integrating multimodal physiological sensing, mobile applications, and cloud-based analytics to support diagnostic assessment and treatment recommendations. Study Chang et al. (2022) proposes Flying Watchdog, an AIoT-based drone surveillance system that combines deep learning and mobile edge computing to perform real-time person recognition, fall detection, and anomaly alerts through continuous video streaming. Study Saleh et al. (2023) introduces an AIoT-based embedded system for Parkinson’s disease prediction using vocal disorder analysis, employing optimized machine learning classifiers to enable accessible, real-time medical screening in both local and remote settings. Study Tseng et al. (2025) develops a portable AIoT fall detection system for outdoor environments, integrating inertial sensing with GPS and NB-IoT communication and using a low-complexity algorithm to support accurate, long-term fall detection and emergency alerting under energy constraints.

Taken together, these studies demonstrate the effectiveness of AIoT for task-specific, real-time healthcare sensing and edge-level intelligence; however, they predominantly focus on isolated applications and single-event detection, often relying on continuous data transmission or cloud-assisted analytics, and do not incorporate digital twin modeling (health state representation), or event-driven system-level architectures needed for scalable, privacy-aware, and population-level healthcare monitoring.

2.3 Survey literature on digital twin and/or AIoT in healthcare

Several survey studies have examined healthcare systems involving AIoT and digital twins, including representative works in Majeed et al. (2023), Baker and Xiang (2023), Lee et al. (2025), Singh and Gill (2023), and Angulo et al. (2020). Study Majeed et al. (2023) provides a comprehensive survey of IoT-, IoMT-, and AIoT-based smart healthcare systems published between 2020 and 2022, with particular attention to security mechanisms, intelligent services, and the emerging role of digital twin concepts in healthcare. The survey reviews enabling technologies, application domains, and open research challenges, and outlines potential directions for integrating AI, IoMT, and DTs to advance smart healthcare delivery. Study Baker and Xiang (2023) presents an extensive survey of AIoT for healthcare, motivated by increasing demands for remote and at-home care. It proposes a unified AIoT architecture encompassing sensing devices, communication technologies, and cross-layer intelligence, and reviews recent research to identify key opportunities and challenges across the AIoT stack, supported by representative healthcare use cases. Study Lee et al. (2025) surveys recent advances in AI-integrated sensing technologies for AIoT systems, focusing on self-powered sensors, wearable AI sensors, nanophotonic sensors, and in-sensor computing architectures that enable energy-efficient and autonomous edge intelligence.

Taken together, these survey papers highlight the growing importance of digital twins, AIoT, and embedded intelligence in future healthcare systems; however, these surveys consistently reveal a gap between conceptual frameworks and operational system design, as none presents a concrete, edge-intelligent digital twin architecture with resource-aware computation for scalable and privacy-preserving healthcare monitoring in real-world deployments.

2.4 Research gaps

Across technical and survey literature, digital twins and AIoT have each demonstrated significant potential for advancing healthcare monitoring, decision support, and personalized care. However, existing digital twin studies largely adopt cloud-centric and application-specific designs, treating sensing systems as passive data sources and relying on continuous data transmission and centralized processing. Conversely, AIoT-focused studies emphasize task-level edge intelligence and embedded sensing efficiency but lack digital twin modeling that can represent evolving patient health over time and across populations. Recent survey papers further acknowledge the promise of combining digital twins and AIoT, yet stop short of proposing operational architectures that unify these concepts through edge intelligence, event-driven communication, and resource-aware system design (see Figure 3).

As a result, critical gaps remain in the literature: (1) there is no integrated, experimentally evaluated framework that connects digital twin healthcare modeling with edge-intelligent AIoT architectures, (2) In particular, prior work does not address how digital twins can be maintained using event-driven updates at the edge while preserving privacy, and (3) nor how such designs fundamentally reshape system scalability in terms of power consumption and communication bandwidth for long-term healthcare monitoring.

These gaps motivate the present work, which proposes an edge-intelligent, event-driven AIoT digital twin architecture designed explicitly to support scalable, privacy-aware, and energy-efficient healthcare monitoring, and provides quantitative analysis of its computational, communication, and population-scale behavior. To address these gaps, we first formalize a digital twin model suitable for event-driven edge intelligence and then describe the proposed system architecture and update mechanisms.

3 System design

This section describes the proposed edge-intelligent AIoT digital twin architecture for healthcare monitoring. The system is designed to support continuous yet efficient monitoring by combining local intelligence at the edge with event-driven communication and digital twin state modeling. The architecture explicitly targets reduced power consumption, reduced communication bandwidth, privacy-aware operation, and scalable population-level deployment.

3.1 System architecture

The proposed system adopts a layered, distributed architecture composed of four tightly integrated components: a sensing layer, an edge intelligence layer, a digital twin layer, and a cloud services layer. Rather than continuously streaming raw sensor data to the cloud, intelligence is deliberately distributed across these layers so that most computation and decision-making occur at the edge, while the cloud supports aggregation and population-level analysis. Figure 4 illustrates the overall architecture and data flow.

The sensing layer consists of wearable and ambient IoT devices that collect multimodal health signals. For example, inertial measurement units (IMUs) can be used to capture movement and posture, with optional inclusion of physiological. Sensor data are sampled at modality-appropriate rates and buffered locally. Importantly, raw sensor streams are not transmitted continuously; instead, they serve as inputs for local processing at the edge, minimizing unnecessary communication and limiting exposure of sensitive data.

The edge intelligence layer performs local feature extraction, inference, and event detection on resource-constrained devices. Sensor streams are segmented into temporal windows and transformed into compact feature representations. Lightweight machine learning models then infer health-related indicators, such as movement stability or activity state, using computationally efficient designs suitable for continuous operation. An event-driven mechanism determines whether observed changes are clinically meaningful, based on threshold crossings, anomaly scores, or state transitions. Although edge processing operates continuously, communication and higher-level updates are triggered only when events occur, substantially reducing both computation frequency and communication overhead. Event-driven communication is a central feature of the proposed architecture. Instead of transmitting raw sensor data or frequent inference outputs, the system generates semantic event messages that summarize meaningful health state changes. Each event includes a compact state label, an associated confidence or risk score, and a timestamp, with optional contextual metadata. Events are transmitted only when triggered, supplemented by low-rate periodic summaries. This approach directly reduces on-device computation, communication bandwidth usage, and cloud ingestion and storage requirements.

The digital twin layer maintains a patient-specific state model that evolves over time based on event-driven updates from the edge. Rather than mirroring raw sensor streams, the digital twin represents estimated health states, historical trends, and associated confidence or uncertainty. By integrating incoming events with historical context, the digital twin preserves coherent health narratives even when data are sparse, noisy, or intermittently unavailable.

The cloud services layer supports population-level functions without participating in continuous per-sample processing. Its primary roles include aggregating digital twin updates across patients, enabling clinical dashboards and analytics, and supporting offline model updates or system configuration. Because communication is semantic and event-driven, cloud resource usage scales with the number of meaningful events rather than raw data volume, enabling efficient monitoring across large patient populations.

Overall, the proposed architecture integrates edge intelligence, event-driven communication, and digital twin modeling to enable efficient, privacy-aware, and scalable healthcare monitoring. By shifting computation and decision-making to the edge and updating digital twins only when clinically meaningful changes occur, the system provides a practical foundation for long-term digital twin deployment. Subsequent sections evaluate this architecture through computational experiments focusing on power consumption, communication bandwidth, and scalability.

3.2 Digital twin state representation and update

To improve reproducibility and clarify how digital twin states evolve under event-driven updates, we formalize the state representation and update mechanism used in the proposed system.

In the proposed architecture, the digital twin does not replicate raw sensor streams. Instead, it maintains a compact, semantic representation of the patient’s health status that evolves over time based on event-driven updates generated at the edge. This design aligns the digital twin with clinically meaningful states rather than continuous data streams.

At time t, the digital twin state is defined as H t = s t , r t , c t, where s t denotes the inferred health state (e.g., normal activity, instability, fall event), r t represents a confidence or risk score associated with the inferred state, and c t denotes contextual metadata such as timestamps, activity context, or sensor reliability indicators.

Edge intelligence continuously analyzes sensor streams using lightweight monitoring features and event-triggered inference. When a clinically meaningful change is detected, the edge device generates a semantic event update u k = s k , r k , τ k, where s k is the detected health state label, r k is the associated confidence or risk score, and τ k is the event timestamp.

These event updates are transmitted to the digital twin layer and used to update the patient’s virtual state representation. The evolution of the digital twin state is modeled as H t = f H t − 1 , u k, where f · denotes a state integration function that incorporates the newly observed event with the historical twin state. In this study, f · is treated as an abstract integration mechanism that conceptually combines incoming semantic updates with prior twin states. The simulation framework focuses on the timing and cost of semantic updates rather than detailed clinical semantics of state evolution. More sophisticated implementations, such as temporal smoothing or transition validation, are left for future work.

In this study, a simple baseline integration mechanism can be defined as a last-event update rule, where the current twin state directly reflects the most recent semantic event, H t = u k. Alternatively, a temporally smoothed update can be considered, H t = α H t − 1 + 1 − α u k. For example, in a healthcare monitoring scenario, u k may represent a semantic event derived from physiological signals such as heart rate, respiration, or body temperature, indicating a change in the patient’s condition. The digital twin state H t can then reflect the patient’s overall health status (e.g., stable, mild, or severe), updated either instantly or in a temporally smoothed manner. These simple formulations illustrate how the integration function may operate in practice, while more sophisticated models are left for future work.

Structurally, the digital twin can be implemented as a time-indexed state record, TwinRecord = {patient_id, timestamp, state_label, confidence_score, contextual_metadata} . This representation enables efficient storage, querying, and longitudinal analysis while avoiding continuous transmission of raw sensor data. Because updates occur only when meaningful events are detected (or when periodic summaries are generated), the digital twin remains synchronized with the physical system while maintaining low communication overhead.

In the present simulation framework, semantic event updates are abstracted as compact state messages containing at least a state label, confidence or risk score, and timestamp. Periodic summaries provide low-rate continuity for longitudinal twin maintenance.

The event-driven update process corresponds directly to the semantic event generation mechanism illustrated in Algorithm 1. In this framework, edge intelligence performs continuous monitoring, while the digital twin evolves through sparse but informative state updates rather than raw data replication.

This formalization clarifies how the digital twin maintains a compact and operational health state model while remaining consistent with the event-driven edge architecture proposed in this work.

3.3 Event-driven edge intelligence

Without edge intelligence, digital twin healthcare systems cannot scale operationally. Cloud-only digital twin architectures depend on continuous streaming of raw multimodal data, resulting in unsustainable computation costs, excessive communication overhead, and elevated privacy risks - especially when monitoring large patient populations over long periods. These limitations fundamentally constrain the feasibility of digital twin deployment in real-world healthcare environments. Event-driven edge intelligence addresses these challenges by shifting analysis and decision-making closer to the data source. Instead of processing and transmitting data continuously, edge-intelligent AIoT systems perform local inference and trigger computation or communication only when clinically meaningful changes occur. This paradigm transforms digital twins from data-intensive replicas into adaptive, knowledge-driven models. In other words, event triggers are computed from lightweight features, not the full feature pipeline (see Algorithm 1).

Under this design, sensor streams are analyzed locally to infer health-related indicators. When the inferred state deviates significantly from the previous digital twin state - based on predefined thresholds or anomaly criteria - a semantic update is generated and transmitted. Otherwise, no communication is triggered. This mechanism enables continuous monitoring while avoiding unnecessary computation and data transmission.

Beyond efficiency, event-driven edge intelligence provides privacy by design. Raw sensor data remain local, and only abstracted health state information is transmitted. This minimizes exposure of sensitive physiological and behavioral signals and enables long-term monitoring without continuous surveillance. Privacy is therefore achieved not as a post hoc regulatory constraint, but as an intrinsic system property arising from data minimization.

Scalability emerges naturally from this design. Because communication is event-driven and semantic, per-patient bandwidth usage is reduced, cloud updates are sparse, and edge computation is decoupled from centralized processing. As a result, system resource usage scales with the frequency of clinically meaningful events rather than sensor sampling rates or raw data volume.

Overall, event-driven edge intelligence enables digital twins to operate on knowledge rather than raw data. This shift delivers power-efficient operation on resource-constrained devices, communication-efficient population-scale monitoring, and privacy-preserving long-term deployment. By grounding digital twin updates in meaningful health events, the proposed approach supports scalable, responsive, and responsible healthcare systems.

Algorithm 1 Event-driven digital twin update with lightweight triggering.

4 System analysis

This section analyzes the proposed system from an architectural efficiency perspective. The evaluation focuses on power consumption and communication bandwidth, using computational cost modeling rather than hardware-specific measurements. This approach enables controlled comparison between continuous and event-driven architectures while isolating the impact of communication patterns and update strategies.

To quantitatively evaluate the efficiency and scalability of the proposed architecture, we analyze three representative system configurations that differ in where intelligence is executed and how data are transmitted. These configurations reflect common design choices in healthcare AIoT systems and provide clear baselines for comparison.

4.1 Architectural cases for analysis

We analyze three representative healthcare monitoring architectures: a cloud-continuous baseline, an edge-continuous baseline, and the proposed event-driven edge-intelligent architecture.

4.1.1 Cloud-continuous (B1)

In the cloud-continuous configuration, AIoT sensing nodes continuously transmit raw sensor data to the cloud, where feature extraction and inference are performed centrally. Edge devices perform minimal preprocessing and act primarily as data forwarders. This configuration represents conventional cloud-centric healthcare monitoring systems.

4.1.2 Edge-continuous (B2)

In the edge-continuous configuration, feature extraction and inference are executed locally at the edge for every sensing window. However, intermediate results (e.g., feature vectors or inference outputs) are transmitted continuously to the cloud for aggregation or visualization. This design reduces raw data transmission but still incurs significant computational and communication overhead under long-term monitoring.

4.1.3 Edge-event (proposed, B3)

In the edge-event configuration, the system adopts event-driven edge intelligence. Lightweight feature extraction and trigger evaluation are performed continuously, but full inference and communication are activated only when clinically meaningful events occur. In addition, low-rate periodic summaries are transmitted to maintain longitudinal continuity. This configuration forms the core of the proposed edge-intelligent AIoT digital twin architecture.

4.2 Power consumption modeling and analysis

Power consumption is evaluated as device-side daily energy, defined as the sum of edge computation energy and communication energy. This choice reflects practical deploy-ability, since AIoT sensing nodes are battery-constrained and must operate continuously. Power consumption is estimated using communication activity cost and computational cost.

The total estimated energy consumption E s y s t e m per patient per day is computed as,where E c o m m represents communication-related energy and E c o m p u t e represents computation-related energy.

E s y s t e m = E c o m m + E e d g e − c o m p u t e + E c l o u d − c o m p u t e (1)

First, communication energy E c o m m is modeled as the sum of a fixed per-packet overhead cost and a per-byte transmission cost (Heinzelman et al., 2000; Shnayder et al., 2004; Anastasi et al., 2009) as follows,where N p k t is the number of transmitted packets per day, B t x is the total transmitted data volume in bytes per day. This distinction is essential for accurately characterizing AIoT systems, where energy consumption may be dominated either by large data volumes (cloud-continuous design) or by frequent small transmissions (event-driven design).

E c o m m = E o v e r h e a d + E t x = ϵ p k t · N p k t + ϵ b y t e · B t x (2)

In the model, the term E o v e r h e a d capture the energy for the fixed overhead per packet, independent of payload size, including radio wake-up and synchronization and packet header. This term E o v e r h e a d matters as even a small message costs energy, and frequent small packets can dominate energy consumption. This term explains why event-driven systems must also reduce packet frequency, not just payload size. Meanwhile, the term E t x represents the energy required to physically transmit bits over the radio. This term E t x increases linearly with data volume and dominates when large payloads are transmitted (e.g., raw sensor streams) This term explains why cloud-continuous systems are energy-expensive. Whereas cloud-centric baselines transmit continuous raw or feature-level data, the proposed system transmits only semantic event messages and periodic summaries, which alone contribute to B t x. Even if sent in a small number of packets, the associated bit-level cost is unavoidable.

Second, edge compute energy E e d g e − c o m p u t e scales primarily with execution frequency and duty cycle, rather than model size alone (Banbury et al., 2021; Lane and Georgiev, 2015; Banbury et al., 2020) as follows,where N f e a t, N inf, and N t r i g denote the number of feature extraction executions, full inference executions, respectively, and light weight trigger checks, C f e a t, C inf, and C t r i g represent operation counts per execution of feature extraction and inference, respectively, and ϵ o p e d g e is the energy per operation in edge devices. This model reflects the fact that feature extraction and trigger checks occur continuously, full inference is expensive and should be minimized, and total edge compute energy scales with execution frequency, not model size alone. Edge devices in AIoT healthcare systems (e.g., microcontroller-based sensing nodes) operate under strict energy and resource constraints. For such devices, compute energy is dominated by arithmetic and lightweight control operations executed at fixed voltage and frequency.

E e d g e − c o m p u t e = ϵ o p e d g e · N f e a t · C f e a t + N inf · C inf + N t r i g · C t r i g , (3)

In event-driven mode B3, event triggers are computed from lightweight features and run the full feature pipeline as follows,where N w and N e are the numbers of windows/day and events/day, respectively, and C f e a t , l i g h t ≪ C f e a t , h e a v y. Meanwhile, B2 is paying full feature extraction every window (i.e., N f e a t = N inf = N w), and N t r i g = 0,

E e d g e − c o m p u t e B 3 = ϵ o p e d g e · ( N w · C f e a t , l i g h t + N e · C f e a t , h e a v y + N e · C inf + N w · C t r i g ) , (4)

E e d g e − c o m p u t e B 2 = ϵ o p e d g e · N w · C f e a t , h e a v y + N w · C inf . (5)

Third, cloud compute energy E c l o u d c o m p u t e is modeled as:where N inf ⁡ c l o u d the number of inferences executed centrally, and ϵ o p c l o u d denotes the energy cost per operation in cloud servers. Cloud servers typically offer lower energy per operation due to optimized hardware (i.e., ϵ o p c l o u d ≪ ϵ o p e d g e), and compute energy per inference is lower. However, total cloud energy grows rapidly with patient population and data rate. Inference is executed for every incoming window or data packet, and feature extraction may also occur centrally while cloud resources are shared across users.

E c l o u d − c o m p u t e = ϵ o p c l o u d · N inf ⁡ c l o u d · C inf (6)

Importantly, power consumption is evaluated as device-side daily energy, defined as the sum of edge computation energy and communication energy (i.e., E c o m m + E e d g e − c o m p u t e). This choice reflects practical deploy-ability, since AIoT sensing nodes are battery-constrained and must operate continuously. For cloud-continuous baselines, device compute is assumed negligible and energy is dominated by raw data transmission; for edge-continuous baselines, energy includes per-window feature extraction and inference plus frequent feature uploads; and for the proposed edge-event system, lightweight trigger checks run continuously while full inference and transmission occur only upon detected events, yielding energy that scales with event frequency rather than sampling rate.

4.3 Communication bandwidth analysis

Communication bandwidth usage is analyzed under both edge computation and cloud computation assumptions to highlight how architectural choices affect data transmission requirements. Bandwidth usage per patient per day is computed aswhere m indexes message types (raw sensor data, feature vectors, semantic event updates, periodic summaries), N m is the number of messages of type m transmitted per day, and S m is the size of each message.

B d a y = ∑ m ∈ M N m · S m (7)

Under the edge computation assumption, feature extraction and inference are performed locally on the AIoT device, and communication follows an event-driven strategy. Only semantic event updates triggered by detected falls or instability events, along with low-rate periodic summaries for digital twin continuity, are transmitted. Consequently, message frequency is governed by event occurrence rather than sensor sampling rate, and message sizes remain compact due to semantic representation. High-rate sensor streams, such as IMU data, remain local and do not contribute directly to network traffic, resulting in substantially reduced bandwidth usage compared to continuous transmission architectures.

Under the cloud computation assumption, feature extraction and inference are centralized, requiring edge devices to transmit data necessary for cloud-side processing. This typically involves continuous transmission of raw sensor data or high-rate feature vectors. In such designs, both N m and S m scale with sensor sampling and windowing frequency, leading to significantly higher bandwidth consumption, particularly for high-frequency modalities such as inertial sensing. Communication bandwidth becomes the dominant system resource, constraining scalability and increasing cloud ingestion and storage demands.

When multiple sensing modalities are present, cloud-centric architectures are typically dominated by the highest-rate sensor stream, with additional sensors contributing linearly to bandwidth usage. In contrast, the proposed edge-intelligent, event-driven architecture integrates multimodal data primarily through local inference, while network communication remains bounded by the number of detected events and periodic summaries. This decoupling enables scalable integration of additional sensors without proportional increases in communication bandwidth, supporting population-scale digital twin monitoring.

4.4 Latency and reliability considerations

In addition to energy efficiency and bandwidth consumption, practical healthcare monitoring systems must also consider system responsiveness and reliability. Although the primary focus of this study is architectural efficiency, it is useful to examine how the proposed event-driven edge architecture affects latency and reliability.

Event detection latency in the proposed system is bounded by the sensing window length and the execution time of edge AI inference. Let L denote the sensing window length and T inf the inference execution time. The event detection latency can be approximated as T d e t e c t ≈ L + T inf, because an event must first be captured within a sensing window before inference is executed. Since inference occurs locally on the edge device, communication delay does not affect the initial detection step. Event propagation latency includes both edge inference and communication delay. Once an event is detected, a compact semantic update is transmitted to the digital twin layer and cloud services. The end-to-end event propagation delay can therefore be approximated as, T e v e n t ≈ L + T inf + T c o m m, where T c o m m denotes the transmission delay of the event message. Because semantic event updates are small and infrequent compared to continuous raw data transmission, T c o m m is typically lower and less variable than in cloud-centric architectures.

Reliability in event-driven systems can be characterized by the probability of successful event detection and transmission. Detection reliability depends primarily on sensing coverage and the accuracy of the edge inference model, while communication reliability depends on network packet delivery probability. Since the proposed architecture transmits only sparse semantic event messages rather than continuous sensor streams, the communication channel experiences significantly lower congestion. This property can improve communication reliability by reducing packet loss and retransmission overhead compared to continuous transmission architectures.

Overall, the event-driven architecture provides a bounded latency structure while reducing communication congestion. Detection latency is primarily determined by sensing window length and inference time, and communication latency remains limited due to the compact size and low frequency of semantic event messages. A detailed empirical evaluation of latency and reliability using hardware edge devices and real-world network conditions is left for future work.

5 Simulation

This section presents simulation-based evaluation results for the proposed edge-intelligent, event-driven digital twin system. Simulations are conducted to assess power consumption, communication bandwidth, and scalability under both edge computation and cloud computation assumptions, using a fall detection scenario as the representative healthcare application.

5.1 Simulation settings

This section describes the simulation configuration used to generate the power, bandwidth, sensitivity, and scalability. All parameters are chosen to ensure internal consistency across simulations and to reflect realistic AIoT healthcare monitoring conditions (see Table 2). We compare three architectural baselines: B1, B2, and B3 (see Section 4.1).

TABLE 2. Sensor and processing parameters in simulation.

5.1.1 Application scenario

The simulation models fall detection and near-fall instability monitoring using AIoT-based inertial sensing, a representative scenario for long-term remote healthcare monitoring of elderly individuals and rehabilitation patients. Such monitoring is commonly deployed in assisted living environments, home-based elderly care, and post-discharge rehabilitation programs, where continuous observation of patient mobility and fall risk is clinically important. In these settings, wearable inertial sensors allow early detection of instability events that may indicate deterioration in balance, medication side effects, or rehabilitation progress. This use case captures a key characteristic of real-world healthcare monitoring: continuous sensing with sparse clinically meaningful events. Each simulated patient is equipped with a single AIoT sensing node containing a 3-axis accelerometer and a 3-axis gyroscope (six channels total). Sensor activity represents daily movement patterns interspersed with rare fall events and more frequent instability episodes, which trigger event-driven processing in the proposed architecture.

5.1.2 Sensor sampling and windowing

IMU data are sampled at 100 Hz, consistent with prior fall detection and activity recognition studies. Sensor streams are segmented into overlapping windows of length L ∈ {1, 3, 5, 7.5, 10} seconds, with a fixed stride of L / 2. These window lengths are used directly in the energy vs. window size and bandwidth vs. window size results, as well as in sensitivity analyses evaluating the effect of windowing on computation and communication costs.

5.1.3 Event generation model

Event-driven behavior is modeled using a stochastic trigger process designed to maintain a constant expected event rate λ per unit time across different window sizes. A baseline trigger probability of p = 0.005 per window with a 0.5 s stride corresponds to an average rate of approximately 0.01 events/second. For other window lengths, the per-window trigger probability is scaled proportionally to the stride. To avoid repeated detections during sustained abnormal conditions, a cooldown period of 60 s is enforced following each detected event, implemented as a stride-dependent number of suppressed windows. In addition to event-triggered updates, the system transmits periodic summary messages every 10 min (144 summaries per day) to preserve digital twin continuity under long-term monitoring.

Event frequency sensitivity results are obtained by sweeping the event rate over λ ∈ {0.001, 0.005, 0.01, 0.05, 0.1} events/sec, which directly corresponds to the energy vs. event frequency and bandwidth vs. event frequency results.

5.1.4 Power consumption modeling

Device-side energy consumption is computed as the sum of edge computation energy and communication energy, based on operation counts and transmitted data volume. Cloud-side computation is excluded from device energy accounting, consistent with the reported results.

5.1.5 Lightweight feature sensitivity analysis

To evaluate robustness of the proposed architecture, a sensitivity analysis varies the cost of lightweight monitoring features relative to full feature extraction. This is parameterized by k = C f e a t , l i g h t / C f e a t , h e a v y, where k ∈ {0.01, 0.05, 0.1, 0.5, 1}. Results are reported for window sizes of 1, 5, and 10 s, corresponding to the nested plots comparing total energy consumption as a function of k.

5.1.6 Population-scale simulation

Scalability is evaluated by extrapolating per-patient simulation results to populations of 100, 300, 1,000, 3,000, and 10,000 patients, assuming independent operation of AIoT sensing nodes. Aggregate device-side energy consumption and communication bandwidth are computed linearly from per-patient values, yielding the population-scale.

The population-scale analysis assumes independent operation of patient-side AIoT sensing nodes and linear aggregation of per-device energy and bandwidth metrics. This abstraction is intended to isolate architectural differences between continuous and event-driven monitoring designs. The present model therefore does not explicitly represent shared cloud resource contention, network queuing effects, or multi-tenant digital twin service constraints. These distributed system factors are important considerations for large-scale deployments and represent a natural direction for future system-level modeling and evaluation.

All modeling assumptions, parameter ranges, and simulation configurations are explicitly defined in this section and summarized in Table 2 to ensure reproducibility. This formulation also enables analytical extension of the evaluation framework to other AIoT healthcare monitoring scenarios.

5.2 Simulation results

This section presents the simulation results comparing the three architectures in terms of device-side energy consumption, communication bandwidth, sensitivity to system parameters, and scalability across patient populations.

5.2.1 Power consumption results

Figure 5 compares device-side power consumption across the three architectures as a function of event frequency and window size, respectively. In the baseline B1, total energy usage is independent of event frequency (in the left figure) and window size (in the right figure). In the baseline B2, total energy usage is independent of event frequency, and the window size further reduces energy consumption by decreasing the number of trigger evaluations per day due to fewer feature uploads and inference executions. The proposed architecture B3 consistently achieves the lowest power consumption in both cases by executing lightweight processing continuously while activating full inference only upon clinically meaningful events. As a result, energy consumption in B3 scales with event frequency, remaining substantially lower than B1 and B2 under sparse event conditions. Thus, these results demonstrate that event-driven edge intelligence effectively limits device-side power consumption, enabling energy-efficient, long-term AIoT healthcare monitoring compared to continuous cloud- or edge-centric.

Figure 6 illustrates the impact of lightweight feature extraction cost on device-side energy consumption. The left figure compares total daily energy consumption of the baseline B2 and the architecture B3 as a function of the cost ratio k. Energy consumption in B2 remains constant across all values of k, as full feature extraction and inference are executed at every window regardless of trigger cost. In contrast, B3 exhibits a gradual increase in energy consumption as k increases, reflecting higher per-window monitoring cost; however, B3 consistently remains below B2 across the entire range of k, confirming the robustness of event-driven inference gating even when lightweight features approach the cost of full feature extraction. The right figure shows total energy consumption of B3 for different window sizes (1, 5, 10 s). For all values of k, longer window sizes result in lower energy consumption due to fewer trigger evaluations per day. While increasing k raises energy consumption for all window sizes, the relative ordering remains consistent, and B3 maintains low energy usage across a wide range of monitoring resolutions. Thus, these results demonstrate that the proposed architecture is resilient to variations in trigger complexity and windowing choices, and that its energy efficiency is primarily driven by event-gated inference rather than lightweight feature cost alone.

5.2.2 Communication bandwidth results

Figure 7 illustrates the communication bandwidth requirements of the three architectures as a function of event frequency (in the left figure) and window length (in the right figure). The results highlight fundamental differences in how bandwidth scales with sensing and inference strategies.

In the baseline B1, bandwidth consumption constant across all event frequencies (in the left figure) and window sizes (in the right figure), dominated by continuous raw IMU streaming and remains. Because all sensor data are transmitted regardless of clinical relevance, bandwidth usage is strictly tied to the sensing rate rather than system dynamics. Similarly, the baseline B2 exhibits high and nearly constant bandwidth consumption (in the left figure) due to frequent per-window feature uploads, with only modest reduction as window length increases and the number of windows per day decreases (in the right figure).

In contrast, the proposed architecture B3 shows bandwidth consumption that scales primarily with event frequency (in the left figure) rather than sensing frequency (in the right figure). As shown in the left figure, increasing event frequency leads to higher bandwidth usage due to more frequent semantic event transmissions, while remaining orders of magnitude lower than B1 and B2 under sparse-event conditions. The right figure demonstrates that increasing window length further reduces bandwidth consumption in B3 by lowering the number of trigger evaluations and event opportunities per day, while periodic summaries contribute a fixed, low-rate baseline.

Thus, these results demonstrate that event-driven communication effectively decouples bandwidth requirements from raw sensor sampling rates, enabling scalable, long-term digital twin monitoring. Compared to continuous cloud- or edge-centric designs, the proposed architecture significantly reduces communication overhead while preserving timely delivery of clinically meaningful information.

5.2.3 Scalability results

Figure 8 presents the scalability analysis of the three architectures in terms of aggregate device-side energy consumption (in the left figure) and aggregate communication bandwidth (in the right figure) as the number of monitored patients increases. The results are obtained by linear extrapolation of per-patient simulation outcomes, reflecting independent operation of AIoT sensing nodes across patients.

In the left figure, the baselines B1 and B2 scale linearly with the number of patients for aggregate device energy. As patient count increases, these continuous architectures rapidly accumulate high total energy consumption, making large-scale deployment increasingly inefficient. In contrast, the proposed architecture B3 exhibits substantially lower aggregate energy across all population sizes. Because full inference and communication are triggered only by clinically meaningful events, total energy in B3 scales with event frequency rather than raw sensing rate, resulting in orders-of-magnitude reductions at feasible scale.

In the right figure, a similar trend is observed for aggregate bandwidth consumption. In B1 and B2, bandwidth increases linearly with the number of patients at a high slope, driven by continuous raw data or feature transmission. At large scales, this leads to extremely high aggregate data volumes that place significant burden on network infrastructure and cloud ingestion pipelines. By contrast, B3 maintains very low aggregate bandwidth even as the number of patients grows, since communication is limited to sparse semantic event updates and low-rate summaries. Consequently, the slope of bandwidth growth in B3 is significantly lower, enabling scalable monitoring of large cohorts.

It should be noted that the scalability analysis assumes independent sensing nodes and linear aggregation of device-side metrics. While this abstraction enables controlled comparison between architectural designs, real-world deployments may introduce additional constraints such as shared cloud processing resources, network congestion, and multi-tenant digital twin management.

Thus, these results demonstrate that scalability in digital twin healthcare systems is fundamentally constrained by continuous communication and inference. By aligning computation and communication with clinically meaningful events, the proposed event-driven edge architecture enables population-scale deployment with manageable energy and bandwidth requirements, supporting practical and sustainable large-scale healthcare monitoring.

6 Discussion

The simulation results suggest that the proposed event-driven, edge-intelligent AIoT digital twin architecture substantially improves energy efficiency, communication bandwidth efficiency, and scalability for practical deployment of continuous healthcare monitoring systems. Across all evaluated scenarios, the key determinant of system efficiency is whether computation and communication are driven by clinically meaningful events rather than continuous data streams.

Power consumption analysis shows that cloud-continuous (B1) and edge-continuous (B2) architectures incur high and largely invariant energy costs due to sustained communication and/or continuous inference. In contrast, the proposed event-driven architecture (B3) indicates potential for orders-of-magnitude lower device-side energy consumption by decoupling inference from sensing frequency. Energy in B3 scales with event frequency and window size, remaining low under sparse-event conditions typical of real-world healthcare monitoring. Sensitivity analysis further confirms that this advantage is architectural: even when lightweight monitoring approaches the cost of full feature extraction, event-gated inference preserves energy efficiency.

Communication bandwidth exhibits a similar pattern. Continuous architectures generate large, constant data streams proportional to sensor sampling rates, resulting in high and inflexible bandwidth requirements. By transmitting only semantic event updates and low-rate summaries, the proposed system limits communication to clinically relevant information, enabling bandwidth usage that scales with event frequency rather than raw data volume. This design simultaneously reduces infrastructure load and limits exposure of sensitive raw data, aligning efficiency with privacy preservation.

Scalability analysis highlights the population-level implications of these results. Continuous cloud- and edge-centric systems accumulate prohibitive aggregate energy and bandwidth costs as patient numbers grow. In contrast, the proposed architecture scales linearly with patient count but with a substantially lower slope, governed by event frequency. This enables population-scale digital twin monitoring with manageable device, network, and cloud resource requirements.

Thus, these findings suggest that event-driven edge intelligence can serve as an important architectural principle for scalable, privacy-aware digital twin healthcare systems. By aligning intelligence with clinical relevance rather than data volume, the proposed approach provides a practical pathway toward long-term, large-scale deployment.

Although this study focuses primarily on architectural efficiency in terms of power consumption and communication bandwidth, system responsiveness and reliability are also important considerations in practical deployments. In the proposed architecture, event detection latency is bounded by the sensing window length and local inference time because events are detected directly at the edge rather than in the cloud. This provides a predictable latency bound compared to cloud-centric architectures where network transmission delays may dominate. In addition, the event-driven communication model reduces network congestion by transmitting only sparse semantic updates rather than continuous sensor streams, which can improve communication reliability and reduce packet loss. Future work will evaluate these latency and reliability characteristics through hardware experiments and real-world deployments.

In practical healthcare deployments, the proposed event-driven digital twin system can support several clinical monitoring workflows. For example, in home-based elderly monitoring or assisted living environments, detected instability or fall events can trigger automated alerts to caregivers or remote monitoring platforms. In rehabilitation contexts, repeated instability events may indicate deterioration in balance or incomplete recovery and can prompt therapist review, adaptive exercise scheduling, or medication reassessment. Because the system transmits compact semantic state updates rather than continuous sensor streams, clinicians can focus on clinically meaningful events while maintaining longitudinal visibility into patient mobility trends through the digital twin representation.

The scalability analysis presented in this study adopts a simplified linear aggregation model to isolate the architectural impact of event-driven edge intelligence. In large-scale deployments, additional factors such as cloud resource sharing, network queuing effects, and multi-tenant digital twin management may influence system performance. Modeling these distributed system dynamics represents an important direction for future research. These factors may introduce non-linear scaling behavior in practical deployments. While the current analysis assumes independent nodes and linear aggregation, modeling such effects represents an important direction for future distributed systems evaluation.

These findings highlight the importance of architectural design choices in enabling scalable and sustainable digital twin healthcare systems beyond purely algorithmic improvements.

This study focuses on architectural modeling and system-level analysis rather than hardware-specific validation. The objective is to provide a principled evaluation of architectural trade-offs under realistic constraints. The current study is based on simulation and analytical modeling, and does not explicitly capture hardware-level overhead such as memory access cost, device-specific energy consumption, or network protocol inefficiencies. In addition, sensor noise and model uncertainty may introduce false positives and false negatives in event detection, affecting system responsiveness and digital twin updates. While event-driven updates and periodic summaries help mitigate these effects, detailed evaluation under noisy conditions and real hardware platforms remains an important direction for future work. Such validation through hardware prototyping and real-world deployment remains an important direction for future work.

In addition to efficiency and system robustness, the proposed architecture also provides inherent privacy benefits through data minimization. In event-driven architectures, only semantic state updates are transmitted instead of raw sensor data, significantly reducing the information content exposed to external systems. This reduction can be interpreted as a decrease in mutual information between raw sensor signals and transmitted data. This perspective provides a proxy interpretation of privacy preservation at the system level, even in the absence of formal guarantees. Incorporating formal privacy metrics such as mutual information or differential privacy is an important direction for future work.

7 Conclusion

This paper presented an event-driven, edge-intelligent AIoT digital twin architecture for scalable and privacy-aware healthcare monitoring. By shifting intelligence from continuous cloud-centric processing to event-gated edge execution, the proposed system aligns computation and communication with clinically meaningful changes rather than raw sensor data streams.

Through comprehensive computational simulations, the proposed architecture significantly reduces device-side energy consumption and communication bandwidth compared to continuous cloud- and edge-centric baselines. Power and bandwidth scale with event frequency rather than sensing frequency, enabling efficient long-term operation under realistic monitoring conditions. Population-scale analysis further showed that aggregate system costs grow at a substantially lower rate, supporting practical deployment across large patient cohorts.

Beyond efficiency gains, the proposed design reframes privacy as a system property enabled through data minimization and semantic digital twin updates, rather than as a post hoc protection mechanism. These results indicate that event-driven edge intelligence is a foundational architectural principle for future digital twin healthcare systems, enabling scalable, trustworthy, and sustainable monitoring in real-world environments.

Future work will include hardware validation and extend this framework in several directions. First, hardware deployment on resource-constrained edge AIoT devices will be conducted to validate the energy and communication models under real operating conditions. Second, the architecture will be evaluated using real sensor traces collected from wearable inertial devices to assess event detection performance and system responsiveness. Third, pilot deployments in practical healthcare contexts such as assisted living environments or remote rehabilitation monitoring will be explored to investigate integration with clinical workflows and care management platforms.

This work positions event-driven edge intelligence as a key enabler for bridging the gap between digital twin modeling and scalable AIoT healthcare deployment.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Generative artificial intelligence tools were used to assist with language refinement and preliminary code drafting during manuscript preparation. The authors take full responsibility for the originality, accuracy, and integrity of all technical content and conclusions.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Weiwei Jiang, Beijing University of Posts and Telecommunications (BUPT), China

Reviewed by: Lidia Bajenaru, National Institute for Research and Development in Informatics, Romania

Bhanuprakash Madupati, Governors State University, United States