Edge-Enhanced Federated Optimization for Real-Time Silver-Haired Whirlwind Trip

Abstract

We propose an edge-enhanced federated learning framework for real-time itinerary optimization in elderly oriented adventure tourism, addressing the critical need for adaptive scheduling that balances activity intensity with health constraints. The system integrates lightweight convolutional neural networks with a priority-based scheduling algorithm, processing participant profiles and real-time biometric data through a decentralized computation model to enable dynamic adjustments. A modified Hungarian algorithm incorporates physical exertion scores, temporal proximity weights, and health risk factors, then optimizes activity assignments while respecting physiological recovery requirements. The federated learning architecture operates across distributed edge nodes, preserving data privacy through localized model training and periodic global aggregation. Furthermore, the framework interfaces with transportation systems and medical monitoring infrastructure, automatically triggering itinerary modifications when vital sign anomalies exceed adaptive thresholds. Implemented on NVIDIA Jetson AGX Orin modules, the system achieves 300 ms end-to-end latency for real-time schedule updates, meeting stringent safety requirements for elderly participants. The proposed method demonstrates significant improvements over conventional itinerary planners through its edge computing efficiency and personalized adaptation capabilities, particularly in handling the latency-sensitive demands of intensive tourism scenarios. Experimental results show robust performance across diverse participant profiles and activity types, confirming the system’s practical viability for real-world deployment in elderly adventure tourism operations.

1. Introduction

With the intensification of global population aging trend, the proportion of the elderly population in the tourism market has gradually increased, forming a segmented market with huge potential. However, there are significant differences in the tourism demands of this group from traditional tourism models. The physical condition, health needs and high attention to safety of the elderly have made travel service providers revisit and optimize existing tourism products and services. From a tourism management perspective, this demographic shift represents both a challenge and opportunity for the industry. Elderly tourists increasingly seek experiences that balance adventure with safety, creating demand for what might be termed “calculated risk” tourism experiences. This trend is particularly evident in East Asian markets where so-called “Silver-Haired Special Forces” groups pursue intensive itineraries traditionally associated with younger backpackers. The tourism industry must respond to this demand with innovative approaches that address the unique physiological constraints of older adults while maintaining the excitement and engagement of adventure tourism. Such approaches require seamless integration of real-time health monitoring, dynamic itinerary adjustment, and respectful privacy preservation—needs that existing tourism planning systems are ill-equipped to handle. In recent years, with the rapid development of technology, especially the rise of edge computing and federated learning technologies, new ideas and methods have been provided to solve this problem.

The rapid demographic shift toward aging populations has created new opportunities and challenges in the tourism industry, particularly for specialized segments like the so-called “Silver-Haired Special Forces”—elderly tourists pursuing intensive, adventure-oriented travel experiences. Unlike conventional elderly tourism focused on leisure and relaxation, this emerging market segment demands sophisticated itinerary planning systems capable of balancing high-intensity activities with real-time health monitoring and adaptive scheduling (La Rocca & Fistola, 2018). Traditional tourism recommendation systems often fail to address the unique requirements of this demographic, as they typically prioritize static preferences over dynamic physiological constraints (Lucas et al., 2013). The impetus for this research emerged from empirical observations of the growing “Silver-Haired Special Forces” phenomenon in East Asian tourism markets, where elderly tourists increasingly pursue intensive, adventure-oriented experiences traditionally associated with younger travelers. This trend, coupled with advancements in edge computing and federated learning technologies, revealed a critical need for systems capable of supporting such demanding tourism modalities while prioritizing health constraints and privacy concerns. This research was specifically motivated by the absence of integrated solutions that address both the dynamic health monitoring requirements and the personalized activity scheduling needs of this emerging tourist segment.

By sinking computing power from the cloud to the edge of the network, edge computing can effectively reduce data transmission delays and achieve rapid processing and response to real-time data. This feature is particularly important for travel scenarios for the elderly that require real-time health monitoring and dynamic itinerary adjustment. Despite these technological advancements, significant gaps remain in both literature and practice. First, current systems typically address either health monitoring or activity recommendation in isolation, lacking integrated approaches that dynamically adjust itineraries based on real-time physiological data. Second, privacy concerns often limit the practical implementation of centralized data processing for sensitive health information. Third, existing edge computing solutions lack the computational efficiency needed for complex, multi-objective optimization in resource-constrained tourism environments. Our work directly addresses these gaps by proposing a unified framework that combines health-aware scheduling with privacy-preserving federated learning on efficient edge devices. However, most edge computing applications currently focus on only a single function, such as activity recommendation or health monitoring, and fail to organically combine the two. This separation of treatment leads to insufficient coordination between itinerary planning and participants’ physical condition, which makes it difficult to meet the health and safety needs of the elderly in high-intensity tourism activities. In contrast, our proposed framework achieves tight integration by processing real-time biometric streams (e.g., heart rate, SpO₂) through a decentralized edge architecture, enabling dynamic health-aware scheduling with end-to-end latency under 300 ms for instantaneous feedback and intervention.

This work addresses three fundamental problems in elderly adventure tourism: the separation between health monitoring and activity recommendation, privacy concerns in centralized data processing, and computational limitations of existing edge systems. Our solution integrates lightweight convolutional networks with a modified Hungarian algorithm within a federated learning framework, enabling real-time itinerary optimization that simultaneously preserves privacy, ensures health safety, and maintains computational efficiency. This integrated approach represents a significant advancement over conventional systems that address these challenges in isolation.

In addition, most existing tourism recommendation systems provide activity recommendations based on static preferences, neglecting the dynamic changes in physiological status and health constraints of the elderly during tourism. This single recommendation model cannot adapt to the complexity and diversity of the elderly’s tourism market, nor can it meet the elderly’s expectations for personalized and safe travel experiences. At the same time, current edge computing systems have bottlenecks in computing efficiency, making it difficult to efficiently deploy on resource-constrained mobile travel equipment. The existence of these problems has limited the further development of the tourism market for the elderly and also highlighted the importance and urgency of developing a new tourism planning system.

This study contributes to the tourism management literature by addressing the critical gap between personalized itinerary planning and real-time health adaptability for elderly tourists. While prior research has largely focused on static recommendation systems or isolated health monitoring, our integrated framework leverages edge computing and federated learning to enable dynamic, safe, and private itinerary optimization. This approach not only enhances tourist safety and satisfaction but also offers a scalable model for managing specialized tourism segments in an ethically responsible manner, thereby advancing both theoretical and practical understanding of technology-enabled tourism management for aging populations. The primary objective of this work is to develop a real-time itinerary optimization system specifically designed for elderly adventure tourism that addresses three critical challenges: the separation between health monitoring and activity recommendation, privacy concerns in centralized data processing, and computational limitations of existing edge systems. Our framework aims to provide personalized, health-aware scheduling while maintaining strict privacy guarantees through federated learning and efficient edge computing.

Recent advances in edge computing and federated learning offer promising solutions to these challenges. Federated edge architectures enable distributed processing of sensitive health data while maintaining low-latency responses critical for real-time itinerary adjustments (Li et al., 2023). However, existing approaches often treat activity recommendation and health monitoring as separate processes, leading to suboptimal coordination between planned activities and participants’ physical conditions (Movono et al., 2022; Kim et al., 2021). Moreover, most current systems lack the computational efficiency required for deployment on resource-constrained edge devices commonly used in mobile tourism scenarios (Deng et al., 2021).

We present a novel framework that integrates three key innovations: (1) a lightweight convolutional neural network architecture optimized for edge deployment, capable of processing both activity preferences and real-time health data; (2) a priority-based scheduling algorithm that dynamically adjusts itinerary plans based on continuously updated physiological parameters; and (3) a federated learning model that preserves participant privacy while enabling collaborative improvement of recommendation accuracy across multiple tourism operators. The system specifically addresses the challenges of intensive elderly tourism by incorporating exertion scoring, recovery period estimation, and emergency response triggering mechanisms into its core decision-making processes.

The proposed approach differs fundamentally from prior work in several aspects. First, it moves beyond simple activity recommendation to provide true real-time itinerary optimization, considering not just what activities are desirable but when they should occur based on physiological readiness (Shoval et al., 2018). Second, the system implements a novel hybrid computation model where some processing occurs on personal edge devices (e.g., smartwatches) while more complex optimizations run on nearby infrastructure nodes, achieving an optimal balance between responsiveness and computational capability (Rahimi et al., 2021). Third, the framework introduces an adaptive threshold mechanism that automatically adjusts activity intensity limits based on individual response patterns over time, moving beyond static safety margins common in existing systems (van Rossum et al., 2021).

Our experimental evaluation demonstrates significant improvements over baseline methods in both recommendation quality and system responsiveness. The lightweight CNN achieves 92.3% accuracy in activity preference prediction while requiring only 15 MB of memory, making it suitable for deployment on edge devices. The priority-based scheduler reduces emergency intervention frequency by 38% compared to conventional approaches, while maintaining high participant satisfaction scores. The federated learning component shows particular effectiveness in cold-start scenarios, reaching 85% of peak accuracy with just 20% of the training data required by centralized models.

The remainder of this paper is organized as follows: Section 2 reviews related work in elderly tourism systems, edge computing applications, and federated learning approaches. Section 3 provides necessary background on the key technologies and concepts. Section 4 details our proposed framework and its components. Section 5 and Section 6 present our experimental methodology and results. Finally, Section 7 discusses implications and future research directions, followed by conclusions in Section 8.

2. Related Work

Recent advancements in edge computing and federated learning have opened new possibilities for real-time itinerary optimization in specialized tourism domains. This section examines three key research areas relevant to our work: (1) edge computing architectures for real-time systems, (2) health-aware activity recommendation, and (3) federated learning applications in tourism.

2.1. Edge Computing for Real-Time Systems

Mobile edge computing has emerged as a critical enabler for latency-sensitive applications, particularly in scenarios requiring immediate response to sensor data (Dong et al., 2024; Y. Khan et al., 2024). Prior work has demonstrated the effectiveness of edge architectures in processing health monitoring data with sub-second latency (Gupta et al., 2023). However, existing systems often focus on single-modality processing (e.g., either activity recognition or vital sign monitoring) rather than the integrated multimodal analysis required for comprehensive itinerary optimization (Singh & Chatterjee, 2022). Our framework addresses this limitation through a unified feature extraction pipeline that jointly processes participant profiles and biometric data streams.

Recent advancements in edge computing have demonstrated significant potential for latency-sensitive health monitoring applications (Duan et al., 2022). However, many existing systems focus on single-modality processing rather than integrated multimodal analysis (X. Chen et al., 2024). Our framework addresses this gap through a unified processing pipeline that jointly handles participant profiles and real-time biometric data.

In addition to addressing the single-modality limitation, our framework also incorporates advanced machine learning algorithms to enhance the accuracy and efficiency of real-time decision-making. By leveraging edge computing, we are able to process data locally at the edge of the network, significantly reducing the latency associated with data transmission to remote servers. This is particularly crucial for applications such as emergency response systems, where every millisecond counts. For example, in a scenario involving a patient with a sudden health crisis, our framework can quickly analyze the biometric data from wearable devices and cross-reference it with the patient’s medical history stored in the participant profile. This integrated approach allows for the rapid identification of potential health risks and the generation of personalized recommendations for immediate action our framework is designed to be highly scalable and adaptable to different types of sensor data. As the Internet of Things (IoT) continues to expand, the volume and variety of data generated by sensors are increasing exponentially. Our edge computing architecture can dynamically adjust its processing capabilities to handle this growing data load without compromising on latency. This is achieved through a combination of hardware acceleration techniques and software optimization strategies. For instance, we utilize specialized hardware accelerators such as FPGAs and GPUs to offload computationally intensive tasks, while our software algorithms are optimized for parallel processing and efficient memory management.

Another key advantage of our framework is its ability to provide robust security and privacy protection. Edge computing allows sensitive data to be processed locally, minimizing the exposure of personal information to potential security breaches during transmission. Additionally, our framework employs state-of-the-art encryption and access control mechanisms to ensure that only authorized personnel can access and manipulate the data. This is particularly important in healthcare applications, where patient privacy is of utmost concern. By integrating these security features into our edge computing architecture, we can build a trustworthy system that meets the stringent requirements of modern real-time applications.

2.2. Health-Aware Activity Recommendation

Health monitoring systems have increasingly incorporated machine learning techniques to predict user states and recommend appropriate actions (Alnaim & Alwakeel, 2023). While these approaches show promise in controlled environments, they often fail to account for the dynamic nature of tourism activities and their varying physical demands (Guo & Li, 2024; A. Khan et al., 2023). The closest existing work employs priority-based scheduling for medical applications, but lacks the tourism-specific considerations of activity sequencing and group coordination (Vishal et al., 2023). Our modified Hungarian algorithm extends these concepts by incorporating temporal proximity constraints and group-level optimization objectives.

To address these gaps, our system integrates real-time health data with contextual information about the tourism environment. By analyzing biometric data such as heart rate, step count, and fatigue levels, our algorithm can dynamically adjust activity recommendations to ensure they are both health-conscious and enjoyable. For example, if a user’s heart rate exceeds a safe threshold, the system may suggest a more relaxing activity or a break. Additionally, our approach considers group dynamics, ensuring that activities are synchronized for optimal group enjoyment while accommodating individual health needs. This holistic approach not only enhances user experience but also promotes healthier travel behaviors.

This holistic approach aligns with the integrated tourism experience framework proposed by X. Chen et al, which emphasizes the interconnectedness of physiological, psychological, and environmental factors in tourism satisfaction (X. Chen et al., 2025). Our system embodies this holistic perspective by simultaneously considering multiple dimensions of the tourist experience: real-time physiological states through biometric monitoring, personal preferences through federated learning, environmental conditions through external data feeds, and social dynamics through group coordination algorithms. Unlike reductionist approaches that optimize single dimensions in isolation, our holistic framework recognizes that elderly adventure tourism requires balancing potentially competing objectives—maximizing enjoyment while minimizing health risks, preserving privacy while enabling personalization, and maintaining individual agency while ensuring group cohesion. This multi-dimensional optimization represents a significant advancement over conventional systems that address these aspects separately, often leading to suboptimal trade-offs.

2.3. Federated Learning in Tourism Applications

Federated learning has gained traction in privacy-sensitive domains, including healthcare and location-based services (Yu & Li, 2021). Previous tourism recommendation systems have used centralized learning approaches, requiring sensitive data to be transmitted to cloud servers for processing (Sahu et al., 2025). The present study differs by implementing a hierarchical federated learning scheme where personal preferences are learned locally on edge devices, while broader activity patterns are aggregated across tourism operators without sharing raw participant data.

The proposed system advances beyond existing approaches through its tight integration of these three research directions. Unlike prior edge computing applications that focus solely on either health monitoring or activity recommendation, our framework jointly optimizes both aspects in real-time. The federated learning component enables continuous personalization while preserving privacy—a critical requirement for elderly tourism applications. Furthermore, the system’s hybrid computation model represents a novel approach to balancing processing demands between personal devices and infrastructure nodes, addressing a key limitation in current edge-based tourism systems.

In addition to these innovations, our system also incorporates a robust validation mechanism to ensure the accuracy and reliability of recommendations. By periodically updating the federated learning model with anonymized data from multiple sources, we can capture evolving trends in tourism preferences while maintaining user privacy. This dynamic updating process allows the system to adapt to changing user behaviors and environmental conditions, such as seasonal variations in popular activities or the introduction of new tourist attractions. Moreover, our approach ensures that the computational load is distributed efficiently, reducing the risk of overloading personal devices or network infrastructure. This balance is crucial for maintaining a seamless user experience, especially in resource-constrained environments. Overall, our system represents a significant step forward in creating privacy-preserving, personalized, and efficient tourism applications.

3. Background and Preliminaries

To establish the technical foundation for our proposed framework, this section introduces three key concepts: (1) the role of federated edge computing in latency-sensitive applications, (2) optimization techniques for dynamic resource allocation, and (3) wearable sensor data characteristics relevant to real-time monitoring. These components form the basis for understanding our system’s design decisions and performance characteristics.

3.1. Federated Edge Computing in Latency-Sensitive Applications

Distributed computing paradigms face fundamental trade-offs between response latency and computational capability, particularly when processing time-sensitive data streams from mobile users (Mao et al., 2017). The end-to-end delay in such systems can be decomposed into three primary components:

$$Latency=t_{\mathrm{transmission}}+t_{\mathrm{processing}}+t_{\mathrm{synchronization}}$$

(1)

Here, $t_{\mathrm{transmission}}$ denotes the time for data transfer between devices and edge nodes, typically ranging from 20–100 ms in BLE 5.2 networks; $t_{\mathrm{processing}}$ represents the computation time on edge devices, which we optimize using lightweight CNNs; and $t_{\mathrm{synchronization}}$ refers to the delay in model aggregation across nodes, which we minimize through periodic federated updates.

Federated edge architectures address these challenges by distributing computation across network hierarchy levels while maintaining data privacy through localized model training (Akter et al., 2022). In elderly tourism scenarios, this approach becomes crucial as it enables real-time processing of sensitive health data without requiring centralized collection. The edge layer typically consists of gateway devices with moderate computational resources (e.g., 4–8 CPU cores, 8–16 GB RAM), positioned within 1–2 network hops from end-user devices (Ghobaei-Arani et al., 2019).

3.2. Optimization in Dynamic Resource Allocation

Activity scheduling in our context represents a constrained optimization problem where the system must maximize participant satisfaction while respecting physiological limits. Formally, this can be expressed as:

$$\underset{\mathbf{X}}{\mathrm{m}\mathrm{i}\mathrm{n}}f\left(\mathbf{X}\right)\hspace{1em}\mathrm{s}.\mathrm{t}.\hspace{1em}g\left(\mathbf{X}\right)\le 0$$

(2)

Here, $\underset{\mathbf{X}}{\mathrm{m}\mathrm{i}\mathrm{n}}f\left(\mathbf{X}\right)$ denotes the decision vector comprising activity assignments and timing variables; $f\left(\mathbf{X}\right)$ is the objective function combining preference satisfaction and health safety; and $g\left(\mathbf{X}\right)$ encodes constraints such as recovery periods and mobility limits. For instance, $g\left(\mathbf{X}\right)\le 0$ may represent that the heart rate of participant must not exceed a safe threshold.

Where $\mathbf{X}$ represents the decision variables (activity assignments, timing, etc.), $f\left(\cdot \right)$ captures the optimization objectives (preference matching, health safety), and $g\left(\cdot \right)$ encodes constraints (recovery periods, mobility limits). The combinatorial nature of this problem necessitates efficient approximation algorithms, as exact solutions become computationally intractable for realistic problem sizes (Jiang et al., 2020). Our modified Hungarian algorithm builds upon these foundations by incorporating temporal proximity weights and adaptive health risk factors.

3.3. Wearable Sensor Data for Real-Time Monitoring

Wearable devices in elderly tourism scenarios must maintain reliable data collection despite motion artifacts and environmental variability. The signal-to-noise ratio (SNR) serves as a key metric for assessing data quality:

$$SNR=10{\mathrm{l}\mathrm{o}\mathrm{g}}_{10}\left({\displaystyle \frac{{\sigma }_{\mathrm{signal}}^2}{{\sigma }_{\mathrm{noise}}^2}}\right)$$

(3)

Practical deployments typically require SNR values exceeding 15 dB for reliable vital sign monitoring (B. Hu et al., 2015). Modern devices achieve this through a combination of hardware filtering and algorithmic noise suppression, with advanced systems employing adaptive baselining to account for individual physiological variations (Zhang et al., 2025). These capabilities enable our framework to maintain accurate real-time assessments of participant states even during high-intensity activities.

The integration of these three components—federated edge infrastructure, constrained optimization, and robust sensor data processing—forms the technical basis for our proposed itinerary planning system. Each element addresses specific challenges in elderly adventure tourism while working synergistically to enable real-time, health-aware scheduling. The following section will detail how these foundational concepts are implemented in our framework’s architecture.

This work is theoretically grounded in three interconnected domains: distributed cognitive systems theory, which informs our hierarchical edge computing architecture; constraint satisfaction theory, which underpins our modified Hungarian algorithm for multi-objective itinerary optimization; and privacy-preserving machine learning theory, which guides our federated learning implementation with differential privacy guarantees. These theoretical foundations provide the conceptual framework for addressing the complex interplay between real-time health monitoring, personalized activity recommendation, and privacy preservation in elderly adventure tourism scenarios.

This research is grounded in three core hypotheses: (1) The integration of edge computing and federated learning can enable real-time, privacy-preserving itinerary optimization for elderly adventure tourism; (2) Lightweight convolutional neural networks combined with a modified Hungarian algorithm can efficiently handle multi-objective scheduling under dynamic physiological constraints; (3) Joint modeling of real-time biometric data and activity preferences significantly improves both recommendation relevance and health safety. Methodologically, this work follows a design-science research paradigm, constructing a functional prototype and evaluating its performance through rigorous empirical testing to validate these hypotheses.

4. Federated Edge Computing for Real-Time Itinerary Planning

Schema 1. Federated Edge Computing Architecture. The proposed framework establishes a three-tier computational architecture that distributes processing across personal edge devices, local edge servers, and regional aggregation nodes. This hierarchical design enables real-time itinerary optimization while maintaining strict privacy guarantees through federated learning protocols.

The proposed framework establishes a three-tier computational architecture that distributes processing across personal edge devices, local edge servers, and regional aggregation nodes. This hierarchical design enables real-time itinerary optimization while maintaining strict privacy guarantees through federated learning protocols. The system processes multiple data streams including participant profiles, real-time biometrics, and environmental conditions to generate dynamically adjusted activity schedules.

4.1. Decentralized Edge Computing Architecture for Tourism Optimization

The edge computing layer consists of NVIDIA Jetson AGX Orin modules deployed at tourism sites, each serving 8–12 participants simultaneously. These nodes execute the core itinerary optimization algorithms while interfacing with participants’ wearable devices through Bluetooth Low Energy (BLE) 5.2 connections. The architecture minimizes cloud dependency by processing 92% of computation at the edge layer, reducing end-to-end latency to under 300 ms as required for real-time health interventions. The selection of computational methods was strategically guided by the unique constraints of elderly adventure tourism scenarios. The lightweight convolutional neural network architecture was specifically chosen for its ability to achieve high accuracy (92.3%) while maintaining minimal memory footprint (15 MB), making it suitable for deployment on resource-constrained edge devices. The modified Hungarian algorithm was selected for its efficiency in solving assignment problems with multiple constraints, particularly its capability to incorporate temporal proximity weights and adaptive health risk factors into the optimization process. Furthermore, the federated learning framework was implemented to address critical privacy concerns while enabling collaborative model improvement across distributed tourism operators. These methodological choices collectively ensure the system meets the dual requirements of real-time responsiveness and physiological safety for elderly participants.

The system incorporates a robust offline mode for operation in connectivity-limited environments. Each edge node maintains cached copies of the latest federated model parameters and participant profiles, enabling continuous itinerary optimization without real-time updates. During offline operation, the system adopts a conservative safety-first approach, automatically reducing activity intensity thresholds by 20% and extending recovery periods by 25% to compensate for the absence of real-time model improvements. Local biometric monitoring continues uninterrupted, with emergency intervention capabilities maintained through predefined safety thresholds. All decisions are logged for subsequent synchronization, ensuring seamless transition between online and offline modes while prioritizing participant safety.

The system’s interoperability with existing travel planning platforms is ensured through a standardized API gateway that translates between the internal optimization model and external booking/reservation systems. This gateway supports common tourism data standards such as ONVX and OpenTravel API, allowing bidirectional synchronization of itineraries, participant preferences, and real-time availability. The API layer adds minimal latency (under 50 ms per transaction) and operates independently from the critical health monitoring path, ensuring that safety functions remain uncompromised while enabling integration with broader tourism service ecosystems.

The architecture incorporates robust connectivity resilience mechanisms to handle network disruptions common in remote adventure locations. Each edge node maintains a local cache of recent model parameters and participant profiles, enabling continuous operation during connectivity drops. When disconnections exceed 45 s, the system automatically transitions to a conservative safety mode that reduces activity intensity by 25% and extends recovery periods, prioritizing health safety over personalization. A data synchronization protocol with exponential backoff retry ensures efficient catch-up once connectivity is restored, with health-critical information receiving transmission priority. This approach was validated through network degradation tests showing maintained safety functionality even under 20% packet loss conditions.

The overall system architecture illustrated in Figure 1 integrates multiple core components working in concert. Lightweight convolutional neural networks process participant profiles and real-time health data, while a priority-based scheduling algorithm enables dynamic itinerary adjustments based on changing conditions. An edge computing framework efficiently distributes computational tasks across personal devices and infrastructure nodes, optimizing resource utilization. Complementing this, a federated learning module facilitates privacy-preserving model aggregation across the system without compromising individual data security.

This integrated architecture supports seamless coordination between activity planning, transportation arrangements, and health monitoring systems. The harmonious operation of these components is essential for delivering effective real-time optimization in elderly tourism scenarios, where responsiveness and reliability are critical requirements.

The computational workflow follows a producer-consumer model where wearable devices (producers) stream preprocessed biometric data to edge nodes (consumers) at 1 Hz intervals. Each edge node maintains a local buffer ${\mathbf{D}}_t\in {\mathrm{R}}^{n\times m}$ containing the most recent 60 s of physiological measurements from all connected participants, where $n$ represents the number of vital sign metrics and $m$ denotes the time window length. The buffer updates follow a first-in-first-out (FIFO) policy:

$${\mathbf{D}}_t=\mathrm{u}\mathrm{p}\mathrm{d}\mathrm{a}\mathrm{t}\mathrm{e}\left({\mathbf{D}}_{t-1},{\mathbf{d}}_t\right)$$

(4)

${\mathbf{D}}_t$: The data buffer at time t, which is a matrix $\mathrm{D}\in {\mathrm{R}}^{m\times n}$ containing the most recent physiological measurements (e.g., heart rate, SpO₂) from all connected participants. ${\mathbf{D}}_{t-1}$: The data buffer from the previous time step (t − 1). ${\mathbf{d}}_t$: The latest sensor readings arriving at time t.

This equation describes the real-time data management on the edge nodes (NVIDIA Jetson AGX Orin). Each edge node maintains a rolling buffer of the last 60 s of biometric data from each participant. This sliding window allows the system to perform continuous anomaly detection (e.g., sudden heart rate drop) while strictly limiting memory usage to just 2.4 MB per participant, which is crucial for efficient edge computing.

4.2. Lightweight Model Deployment with MobileNetV3

We adapt the MobileNetV3 architecture for edge deployment by replacing standard convolutions with depthwise separable operations. The model processes participant profiles $\mathbf{P}\in {\mathrm{R}}^{n\times d}$ through a series of inverted residual blocks with squeeze-and-excitation attention:

$${\mathbf{F}}_i=\mathrm{D}\mathrm{W}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}\left({\mathbf{P}}_i\right)+\mathrm{P}\mathrm{o}\mathrm{i}\mathrm{n}\mathrm{t}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}\left(\mathrm{ReLU}\left(\mathrm{BN}\left(\mathrm{DWConv}\left({\mathbf{P}}_i\right)\right)\right)\right)$$

(5)

${\mathbf{F}}_i$: The output feature map of the i-th layer. ${\mathbf{P}}_i$: The input to the i-th layer, which represents the participant’s profile data (e.g., demographics, preferences). DWConv: Depthwise Convolution, a lightweight convolution operation that applies a single filter per input channel to reduce computational cost. $\mathrm{P}\mathrm{o}\mathrm{i}\mathrm{n}\mathrm{t}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}$: A 1 × 1 convolution that combines the features from the depthwise convolution. $\mathrm{BN}$: Batch Normalization, a technique to stabilize and accelerate the training process. $\mathrm{ReLU}$: Rectified Linear Unit, an activation function that introduces non-linearity into the model.

This formula defines the core building block of the lightweight neural network deployed on edge devices. The use of Depthwise Separable Convolutions (DWConv + PointwiseConv) is a key reason why the model is so efficient. It achieves 92.3% accuracy in predicting activity preferences while requiring only 3.2 million parameters (15 MB of storage), making it suitable for running on resource-constrained devices like smartwatches and the Jetson modules.

The model’s memory efficiency stems from two key design choices: (1) channel reduction in bottleneck layers limits feature map dimensionality, and (2) hard-swish activations replace resource-intensive operations like sigmoid functions. These optimizations enable simultaneous execution of four model instances on a single Jetson Orin module (8 GB RAM) without compromising other system functions.

4.3. Dynamic Resource Allocation and Federated Aggregation

Edge nodes dynamically adjust computational resource allocation based on real-time demand signals ${\mathbf{R}}_t\in {\mathrm{R}}^k$, where $k$ represents the number of resource types (CPU, GPU, memory). The allocation policy follows a constrained optimization formulation.

Schema 2. Dynamic Resource Allocation Model. The resource allocation policy follows a constrained optimization formulation defined as:

$$\underset{\mathbf{A}}{\mathrm{m}\mathrm{a}\mathrm{x}}\sum _{i=1}^k{u}_i\left(a_i\right)\hspace{1em}\mathrm{s}.\mathrm{t}.\hspace{1em}\sum _{i=1}^k{a}_i\le C$$

(6)

where $u_i\left(\cdot \right)$ denotes utility functions for each resource type, $a_i$ represents allocated resources, and $C$ indicates total available capacity. The system solves this optimization every 5 s using a greedy approximation algorithm with $O\left(k\mathrm{l}\mathrm{o}\mathrm{g}k\right)$ complexity. $u_i\left(a_i\right)$ represents the utility function for the i-th resource type. It quantifies the benefit or performance gained from allocating $a_i$ amount of that resource. Where k represents the number of different resource types (e.g., CPU, GPU, memory).

The framework incorporates real-time external data feeds including weather APIs, venue availability services, and transportation status updates to detect disruptions. When unexpected changes occur (e.g., activity cancellations, severe weather, transportation delays), the system immediately triggers a re-optimization cycle that: (1) evaluates alternative activities from a pre-validated backup database filtered by current participant physiological states, (2) adjusts temporal sequencing to accommodate new constraints while maintaining recovery requirements, and (3) calculates rerouting options considering proximity, accessibility, and health suitability. This adaptive capability was validated during testing through simulated disruption scenarios, maintaining 89% itinerary continuity despite 30% activity cancellation rates.

The utility functions $u_i\left(\cdot \right)$ incorporate domain-specific weights derived from expert consultation and empirical analysis of participant data. Health safety constraints receive the highest weighting (0.45), followed by preference matching (0.35) and temporal continuity (0.20), reflecting the priority of physiological well-being in elderly adventure tourism. These weights were validated through A/B testing showing a 38% reduction in emergency interventions compared to equal weighting approaches.

For medical emergencies such as sudden blood pressure drops or cardiac events, the system implements an immediate response protocol that overrides all other optimization objectives. When vital sign anomalies exceed adaptive emergency thresholds (derived from individual baselines and population norms), the system: (1) triggers audible and haptic alerts on both participant wearables and guide devices, (2) automatically reroutes the itinerary to the nearest medical facility while considering terrain constraints, (3) establishes emergency communication channels with local emergency services, transmitting critical health data and location information, and (4) initiates group-level contingency plans to ensure companion participants are appropriately managed. This emergency protocol operates with the highest system priority, preempting any ongoing computational tasks to ensure sub-second response times.

Federated aggregation occurs at 30-min intervals, combining local model updates ${\mathbf{w}}_k$ from $K$ edge nodes through weighted averaging:

$${\mathbf{w}}_g^{t+1}=\sum _{k=1}^K{\displaystyle \frac{n_k}{N}}{\mathbf{w}}_k^t+\eta \nabla \mathcal{L}\left({\mathbf{w}}_g^t\right)$$

(7)

where $K$ represents the total number of edge nodes, $n_k$ is the number of local samples on the k-th node, where $n_k$ denotes the number of samples at node $k$, $N$ represents the total sample count, and $\eta$ controls the global learning rate. The term $\eta \nabla \mathcal{L}\left({\mathbf{w}}_g^t\right)$ introduces a global gradient-based refinement, where $\eta$ is the global learning rate and $\nabla \mathcal{L}\left({\mathbf{w}}_g^t\right)$ is the gradient of the loss function evaluated on the current global model ${\mathbf{w}}_g^t$.

This additional step enhances the model’s adaptability and helps mitigate potential performance degradation due to non-IID data distributions across nodes. As highlighted in the experimental results, this aggregation mechanism reduces communication overhead by 73% compared to synchronous federated averaging and achieves 80% of peak recommendation accuracy within just 15 federated rounds—significantly outperforming centralized baselines while rigorously preserving data privacy through localized training and secure aggregation protocols.

4.4. Federated Learning Protocol for Preference Prediction

The federated training process employs a hybrid loss function combining cross-entropy for activity classification and mean squared error for exertion score regression.

Schema 3. Privacy-Preserving Federated Learning Protocol. The federated training process employs a hybrid loss function combining cross-entropy for activity classification and mean squared error for exertion score regression:

$$\mathcal{L}=\alpha {\mathcal{L}}_{CE}+\left(1-\alpha \right){\mathcal{L}}_{MSE}$$

(8)

$\mathcal{L}$: The total hybrid loss function that is minimized during the local training process on each edge node. This single loss function combines two distinct objectives. $\alpha$: A weighting hyperparameter that balances the contribution of the two loss components. Its value ranges between 0 and 1, determining the relative importance of the classification task versus the regression task during model training. ${\mathcal{L}}_{CE}$: The Cross-Entropy loss. This component is standard for classification tasks. It measures the difference between the predicted probability distribution over activity classes (e.g., hiking, museum visit) and the true participant preference, thereby directly optimizing for recommendation accuracy. ${\mathcal{L}}_{MSE}$: The Mean Squared Error loss. This component is typically used for regression tasks. In this context, it calculates the error between the predicted exertion score for an activity and the actual or expected physiological exertion level, ensuring the model learns to accurately estimate the physical demand of recommended activities.

The hybrid loss function $\mathcal{L}=\alpha {\mathcal{L}}_{CE}+\left(1-\alpha \right){\mathcal{L}}_{MSE}$ enables joint optimization of activity preference prediction (via cross-entropy ${\mathcal{L}}_{CE}$ and physiological exertion estimation (via mean squared error ${\mathcal{L}}_{MSE}$. This dual objective ensures recommendations balance participant enjoyment with health safety, a core innovation of the federated learning framework. where $\alpha$ balances the contribution of each objective. Local training occurs on edge nodes using participant-specific data batches, with gradients clipped at ${\mathcal{l}}_2$-norm 1.0 to ensure differential privacy.

To mitigate demographic biases, we employ a fairness-constrained aggregation scheme that modifies the standard federated averaging approach. Rather than relying solely on sample-based weighting $\left(n_k/N\right)$, we incorporate demographic parity regularizers that minimize performance disparities across subgroups. The aggregation weights are adjusted using a fairness-aware weighting function:

$$w_g=\sum ({\lambda }_d\ast n_{k,d}/N_d)\ast w_k$$

(9)

$w_g$ represents the updated global model parameters. This is the new model obtained through aggregation after fairness adjustment. ${\lambda }_d$ refers to the fairness coefficient of the DTH demographic group (such as a specific age group, gender, or health status). This is a key parameter used to artificially adjust the contribution weight of this group in the global model to correct potential biases. $n_{k,d}$ refers to the number of samples belonging to the DTH demographic group at the KTH edge node. $w_k$ refers to the local model parameter update of the KTH edge node.

Formula (9) reflects that the paper has deeply integrated the concept of Responsible AI into the technical solution. It not only optimizes the accuracy and efficiency of the model, but also actively and mathematically ensures that the model’s decisions (activity recommendations) are fair and unbiased for different subgroups of elderly users, which is the key to whether the system can be ethically and responsibly deployed in the real world.

The system implements secure aggregation through additive homomorphic encryption, preventing the central server from accessing individual model updates. Each edge node $k$ encrypts its update ${\mathbf{w}}_k$ with a shared secret key $s_k$:

$${\mathbf{w}}_k^{enc}={\mathbf{w}}_k+s_k\mathbf{1}$$

(10)

${\mathbf{w}}_k^{enc}$ represents the encrypted local model parameter update sent by the KTH edge node. This is a string of ciphertext that cannot be directly interpreted by the server. ${\mathbf{w}}_k$ displays the unencrypted update of the original model parameters obtained by the local training of the KTH edge node. This is sensitive information that needs protection. $s_k$ represents a shared secret key generated by the KTH edge node. This key is part of a secure aggregation protocol, usually generated together with a central server or other nodes, and ensures that it can be offset after aggregation. Here, $\mathbf{1}$ represents a vector where all elements are 1, and its dimension is the same as the model parameter vector ${\mathbf{w}}_k$ is the same. Its function is to enable the secret key $s_k$ to evenly “mask” every parameter in the model update vector.

In conclusion, Formula (10) represents the ultimate line of defense in the design of system privacy protection. It not only relies on the distributed architecture of federated learning, but also introduces cryptographic primients to ensure that the sensitive health data of participants is protected to the greatest extent throughout the entire data lifecycle—from local training to transmission and then to global aggregation, thereby building a real-time health perception recommendation system that is both efficient and secure, and in line with ethical standards.

The server aggregates encrypted updates before decryption, ensuring no single participant’s data can be reconstructed from the transmitted information. This cryptographic protection adds only 12 ms to the overall aggregation latency while providing provable privacy guarantees.

The federated edge architecture demonstrates particular effectiveness in cold-start scenarios, where new participants benefit from collective knowledge without sharing raw data. Experimental results show the system reaches 80% of maximum recommendation accuracy after just 15 federated rounds, compared to 40 rounds required by centralized training approaches. This rapid convergence stems from the diversity of local datasets across tourism operators, which collectively cover a wider range of activity patterns and physiological responses.

5. Experimental Setup and Methodology

This study employs a comprehensive mixed-methods research design to evaluate the proposed edge-enhanced federated learning framework. The research flow adopted a systematic, multi-stage process: (1) System Construction: Development and integration of the edge computing framework, lightweight CNN models, and federated learning protocol; (2) Controlled Laboratory Testing: Evaluation of computational efficiency, latency, and accuracy metrics under simulated conditions; (3) Simulated User Scenarios: Validation of recommendation quality, health safety compliance, and personalization capabilities using standardized benchmarks and the collected dataset; (4) Comparative Analysis: Performance comparison against established baseline methods across defined metrics. This structured approach ensures comprehensive validation of both the technical performance and practical applicability of the proposed framework. The selection of computational methods was strategically guided by the unique constraints of elderly adventure tourism scenarios. The lightweight convolutional neural network architecture was specifically chosen for its ability to achieve high accuracy (92.3%) while maintaining minimal memory footprint (15 MB), making it suitable for deployment on resource-constrained edge devices. The modified Hungarian algorithm was selected for its efficiency in solving assignment problems with multiple constraints, particularly its capability to incorporate temporal proximity weights and adaptive health risk factors into the optimization process. Furthermore, the federated learning framework was implemented to address critical privacy concerns while enabling collaborative model improvement across distributed tourism operators. These methodological choices collectively ensure the system meets the dual requirements of real-time responsiveness and physiological safety for elderly participants. This rationale provides a clear guideline for other researchers seeking to adapt our framework to similar domains with stringent latency, privacy, and resource constraints. To rigorously test the accuracy of our experiments, we implemented a multi-layered validation strategy. This included the use of standardized benchmarks, controlled noise injection to simulate real-world sensor variability, stratified participant sampling to ensure demographic representativeness, and statistical methods such as mixed-effects modeling with Bonferroni-Holm correction. Furthermore, we conducted ablation studies to isolate the impact of key components (e.g., federated learning, dynamic thresholds) on overall accuracy, ensuring that each element’s contribution was quantitatively assessed and validated. The methodology integrates experimental validation of technical performance metrics with user-centered evaluation of system effectiveness, complemented by systematic comparative analysis against established baseline methods. The research design follows a rigorous systematic approach to ensure objectivity, reproducibility, and comprehensive assessment across all relevant dimensions of system performance, addressing both technical capabilities and practical applicability in real-world elderly tourism scenarios.

To validate the proposed framework’s effectiveness, we designed comprehensive experiments evaluating three key aspects: (1) recommendation accuracy under varying participant conditions, (2) system responsiveness in real-world deployment scenarios, and (3) computational efficiency across different edge hardware configurations. The evaluation methodology follows established protocols for federated learning systems while incorporating tourism-specific performance metrics (AzharShokoufeh et al., 2025).

The experimental methodology employs a hierarchical validation framework that encompasses multiple evaluation perspectives. This framework integrates quantitative assessment of computational efficiency, latency, and resource utilization metrics under controlled laboratory conditions with evaluation of recommendation accuracy, health safety compliance, and personalization capabilities using standardized benchmarks. Additionally, the methodology incorporates measurement of perceived safety, satisfaction, and usability through structured participant surveys and interviews, while also facilitating systematic comparison against state-of-the-art baseline methods using predefined evaluation metrics. This integrated multi-perspective approach ensures comprehensive validation of both technical performance and practical applicability.

5.1. Dataset and Participant Profiles

The study utilized a multimodal dataset comprising 1248 elderly adventure tourism participants across 12 different activity categories, collected through partnerships with specialized tour operators (Qi & Han, 2024). Each participant profile includes demographic information (age 65–82, mean 71.3 ± 4.7) (

Table 1. The following table summarizes the participant distribution across key demographic strata.

Demographic Characteristic	Strata	Percentage	Target Population Benchmark
Age Group	65–69 years	32.5%	31.8%
	70–74 years	28.7%	29.2%
	75–82 years	38.8%	39.0%
Gender	Female	52.3%	53.1%
	Male	47.7%	46.9%
Health Status	Robust	41.2%	40.5%
	Moderate comorbidities	35.8%	36.2%
	Managed chronic conditions	23.0%	23.3%

), baseline physiological measurements, and activity preference ratings collected through standardized questionnaires. Participants were selected through a stratified sampling approach designed to ensure broad representativeness across key demographic and health dimensions. The selection criteria included: (1) age distribution across the 65–82 range with proportional representation from 5-year age cohorts; (2) balanced gender representation (52% female, 48% male); (3) diverse health status levels based on pre-screening using the Functional Comorbidity Index; and (4) varied previous adventure tourism experience from novice to experienced.

The sampling framework was developed Participants were recruited through a multi-stage process: initial screening via health questionnaires, followed by physical readiness assessments conducted by certified geriatric fitness specialists, and finally, informed consent procedures emphasizing the study’s data privacy protections.

To ensure broad-scale representativeness, we implemented quota sampling based on population parameters derived from the World Health Organization’s global aging reports and tourism demographic studies. This approach ensured proportional representation across age subgroups, gender, and health status categories, providing a participant pool that reflects the diversity of the elderly adventure tourism population. The final cohort’s characteristics were validated against known population parameters using chi-square goodness-of-fit tests (p > 0.05 for all demographic variables), confirming the sample’s representativeness.

Real-time biometric data streams were captured at 10 Hz sampling rate using FDA-cleared wearable devices (Khayyat et al., 2024), including:

• Heart rate variability (RMSSD 28.4 ± 11.2 ms)
• Blood oxygen saturation (SpO₂ 96.2 ± 2.1%)
• Galvanic skin response (GSR 5.1 ± 2.8 μS)
• Triaxial acceleration (3.2 ± 1.5 m/s² peak)

All participants provided comprehensive informed consent through a multi-stage process that explicitly detailed how their data would be used for real-time itinerary optimization. The consent protocol included visual aids demonstrating the federated learning approach and clear explanations of how biometric data would influence activity recommendations. Participants were informed of their right to access, correct, or delete their data at any point, and could view a real-time dashboard showing how their physiological data was being used to adjust their itinerary. User feedback was collected through multiple channels: structured post-activity surveys (5-point Likert scale), semi-structured interviews focusing on comfort and safety perceptions, and real-time preference adjustments during activities. This feedback directly influenced several system improvements: (1) the exertion threshold algorithm was adjusted to reduce false positives after participants reported overly conservative scheduling, (2) the notification system was modified to provide clearer explanations for itinerary changes based on user confusion feedback, and (3) the activity database was expanded to include more low-intensity alternatives based on participant suggestions. This iterative feedback process ensured the system evolved to better meet user needs while maintaining safety standards. This transparency framework ensured participants maintained agency over their data while understanding the benefits of personalized health-aware scheduling.

The dataset was partitioned into training (70%), validation (15%), and test (15%) sets while maintaining balanced distributions across age groups and activity types. To simulate real-world conditions, we introduced controlled noise artifacts (20 dB SNR) matching levels observed in field deployments (Yang et al., 2025).

Data collection followed a rigorously defined protocol to ensure objectivity and reproducibility. Quantitative performance metrics were automatically logged by the system at predefined intervals, while qualitative feedback was collected through standardized instruments administered by trained researchers blind to the experimental conditions. All data underwent a three-stage validation process: (1) automated sanity checks for outlier detection, (2) manual verification of a randomly selected 10% subset, and (3) cross-validation against secondary data sources where available.

Statistical analysis employed a mixed-effects model approach to account for both within-participant and between-participant variability. All comparative analyses included correction for multiple comparisons using the Bonferroni-Holm method, and effect sizes were calculated using Cohen’s d for continuous variables and Cramer’s V for categorical variables. The analysis plan was pre-registered before data collection to prevent data-dependent analytical choices, and all analytical code is available for reproducibility verification.

5.2. Baseline Methods and Evaluation Metrics

The proposed framework was compared against three established approaches to evaluate its performance relative to existing methods.

• Centralized Cloud-Based Recommender (CCR): A conventional tourism recommendation system using centralized processing (Lu et al., 2023)
• Edge-Only Health-Aware Scheduler (EHS): A recent edge computing approach without federated learning capabilities (Rui et al., 2023)
• Hybrid Federated Baseline (HFB): A basic federated learning implementation without dynamic resource allocation (Hardy & Aryal, 2020)

Evaluation metrics were selected to capture both recommendation quality and system performance:

$$\mathrm{Precision}@\mathrm{k}={\displaystyle \frac{\left|{\mathrm{Recommended}}_k\cap {\mathrm{Preferred}}_k\right|}{k}}$$

(11)

$$\mathrm{Health}\mathrm{Safety}\mathrm{Score}=1-{\displaystyle \frac{\sum _{i=1}^n\mathbb{I}\left({\mathrm{Violation}}_i\right)}{n}}$$

(12)

$$\mathrm{Response}\mathrm{Latency}=t_{\mathrm{input}}-t_{\mathrm{output}}$$

(13)

In Formula (11), Precision@k represents the accuracy of the first k recommendation results of the evaluation recommendation system. Its value ranges from 0 to 1. The higher the value, the better the recommended quality. In the formula, k represents a predefined integer, indicating the first k recommended activities considered (for example, Top-3 or Top-5 recommendations). ${\mathrm{Recommended}}_k$ represents the set of activities that the user actually prefers or selects (in the experiment, it is real labeled data).

In Formula (12), $\mathrm{Health}\mathrm{Safety}\mathrm{Score}$ represents a standardized score used to quantify the overall performance of the system in adhering to health constraints. The closer the score is to 1, the safer the system is. $n$ is displayed in the article as the total number of evaluations or the total number of activities. In the article, ${\mathit{Violation}}_i$ is displayed as an indicator function, indicating whether a violation of health constraints occurred in the i-th assessment (for example, activities beyond the physiological tolerance of the participants were arranged). If a violation occurs, its value is 1; otherwise, it is 0.

Formula (13) represents Response Latency, where $\mathrm{Response}\mathrm{Latency}$ indicates the end-to-end response latency of the system, that is, the total time elapsed from input to output. The smaller this value is, the faster the system responds. $t_{\mathrm{input}}$ represents the input timestamp, indicating the time point when sensor data or requests arrive and the system begins processing. $t_{\mathrm{output}}\hspace{0.25em}$ refers to the output timestamp, indicating the point in time when the system generates recommendations or intervention measures and completes the output.

5.3. Implementation Details

The system was implemented on NVIDIA Jetson AGX Orin (32 GB) modules for edge nodes and Samsung Galaxy Watch4 for participant devices. The federated learning protocol used a 5-layer CNN with the following hyperparameters:

$$\mathrm{Learning}\mathrm{Rate}=0.001\times {\displaystyle \frac{1}{1+0.01\times \mathrm{r}\mathrm{o}\mathrm{u}\mathrm{n}\mathrm{d}}}$$

(14)

$$\mathrm{Batch}\mathrm{Size}=32\times \mathrm{m}\mathrm{i}\mathrm{n}\left(1,{\displaystyle \frac{\mathrm{Memory}}{1.5\mathrm{GB}}}\right)$$

(15)

Formula (14) indicates the attenuated learning rate scheduling. $\mathrm{Learning}\mathrm{Rate}$ represents the learning rate used when training the local model in each round of federated learning. It is a hyperparameter that controls the step size of the model parameters adjusted according to the gradient of the loss function. 0.001 represents the initial learning rate. This is the benchmark learning rate value used at the beginning of the training. “round” indicates the current index of federated learning rounds. As the training progresses, this number will keep increasing. 0.01 represents the attenuation coefficient. This value controls the rate at which the learning rate decreases as the number of training rounds increases.

Formula (15) represents dynamic batch size adjustment. Batch Size represents the number of training samples used in one iteration during local training. 32 refers to the base batch size. This is a commonly used value that strikes a balance between model performance and memory consumption.

“min” represents the function of taking the minimum value. It is used to ensure that the calculation result does not exceed the benchmark batch size. “Memory” represents the available memory size of the current edge node. This is a dynamic variable that depends on the hardware configuration of the device and the current system load. 1.5 GB is displayed as the memory threshold. This is a preset constant representing the approximate memory capacity required to run the base batch size (32) smoothly.

Local training ran for 5 epochs per federated round with early stopping (patience = 2). The modified Hungarian algorithm was configured with temporal window $w=15$ min and exertion limit $\theta =0.7$ based on pilot studies (Devries et al., 1989). These parameter values were rigorously selected through an iterative optimization process: the exertion limit $\theta =0.7$ corresponds to approximately 70% of age-predicted maximum heart rate reserve, aligning with established guidelines for moderate-intensity exercise in elderly populations. The 15-min temporal window (w) was determined to provide adequate physiological recovery between activities while maintaining itinerary continuity. Pilot studies with our participant cohort confirmed that these parameters optimally balance safety constraints with activity enjoyment, achieving a 92.7% health safety score while maintaining high participant satisfaction. All experiments were repeated 5 times with different random seeds to ensure statistical significance. Interoperability testing was conducted with three major travel platform simulators (Amadeus GDS, Booking.com API, and Expedia Partner Central) showing successful integration rates of 94.3%, 91.7%, and 89.8%, respectively. The API gateway handled concurrent requests with 99.2% success rate at peak loads of 150 requests/second, demonstrating production-ready compatibility.

5.4. Simulation Environment

To evaluate scalability, we developed a tourism scenario simulator modeling:—Participant movement patterns (Levy walk distribution)—Wireless channel conditions (3GPP TR 38.901 UMi model)—Edge node placements (Poisson point process, λ = 0.0005).

The simulator incorporated device heterogeneity by varying:—CPU capabilities (1–4 cores @ 1–2.5 GHz)—Memory constraints (512 MB–4 GB)—Network bandwidth (10–100 Mbps).

This setup allowed controlled testing of the system’s adaptive capabilities under realistic resource constraints while maintaining reproducibility across experimental runs.

6. Experimental Results and Analysis

6.1. Recommendation Accuracy and Personalization

Evaluating the system’s core functionality, our lightweight CNN architecture achieved superior performance in activity recommendation accuracy compared to baseline methods. As shown in

Table 2. Comparative recommendation accuracy across methods.

Method	Precision@5	Model Size (MB)	Memory Usage (MB)
Centralized Cloud (CCR)	89.10%	47	310
Edge-Only Scheduler (EHS)	85.70%	29	190
Hybrid Federated (HFB)	88.40%	38	245
Proposed Framework	92.30%	15	98

, the proposed framework maintains 92.3% precision@5 while reducing model size by 68% compared to conventional approaches. These results align with recent findings in edge-based recommendation systems (Sun et al., 2020) while significantly improving model efficiency. The 68% reduction in model size compared to conventional approaches demonstrates the effectiveness of depthwise separable convolutions, supporting similar efficiency gains reported in mobile-optimized architectures (Tran et al., 2025). This efficiency stems from the depthwise separable convolutions and channel reduction techniques described in Section 4.2.

The federated learning component demonstrates particular effectiveness in cold-start scenarios, where new participants benefit from collective knowledge without compromising privacy. As illustrated in Figure 2, our framework reaches 80% of maximum accuracy within 15 federated rounds, compared to 40 rounds required by centralized approaches. This accelerated convergence results from the diversity of local datasets across tourism operators, which collectively cover broader activity patterns.

6.2. Real-Time Performance and Latency

Meeting stringent response time requirements proves critical for health-sensitive itinerary adjustments. The priority-based scheduling algorithm reduces emergency intervention frequency by 38% compared to conventional approaches, while maintaining sub-300 ms latency for critical path operations. Figure 3 demonstrates how the system dynamically adapts activity intensity based on real-time physiological feedback, preventing hazardous exertion levels while maximizing participant engagement.

The edge computing architecture achieves this responsiveness through optimized resource allocation, as formalized in Equation (6). Benchmark tests show the system processes concurrent biometric streams from 12 participants while utilizing only 73% of available edge node resources, leaving sufficient headroom for peak demand scenarios. This efficiency enables deployment on modest hardware configurations without sacrificing safety margins.

Under simulated poor connectivity conditions (5 dB SNR, 20% packet loss), the system maintained functional safety monitoring with 95% reliability, though end-to-end latency increased to 650 ± 120 ms due to retransmission requirements. The local caching mechanism ensured continuous itinerary optimization during complete network outages, with synchronization completing within 2.3 ± 0.8 s upon connectivity restoration. These results demonstrate the system’s robustness for deployment in remote adventure tourism locations with intermittent network coverage.

In extended offline operation tests (up to 24 h without connectivity), the system maintained 88.2 ± 3.1% of its safety effectiveness and 82.5 ± 4.7% of recommendation accuracy compared to online operation. The conservative safety adjustments successfully prevented all critical health incidents during offline testing, though with a 15.3% increase in conservative itinerary modifications. The data synchronization protocol efficiently reconciled offline decisions upon reconnection, completing within 3.2 ± 1.1 min after 24 h of disconnected operation, demonstrating the practical viability for multi-day adventures in remote locations.

As shown in

Table 3. Comparison of real-time performance metrics across different system components.

System Component	Average Latency (ms)	Peak Throughput (requests/sec)	Energy Consumption (W)
Wearable Data Collection	45 ± 12	10	2.1
Edge Node Processing	110 ± 25	8	18.7
Federated Aggregation	145 ± 30	2	7.2
Emergency Intervention	280 ± 50	1	22.4

, the latency breakdown demonstrates how the proposed edge-enhanced architecture achieves its 300 ms end-to-end target (Section 4.1) through optimized workload distribution. Wearable devices handle initial data preprocessing with minimal latency (45 ms), while the edge nodes’ 110 ms processing time reflects the efficiency of the lightweight CNN operations (Section 4.2). The federated aggregation’s 145 ms latency—73% lower than conventional approaches (Section 6.4)—validates the cryptographic optimizations described in Equation (9). Notably, emergency interventions consume 22.4 W (approaching the system’s 28 W limit from Section 6.4), highlighting the priority-based scheduler’s effectiveness in reducing such events by 38% (Section 6.2). These metrics collectively confirm the framework’s ability to balance real-time responsiveness with energy efficiency in elderly tourism scenarios.

6.3. Health Safety and Constraint Satisfaction

The modified Hungarian algorithm’s incorporation of temporal proximity weights and adaptive health risk factors yields significant improvements in physiological constraint satisfaction. Experimental results indicate a 92.7% health safety score across all test scenarios, with violations primarily occurring during unexpected environmental changes (e.g., sudden weather shifts). The density contour in Figure 4 reveals the system’s effective avoidance of high-risk activity assignments, concentrating recommendations in the optimal exertion-risk quadrant.

Notably, the framework’s adaptive threshold mechanism automatically adjusts safety margins based on individual response patterns. Participants showing consistent physiological stability receive 18% more high-intensity activity recommendations than those with variable responses, demonstrating the system’s personalized risk calibration capabilities.

6.4. Computational Efficiency and Resource Utilization

To quantitatively evaluate the computational efficiency of the proposed framework,

Table 4. Computational resource utilization during peak operation (12 concurrent participants).

Resource Type	Average Utilization (%)	Peak Utilization (%)	Memory Allocation (MB)
CPU Cores	63 ± 8	89	-
GPU Cores	55 ± 6	78	-
Shared Memory	-	-	98 ± 12
Model Cache	-	-	15 ± 2
Communication Buffer	-	-	2.4 ± 0.3

presents detailed resource utilization metrics during peak operation with 12 concurrent participants (as benchmarked in Section 6.2). The data reveals three critical characteristics: (1) The CPU/GPU utilization rates (63%/55% average) demonstrate effective load balancing across edge nodes, staying well below the 89%/78% safety thresholds—a design feature discussed in Section 4.3’s dynamic allocation model (Equation (6)); (2) Memory allocation remains stable at 98 MB (shared) + 15 MB (model cache), consistent with the MobileNetV3 optimization claims in Section 4.2; and (3) The per-participant communication buffer (2.4 MB) validates the FIFO policy’s efficiency (Equation (4)) while meeting the 300 ms latency target. Notably, these metrics collectively achieve the 28 W power budget referenced in Section 6.4, proving the system’s viability for resource-constrained tourism deployments.

The edge deployment’s resource efficiency metrics substantiate its practical viability for tourism operations. Memory usage remains stable at 98 MB per edge node during normal operation, with peak utilization not exceeding 1.8 GB even during simultaneous model updates. Energy consumption measurements show the complete system (edge node + 8 wearables) requires only 28 W during active itinerary optimization—comparable to a laptop computer.

Communication overhead analysis reveals the federated protocol’s bandwidth efficiency, transmitting just 4.7 KB per model update compared to 320 KB for full gradient exchanges in conventional approaches. This 98% reduction enables reliable operation in bandwidth-constrained environments typical of remote tourism locations.

6.5. User Satisfaction, Cost Efficiency, and Scalability

Beyond technical metrics, we evaluated the system’s practical impact through user satisfaction surveys, cost efficiency analysis, and scalability tests. User satisfaction was measured via a standardized questionnaire administered to 120 participants after simulated 3-day tours, assessing perceived safety, enjoyment, and personalization on a 5-point Likert scale. Cost efficiency was evaluated by comparing the total cost of ownership (TCO) per participant per day, including hardware, energy, and data transmission costs. Scalability was tested by simulating scenarios with 50 to 200 concurrent participants and measuring end-to-end latency.

As summarized in

Table 5. Comparison of user satisfaction, cost efficiency, and scalability metrics.

Metric	Proposed Framework	CCR	EHS	HFB
User Satisfaction (1–5 scale)	4.6 (±0.3)	3.8 (±0.4)	4.1 (±0.3)	4.2 (±0.3)
Cost Efficiency (TCO $/user/day)	2.38	3.51	2.95	2.8
Latency @ 100 users (ms)	285 (±35)	520 (±120)	310 (±40)	450 (±75)
Max Scalable Users (<300 ms)	150	75	125	100

, our framework achieved a mean satisfaction score of 4.6/5 (±0.3), significantly higher than all baseline methods. The edge-enhanced architecture reduced TCO by 32% compared to the cloud-centric CCR approach, primarily due to minimized bandwidth usage and efficient local processing. The system maintained its sub-300 ms latency target for up to 150 users, with a graceful degradation to 450 ms at 200 users, demonstrating robust scalability for typical elderly tour group sizes. These results confirm the framework’s practical advantages in real-world deployment, balancing high performance with positive user experience and operational cost-effectiveness.

6.6. Ablation Study

To isolate the contribution of individual system components, we conducted controlled experiments with modified configurations.

Table 6. Ablation study results (normalized metrics).

Configuration	Accuracy	Latency	Safety
Full System	1	1	1
Without Federated Learning	0.83	0.97	0.91
Static Thresholds	0.95	1.02	0.87
Centralized Optimization	0.89	0.62	0.93

presents the ablation study results comparing the performance of different system configurations using normalized metrics. The full system achieves optimal performance across all metrics (normalized to 1). Removing federated learning causes the most significant accuracy drop (17%, to 0.83) while minimally affecting latency (0.97), validating the critical role of federated learning in maintaining recommendation quality while preserving privacy, as emphasized in Section 4.4. The static threshold configuration shows a 13% safety score reduction (0.87), demonstrating the importance of the dynamic threshold mechanism described in Section 4.3 for real-time health constraint adaptation. While centralized optimization improves latency (0.62), it compromises the system’s scalability and privacy—a key trade-off that the proposed edge architecture deliberately avoids, as discussed in Section 6.4 regarding the advantages of distributed computation. These results quantitatively confirm the framework’s design choices in balancing accuracy, responsiveness, and safety through its integrated components.

Our findings have significant implications for both tourism management practice and technology development. The 38% reduction in emergency interventions demonstrates that real-time adaptive systems can substantially enhance safety without compromising experience quality, potentially establishing new industry standards for elderly adventure tourism. This improvement in safety monitoring performance is consistent with recent advancements in real-time health intervention systems (S. Khan et al., 2020), though the current framework achieves significantly lower latency through edge optimization. The sub-300 ms response time represents a notable advancement over cloud-based systems typically reporting 500–800 ms latency (J.-X. Hu et al., 2017). The successful deployment on resource-constrained edge devices (Section 6.4) validates the practical viability of AI-driven personalization in remote tourism settings. Furthermore, our federated learning approach offers a blueprint for privacy-preserving tourism analytics that could transform how the industry handles sensitive customer data. These advancements collectively contribute to a new paradigm of responsible tourism technology that balances innovation with ethical considerations and safety requirements, setting a benchmark for future developments in smart tourism ecosystems.

The results demonstrate that federated learning contributes most significantly to recommendation accuracy (17% drop when removed), while the dynamic threshold mechanism proves crucial for health safety (13% improvement). Interestingly, centralized optimization actually reduces latency but at the cost of privacy and scalability—a trade-off our edge architecture deliberately avoids.

7. Discussion and Future Work

7.1. System Limitations and Scalability Constraints

The experimental results demonstrate the achievement of our primary research objectives: the system successfully provides real-time itinerary optimization (300 ms latency objective met), maintains privacy through federated learning (no raw data exchange required), and ensures safety while preserving user satisfaction (92.7% safety score, 4.6/5 satisfaction rating). The performance metrics validate our engineering approach and technical choices, confirming that the developed framework effectively addresses the practical challenges identified in elderly adventure tourism.

While the proposed framework demonstrates strong performance in controlled experiments, several practical limitations emerge when considering large-scale deployment. The current architecture assumes continuous wireless connectivity between wearables and edge nodes, which may not hold in remote adventure tourism locations with intermittent network coverage (Liu et al., 2019). Field tests revealed that signal dropout exceeding 45 s triggers conservative fallback scheduling, potentially reducing itinerary personalization. These connectivity challenges in remote tourism locations have been previously documented (Liu et al., 2019), but the current framework’s conservative fallback approach provides a novel solution that maintains safety while minimizing service disruption. The 95% reliability under poor connectivity conditions compares favorably with existing edge systems typically reporting 80–85% reliability in similar conditions (Y. Khan et al., 2024). Moreover, the system’s reliance on medical-grade wearables presents cost barriers for widespread adoption, as current devices remain prohibitively expensive for many tourism operators (S. Chen et al., 2016). To address cost barriers, our framework incorporates device-agnostic compatibility that supports a range of wearable tiers, from consumer fitness trackers to medical-grade devices. While validation used medical-grade equipment for research rigor, testing with consumer devices ($200–500 range) demonstrated maintained 87% safety effectiveness, making broader adoption economically viable. We propose a rental-based business model where operators provide wearables as part of premium packages, distributing costs across multiple users. Additionally, the rapid cost reduction in biometric sensor technology suggests these devices will become increasingly accessible, with projected 40% price decreases within three years based on market trends.

The federated learning component faces inherent scalability challenges when expanding beyond specialized tour groups to mass-market applications. As participant diversity increases, the global model must accommodate wider variations in physiological responses and activity preferences, potentially requiring more complex architectures that conflict with edge deployment constraints (Moshawrab et al., 2023). Preliminary tests with heterogeneous populations show model accuracy drops by 8–12% when extending beyond the original target demographic (65–82 years), suggesting the need for demographic-aware personalization techniques.

7.2. Ethical Considerations in Elder-Centric Federated Learning

The system’s health monitoring capabilities raise important ethical questions regarding data ownership and algorithmic transparency. While federated learning preserves raw data privacy, the aggregated models still encode sensitive patterns about elderly participants’ physiological limits and activity tolerances (Babar et al., 2024). Current implementations lack mechanisms for participants to audit or contest how their data contributions influence recommendations—a critical feature for maintaining trust in automated itinerary planning.

Our framework addresses this limitation through a transparent consent interface that allows participants to visualize how their data influences recommendations in real-time. The system provides explanatory feedback when itinerary adjustments are made based on physiological data, such as “Your heart rate variability suggests needing additional recovery time, so we’ve rescheduled the hiking activity for tomorrow morning.” This approach maintains algorithmic transparency while respecting privacy through federated learning, creating a balance between personalized optimization and participant agency that was particularly appreciated by our elderly cohort during user feedback sessions.

Furthermore, the dynamic threshold system’s risk calculations could inadvertently reinforce age-related biases if not carefully calibrated. Early deployments revealed instances where overly conservative algorithms systematically underestimated capable participants’ physical limits, creating self-fulfilling prophecies of reduced activity capacity (Taati et al., 2019). As noted in our experimental results, the current framework is specifically designed and optimized for the elderly adventure tourism segment (“Silver-Haired Special Forces”). The physiological models, exertion thresholds, and activity databases are tailored to this demographic. While the underlying architecture—integrating federated learning, edge computing, and real-time optimization—is broadly applicable, adapting it to younger or less healthy populations would require recalibrating these core components with new demographic-specific data. For example, applying the system to participants with chronic conditions would necessitate incorporating disease-specific risk models and activity contraindications. Future work will explore this promising direction of creating a more generalized, adaptive health-aware tourism platform by developing demographic-aware model personalization techniques that can dynamically adjust to the user’s age, health profile, and fitness level.

Developing fairness-aware adaptation mechanisms that distinguish between genuine physiological constraints and transient performance variations remains an open challenge. To address potential biases in federated learning, our framework incorporates multiple fairness preservation mechanisms including demographic-aware aggregation that weights contributions based on population representativeness rather than data quantity, complemented by continuous monitoring of performance metrics across age, gender, and health status subgroups using disparity detection algorithms. The system also performs adaptive re-calibration of recommendation thresholds when bias exceeding 5% is detected. Our participant cohort (n = 1248) was specifically designed to include diverse elderly subgroups through stratified sampling across age deciles (65–74, 75–82), gender balance (52% female, 48% male), and varying health conditions to ensure model fairness and generalizability across the target population.

To ensure the responsible deployment of our system across diverse cultural and regulatory environments, we propose comprehensive implementation guidelines that address varying cultural perceptions of elderly care and privacy norms. These guidelines include adaptable threshold settings that respect regional differences in risk tolerance and activity preferences. For regulatory compliance, our framework incorporates modular design principles that facilitate adherence to jurisdiction-specific data protection regulations such as GDPR, CCPA, and emerging AI governance frameworks. We further establish a continuous monitoring and evaluation protocol that tracks system performance metrics across three dimensions: safety effectiveness (emergency intervention rates), user satisfaction (experience quality scores), and ethical compliance (bias detection metrics). This tripartite evaluation framework enables tourism operators to maintain optimal system performance while ensuring ongoing compliance with evolving regulatory requirements and ethical standards across different operational contexts.

This research offers significant practical benefits for both tourism operators and elderly tourists. For operators, the 38% reduction in emergency interventions translates to substantially lower liability risks and insurance costs, while the edge-compatible architecture minimizes infrastructure investments. The federated learning approach enables collaborative improvement across operators without compromising competitive advantages or violating data protection regulations. For elderly tourists, the system provides unprecedented safety assurance through real-time health monitoring while maintaining full privacy control over their sensitive data. The personalized itinerary optimization ensures that participants can engage in adventure activities matched to their physical capabilities, enhancing both safety and enjoyment. These practical benefits position our framework as a transformative solution for the growing silver-haired adventure tourism market.

7.3. Extended Applications in Smart Tourism Ecosystems

The framework’s underlying technologies show promising potential beyond specialized elderly tourism. The hybrid edge-cloud architecture could enhance general smart tourism systems by enabling real-time group coordination across heterogeneous attractions (Rui et al., 2023). For instance, integrating transportation scheduling with activity recommendations could optimize entire vacation itineraries while accounting for participants’ evolving energy levels and interests.

The health-aware recommendation engine also has natural extensions to other demographic groups with specific physiological constraints, such as adaptive tourism for individuals with chronic conditions or disability-inclusive activity planning (Q. W. Khan et al., 2024). Future iterations could incorporate additional data modalities like environmental sensors (air quality, UV exposure) to create comprehensive well-being-oriented tourism ecosystems. Future work will focus on developing optimized algorithms for lower-cost wearable devices, exploring sensor fusion techniques that compensate for individual sensor limitations through multi-modal data integration. We are also investigating partnership models with wearable manufacturers to develop purpose-built devices specifically for adventure tourism that balance cost and accuracy requirements. These efforts aim to reduce the per-participant equipment cost below $100 while maintaining safety standards, significantly enhancing accessibility for diverse tourism markets.

The federated learning approach’s privacy-preserving characteristics make it particularly suitable for cross-border tourism collaborations, where data sovereignty regulations often prevent centralized data aggregation. By enabling knowledge sharing between international tourism operators without raw data exchange, the system could accelerate the development of culturally used recommendation models while complying with strict data protection regimes like GDPR.

7.4. Practical Implications and Beneficiaries

Our edge-enhanced federated optimization framework represents a paradigm shift in how tourism services can be delivered to elderly adventure tourists. By seamlessly integrating real-time health monitoring with dynamic itinerary plan-ning, the system enables tourism operators to offer personalized, safety-assured experiences at scale. For the hospitality sector, this technology facilitates the creation of value-added services that extend beyond traditional accommodation offerings, allowing hotels and resorts to differentiate themselves in the competitive silver tourism market. The federated learning architecture particularly addresses industry-wide concerns about data privacy and security, providing a compliant framework for collaboration across different service providers while maintaining competitive advantages. This approach not only enhances operational efficiency through reduced emergency incidents and liability costs but also creates new revenue streams through premium personalized tour packages that cater to the specific needs and capabilities of elderly tourists.

This research offers significant practical benefits for multiple stakeholders in the elderly tourism ecosystem. For tourism operators, the 38% reduction in emergency interventions translates to substantially lower liability risks and insurance costs, while the edge-compatible architecture minimizes infrastructure investments. The federated learning approach enables collaborative improvement across operators without compromising competitive advantages or violating data protection regulations. For elderly tourists, the system provides unprecedented safety assurance through real-time health monitoring while maintaining full privacy control over their sensitive data. The personalized itinerary optimization ensures that participants can engage in adventure activities matched to their physical capabilities, enhancing both safety and enjoyment. Additionally, healthcare providers benefit from reduced emergency incidents during tourism activities, and technology developers gain insights into implementing privacy-preserving AI systems for sensitive applications. These practical benefits position our framework as a transformative solution for the growing silver-haired adventure tourism market.

7.5. Operational Risks and Mitigation Strategies

The transition from experimental validation to real-world operation introduces several practical risks that must be carefully managed. Connectivity limitations in remote adventure locations represent a significant challenge, as signal dropouts exceeding 45 s trigger conservative fallback scheduling that reduces personalization. To mitigate this, we propose implementing predictive caching of itinerary options based on anticipated network conditions and participant profiles. The system’s reliance on medical-grade wearables presents cost barriers; however, our device-agnostic approach supports a range of wearable tiers, and we anticipate rapid cost reduction in biometric sensor technology (projected 40% decrease within three years). Scalability challenges in federated learning require demographic-aware personalization techniques to maintain accuracy across diverse populations. Most critically, ethical risks concerning algorithmic transparency and potential age-related biases necessitate robust monitoring systems and participant-facing explanatory interfaces that maintain trust while preserving privacy through federated learning.

8. Conclusions

The proposed edge-enhanced federated optimization framework represents a significant advancement in real-time itinerary planning for elderly adventure tourism, successfully addressing the dual challenges of personalization and physiological safety. By integrating lightweight neural networks with a modified Hungarian algorithm, the system achieves superior recommendation accuracy while maintaining the computational efficiency required for edge deployment. The federated learning architecture demonstrates particular effectiveness in preserving participant privacy without sacrificing model performance, reaching 80% of maximum accuracy within just 15 training rounds.

The experimental results demonstrate the achievement of our primary research objectives established in the introduction: (1) The system successfully addresses the separation between health monitoring and activity recommendation through integrated real-time processing, achieving 300 ms end-to-end latency for dynamic itinerary adjustments; (2) Privacy concerns in centralized data processing are resolved through our federated learning architecture, which achieves 80% of maximum accuracy within 15 federated rounds without raw data exchange; (3) Computational limitations of existing edge systems are overcome through lightweight model optimization, maintaining memory usage below 100 MB while processing 12 concurrent participants. These results confirm that all three critical challenges identified in Section 1 have been effectively addressed by our framework.

The primary contributions of this work include: (1) the development of a novel edge-enhanced federated learning framework specifically designed for elderly adventure tourism scenarios, which represents the first integrated solution that simultaneously addresses real-time health monitoring, personalized activity recommendation, and privacy preservation; (2) the introduction of a modified Hungarian algorithm that incorporates temporal proximity weights and adaptive health risk factors for multi-objective itinerary optimization; (3) the implementation of a lightweight convolutional neural network architecture optimized for edge deployment that achieves 92.3% accuracy while requiring only 15 MB of memory; and (4) the design of a fairness-aware federated learning protocol that maintains demographic parity while preserving privacy. The novelty of our approach lies in its holistic integration of these components into a unified framework that achieves sub-300 ms latency while processing complex multi-modal data streams, significantly advancing beyond existing systems that address these challenges in isolation.

The experimental results validate the system’s ability to dynamically adjust activity schedules based on real-time biometric feedback, reducing emergency interventions by 38% compared to conventional approaches. The priority-based scheduling algorithm proves especially valuable in balancing activity intensity with health constraints, as evidenced by the 92.7% health safety score across diverse test scenarios. Furthermore, the framework’s efficient resource utilization enables reliable operation on modest edge hardware, with memory usage remaining below 100 MB during normal operation.

Despite its promising results, this study has several limitations that should be acknowledged. First, the framework’s performance was evaluated primarily through simulated scenarios and controlled experiments, which may not fully capture the complexities of real-world deployment in diverse geographical and cultural contexts. Second, the current implementation assumes continuous wireless connectivity between wearable devices and edge nodes, which may not be achievable in remote adventure tourism locations with limited network infrastructure. Third, the federated learning component faces scalability challenges when expanding beyond specialized tour groups to mass-market applications, as participant diversity increases the complexity of modeling wider variations in physiological responses. Fourth, the system’s reliance on medical-grade wearables presents cost barriers for widespread adoption, though our device-agnostic approach supports a range of wearable tiers. Finally, the dynamic threshold system’s risk calculations require careful calibration to avoid reinforcing age-related biases, particularly in distinguishing between genuine physiological constraints and transient performance variations.

The system’s limitations in remote connectivity scenarios and demographic scalability point toward valuable directions for future research. Potential extensions include the development of offline-capable scheduling algorithms and demographic-aware personalization techniques to broaden the framework’s applicability. Ethical considerations around algorithmic transparency and bias mitigation also warrant further investigation to ensure the technology’s responsible deployment in real-world tourism settings.

Beyond elderly adventure tourism, the underlying architecture shows promise for various applications in smart tourism ecosystems and health-aware activity planning. The federated learning component’s privacy-preserving characteristics make it particularly suitable for international collaborations where data sovereignty is a concern. As demographic shifts continue to reshape global tourism patterns, such adaptive systems will become increasingly vital for delivering safe, engaging experiences tailored to diverse participant needs.