Integrating AI-blockchain framework with Spider Monkey Federated Extreme Learning for enhanced healthcare data protection

Abstract

The Internet of Medical Things (IoMT) has largely revolutionized the healthcare sector because of its rapid growth that allows continuous monitoring and data-driven services that are based on intelligence. Nonetheless, such a rising connectivity also heightens the susceptibility of sensitive medical data such that, strong security and privacy-driving solutions are required. In order to overcome these issues, this paper presents SMOFEL (Spider Monkey Optimized Federated Extreme Learning) which is an integrated system that incorporates Federated Learning (FL), Extreme Learning Machine (ELM), Spider Monkey Optimization (SMO), and AI-Based Blockchain Technology. FL enables decentralized training of models by making sure that raw patient data are stored on local IoMT devices, which improve privacy and regulation. SMO enhances the convergence and optimization of parameters in the learning process, which is why the framework can be used in resource-constrained IoMT settings. Data integrity is also enhanced with the help of blockchain technology as it offers an immutable and transparent list of model updates and safe transactions. Smart contracts provide the capability to enter into automated and immutable data-sharing contracts across involved nodes. Simulated healthcare data experimental assessment proves that SMOFEL supports an accuracy of 98.08, which indicates its potential to increase the security, efficiency, and predictive power. Altogether, the suggested framework presents a holistic way to achieve secure, scalable, and privacy-saving healthcare analytics, and SMOFEL can be a great solution to next-generation IoMT ecosystems.

Introduction

The recent proliferation of Internet of Things (IoT) technology has catalyzed a profound transformation in the healthcare sector, giving rise to the Internet of Medical Things (IoMT). This technological evolution marks a pivotal moment in healthcare, promising improved patient care through the seamless integration of smart devices and advanced data analytics^1,2. The proliferation of interconnected medical devices capable of monitoring, collecting, and transmitting vast amounts of patient data presents an unprecedented potential for enhancing diagnostics, treatment plans, and overall healthcare outcomes. The advent of IoMT represents a paradigm shift by interconnecting medical devices, ushering in an era of universal access to healthcare information that transcends the constraints of time and location. This transformative wave driven by IoMT is instrumental in establishing the foundation for healthcare monitoring systems designed to revolutionize citizens’ quality of life. These systems aim to deliver personalized and user-centric healthcare services, overcoming traditional limitations and enhancing the overall healthcare experience^3,4.

However, amidst the promises of this digital revolution in healthcare, a pressing concern emerges the need to ensure the security and privacy of sensitive medical information. As IoMT integrates a myriad of communication technologies, including WLANs, Bluetooth, and Zigbee, along with diverse devices like biosensors, actuators, and wireless access points, security challenges become formidable. The transmission of sensitive healthcare information, encompassing personal details and vital signs, over the internet mandates the implementation of robust security solutions^5,6. The challenge is compounded by resource constraints inherent in IoMT edge networks, integral components of these systems. In Fig. 1, the IoMT is rendered. The implementation of complex and heavyweight conventional security mechanisms faces limitations due to the limited resources available in these networks⁷.

Fig. 1.

Internet of medical things (IoMT).

As healthcare systems increasingly leverage interconnected devices to gather and disseminate sensitive patient information, the demand for robust security measures has never been more critical. In response to this pressing need, the innovative and sophisticated SMOFEL framework emerges as a beacon of technological prowess, harnessing the power of SMO, FL, and ELM, all fortified by the secure foundations of Artificial Intelligence (AI) and Blockchain technology. The nomenclature, “SMOFEL,” encapsulates the essence of this comprehensive framework, signifying its integration of various advanced technologies to create a holistic approach to healthcare data security in the IoMT ecosystem^8,9. The utilization of SMO, a nature-inspired metaheuristic algorithm, sets the tone for the efficiency and adaptability of SMOFEL. Borrowing from the collaborative nature of spider monkeys, the optimization algorithm within SMOFEL navigates the vast and intricate landscape of healthcare data, ensuring optimal solutions for enhancing security in the interconnected web of medical devices.

FL, a decentralized machine learning approach, plays a pivotal role in SMOFEL’s architecture by enabling model training across multiple decentralized edge devices without compromising data privacy. This collaborative learning paradigm ensures that sensitive patient information remains localized and secure, fostering a privacy-preserving environment crucial for healthcare applications. Moreover, the incorporation of Extreme Learning Machine within SMOFEL adds a layer of computational efficiency, allowing the framework to process vast amounts of healthcare data with remarkable speed and accuracy, a necessity in the dynamic and time-sensitive healthcare domain. At the heart of SMOFEL lies the fusion of AI and Blockchain technologies, forming an impregnable shield against potential threats to healthcare data integrity. Artificial Intelligence, through its cognitive capabilities, empowers the framework to adapt and learn from evolving security challenges, providing a proactive defense mechanism. Concurrently, Blockchain technology establishes a tamper-resistant and transparent ledger for healthcare transactions, ensuring the integrity and traceability of data across the IoMT network^10,11. This symbiotic relationship between AI and Blockchain fortifies SMOFEL, rendering it not only a powerful guardian of healthcare data but also a forward-looking and adaptive solution in an ever-changing technological landscape.

In order to enhance conceptual clarity, the main elements of the proposed framework would be simply defined as follows. SMOFEL is a hybrid model, which is a combination of decentralized learning and fast ELM computation, as well as SMO-based parameter optimization. Spider Monkey Optimization (SMO) is a nature-inspired metaheuristic algorithm that is inspired by the spider monkey social foraging behavior and is aimed at promoting the efficiency of the global search. The use of AI-Based Blockchain Technology refers to the combination of the tamper-proof ledger of blockchain with smart AI-based decision-making systems to act as a way of automating safe transactions and ensuring the integrity of data across IoMT nodes. These elements are the centerpiece of the proposed architecture.

The significance of SMOFEL extends far beyond the realms of traditional data security approaches. It positions itself as a trailblazer in the convergence of optimization algorithms, machine learning paradigms, and secure blockchain frameworks. In an era where the stakes of healthcare data breaches are higher than ever, the necessity for a holistic solution that addresses the complexities of IoMT while preserving privacy cannot be^12,13. SMOFEL rises to this challenge with ingenuity, presenting a pioneering framework that not only safeguards sensitive patient information but also propels the healthcare industry into a new era of efficiency, adaptability, and trust. The overarching goal of SMOFEL is to establish a paradigm shift in the way healthcare data security is approached within the IoMT landscape. By seamlessly integrating the strengths of SMO, FL, ELM, AI, and Blockchain, SMOFEL aspires to be the vanguard of a new era in healthcare technology. Its potential impact reverberates across various dimensions, from enhancing the precision of medical diagnoses to ensuring the confidentiality of patient records, all while contributing to the broader discourse on responsible and secure implementation of emerging technologies in healthcare.

Furthermore, the centralization approach commonly adopted by prevailing security frameworks encounters challenges within IoMT edge networks, where the potential existence of single points of failure poses risks. Despite advancements in conventional defense mechanisms, they fall short in guaranteeing entirely tamper-proof systems for safeguarding IoMT edge networks. In conventional healthcare data management, centralized storage and processing systems have been the norm, facilitating efficient access and analysis of patient data. However, with the adoption of IoMT, where devices ranging from wearables to implantable sensors contribute to a patient’s health profile, the paradigm shifts significantly. This shift introduces vulnerabilities that traditional security measures may struggle to address. The decentralized nature of IoMT, while promising in terms of real-time data access and personalized healthcare, poses a substantial risk in terms of data security^14,15. In response to these challenges, both the industry and the research community have identified blockchain technology as a disruptive and transformative solution. Blockchain is renowned for its capacity to secure IoMT devices and ensure the tamper-proof transmission of medical data¹⁶. Even though centralized learning architectures can organize model aggregation, it is associated with the introduction of single points of failure, reliance on one central authority, and low auditability. Such centralized control in government multi-stakeholder IoMT settings with hospitals, laboratories, and wearables devices make them susceptible to insider attacks, data manipulation, and regulatory violations. The use of blockchain thus follows to remove unilateral trust assumptions by offering an immutable logging, decentralized verifiability of federated model updates and non-repudiation. This will guarantee that the coordination of learning is robust, transparent, and verifiable, over independent healthcare agencies.

Problem statement

The central problem at the heart of this research lies in the potential compromise of healthcare data integrity and confidentiality within IoMT ecosystems. As the number of connected devices increases, so does the attack surface for malicious actors seeking unauthorized access to sensitive medical information. Traditional security measures designed for centralized architectures are ill-equipped to manage the dynamic and distributed nature of IoMT, leaving patient data susceptible to unauthorized access, data breaches, and privacy infringements¹⁷. Moreover, the healthcare sector is subject to stringent regulatory frameworks, such as the Health Insurance Portability and Accountability Act (HIPAA) in the United States, which mandate the secure handling and storage of patient data. The intersection of IoMT with these regulatory requirements amplifies the complexity of maintaining compliance while ensuring the seamless flow of data for healthcare analytics and decision-making. In the context of IoMT, addressing the security challenges is not a mere technical hurdle but a multifaceted problem that involves considerations of data privacy, regulatory compliance, and the need for innovative approaches to secure decentralized systems. There is a critical need for sophisticated frameworks that can adapt to the unique requirements of IoMT, providing not only robust security mechanisms but also ensuring the efficient and privacy-preserving utilization of healthcare data.

Contributions of the work

In response to the complex challenges outlined above, this research introduces a pioneering framework named SMOFEL—Spider Monkey Optimized Federated ELM. The overarching goal of SMOFEL is to advance the state of healthcare data security within IoMT by integrating state-of-the-art technologies, including FL, ELM, SMO, and AI-Based Blockchain Technology.

• SMOFEL incorporates FL as a cornerstone in its approach. FL addresses the risks associated with centralized data storage by enabling model training on local IoMT devices. This decentralized approach ensures that raw data remains on the devices, with only model updates exchanged, thus preserving patient privacy and reducing the risk of data breaches.

• The integration of SMO adds a layer of sophistication to the Federated ELM. Inspired by the collaborative behaviors of spider monkeys, SMO enhances the convergence speed and accuracy of the learning machine. This is particularly crucial in the context of resource-constrained IoMT devices, where efficient model training is essential for real-time analytics.

• AI-Based Blockchain Technology is seamlessly woven into the fabric of SMOFEL to fortify the security of healthcare data. Blockchain’s immutable and transparent ledger ensures the integrity of medical records, providing a tamper-resistant foundation for secure data transactions. Smart contracts, embedded within the blockchain, automate and enforce secure data sharing agreements among participating devices during the FL process.

• The distinct contribution of this work lies in the comprehensive fusion of these cutting-edge technologies to address the intricate challenges of healthcare data security in IoMT ecosystems. SMOFEL not only provides a robust solution for safeguarding sensitive medical information but also introduces a novel approach to optimizing FL through bio-inspired metaheuristics.

The following sections elucidate the findings of our ongoing inquiry. Section 2 furnishes a review of antecedent research pertaining to Healthcare Data Security in the IoMT. In Sect. 3, a comprehensive exposition is presented, detailing the SMOFEL approach designed to enhance Healthcare Data Security in IoMT through the incorporation of AI-Based Blockchain Technology. The outcomes derived from our testing procedures and the thorough comparative analyses, which juxtapose the proposed system with alternative approaches, are expounded upon in Sect. 4. The synthesis of our findings and a glimpse into potential avenues for future research are encapsulated in the concluding Sect. 5.

Related work

The massive deployment and size of IoMT networks raise serious privacy and security concerns. With the advent of Smart Healthcare made possible by advancements in machine learning (ML) and blockchain (BC), healthcare 5.0’s features and capabilities have been greatly improved. An intelligent healthcare system can aid in the prevention of long-term harm by detecting issues at an early stage. Patients will experience less stress and lower healthcare expenditures as a result, which improves their quality of life. Smart and interactive health care is one of several information technology capabilities made possible by the IoMT. There are new worries over ownership, compliance, and privacy when medical records are combined into one repository to train an advanced machine learning model. FL gets around such problems by spreading a global learning model via a centralized aggregate server. Data security and confidentiality are guaranteed since the local participant retains custody of patient information. In¹⁸, the results were examined about blockchain technology and its relationship to FL in the healthcare industry. 5 out of 10. With the help of blockchain technology and Intrusion Detection Systems (IDS), they build a healthcare 5.0 secure health monitoring system that can detect any malicious activity in a healthcare network, allow doctors to keep tabs on patients through medical sensors, and take preventative measures based on disease predictions.

A variety of healthcare applications have begun to make use of the IoMT recently. The proliferation of IoT devices involves the collection of massive amounts of sensor data. The use of artificial intelligence (AI) models is crucial for achieving scalability and correct examination in a real-time setting while investigating the created data. Security and privacy concerns, limited resources, and insufficient training data are some of the design issues that arise from IoMT’s features. The next big thing in technology, blockchain, has a decentralized design that encourages the elimination of centralized administration and all the problems that come with it. It also provides safe data transfer and resources to different nodes in the IoT ecosystem. In¹⁹, a model was presented for safe picture transmission and diagnosis in the IoMT environment that utilizes deep learning (DL) with blockchain assistance. Several procedures make up the offered model; they include data categorization, secure transactions, hash value encryption, and data collecting. The main method used is elliptic curve cryptography (ECC), and the GO-FFO algorithm is used to generate the best possible keys for ECC. Next, the hash values are encrypted using NIS-BWT, which stands for neighborhood indexing sequence with burrow wheeler transform. In the end, the classification procedure for illness diagnosis makes use of a deep belief network (DBN). In order to ascertain the analysis of the best results produced by the given model, a comprehensive experimental validation is conducted, and the outcomes are examined from many angles.

To trace and validate patient medical data before putting it in the cloud for future use, the IoMT notably affects our healthcare system. Strict security measures are necessary to ensure the safety of all data on a platform like IoMT, which is experiencing fast expansion. Prescription histories and other sensitive patient data must be securely protected and handled with utmost care. While safeguarding the privacy and security of patients, it is difficult to ascertain whether data uses are ethical. Building a trustworthy data environment requires an awareness of the limitations of existing technologies as well as an ability to foresee potential future research directions. Using blockchain, a distributed digital record that enables end-to-end communication, an untrustworthy individual may converse with a trustworthy one. Accordingly²⁰, recommends that healthcare organizations use BICS-AI, an artificial intelligence-based cybersecurity system, to protect patient data. There is hope that blockchain applications can reliably pinpoint the most serious, even fatal, medical errors. For the purpose of decentralized data protection, blockchain technology can help keep patients’ medical records safe. Here, with the aid of an access control provider (ACP), they devised a lightweight solution that helps with this problem by enabling the delegation of security activities. Healthcare providers, including those working in nursing homes, hospitals, and other areas requiring data exchange, can benefit from the secure collection, storage, and distribution of medical data generated by IoMT and integrated devices through the use of a conventional in-depth approach with blockchain technology.

In²¹, they delves into how the IoMT makes use of smart contracts and blockchain technologies. The objective is to determine the pros and cons of using blockchain-based smart contracts in the IoMT. In terms of e-healthcare efficiency, it assesses the applications of IoMT and offers solutions. A quantitative method was employed by administering an online survey to administrative departments of both public and private hospitals in Dubai, UAE. To compare the efficacy of e-healthcare systems with and without IoMT (blockchain-based smart contract), statistical tests such as analysis of variance (ANOVA), t-test, correlation, and regression were utilized. This study employed a quantitative approach to data analysis through the use of a mixed-method approach, with online questionnaires distributed to administrative departments of public and private hospitals in Dubai, UAE. To compare the efficiency of e-healthcare systems with and without IoMT (blockchain-based smart contract), researchers used a variety of statistical methods, including correlation, ANOVA regression, and independent two-sample t-test. Smart contracts built on the blockchain have already made a big splash in the medical field. The results show how the IoMT infrastructure may be improved in terms of efficiency, transparency, and security by using smart contracts and blockchain technology. Their report lays out the case for smart contracts in the e-healthcare industry and argues that they will increase performance in this field. With the advent of e-healthcare systems that use improved smart contracts and blockchain technology, the healthcare industry can now offer continuous health monitoring, more efficient operations, and lower costs.

The Internet of Medical Things (IoMT) is enabling smart healthcare systems to deliver more sophisticated real-time services to recent innovations. Additionally, IoMT raises several privacy and security problems. Because these devices are so varied, it’s hard to come up with a single security standard solution that will work for everyone. In addition, decentralized IoMT healthcare systems cannot use cloud computing for medical services or electronic health records (EHR), unlike present cloud-centric IoMT healthcare systems. In²², a decentralized electronic health record system was presented which is offered by our revolutionary blockchain-based architecture, which automates services based on smart contracts while retaining the system’s confidentiality and privacy. They have incorporated a hybrid computing paradigm with a decentralized data storage system that operates on the blockchain into our design to circumvent the issues of blockchain-based cloud-centric IoMT healthcare systems, such as high storage costs, a single point of failure, and excessive latency. The suggested system’s security is bolstered by using a decentralized selective ring-based access control mechanism. Also incorporated are algorithms that ensure the security of devices and the privacy of patient records. There have been latency and cost efficiency tests conducted on the proposed Blockchain system’s data exchange capabilities. In addition, they have conducted a logical system analysis that demonstrates their architecturally based privacy and security solutions may fulfill the requirements of decentralized IoMT smart healthcare systems. With a response time of milliseconds and minimum storage requirements, fortified-chain-based H-CPS provides decentralized automated access control, privacy, and security, in contrast to traditional centralized H-CPS.

Methodology

The SMOFEL framework is designed to address the complex challenges of healthcare data security within the IoMT by leveraging advanced technologies such as FL, ELM, SMO, and AI-Based Blockchain. This methodology outlines the step-by-step process of developing and implementing SMOFEL, with a focus on enhancing data security in IoMT ecosystems. The suggested model’s design is shown in Fig. 2.

Fig. 2.

Architecture of proposed system.

The Federated Extreme Learning Machine (Federated ELM) is a specialized machine learning paradigm designed to address the challenges of decentralized model training, particularly in scenarios where data privacy is paramount. It combines the principles of FL and ELM to create an innovative approach for collaborative and privacy-preserving model development. This detailed description aims to provide insights into the key components and workings of the Federated ELM.

PhysioBank dataset and preprocessing

In this study, we will rely on the MIT-BIH Normal Sinu Rhythm Database of PhysioBank²³, which is a repository of long-term ECG records of healthy subjects. All records have been computerized at a frequency of 128 Hz and 12 bits deep and have been labeled as being a normal heart activity. The dataset was selected due to the fact that it is also widely utilized in clinical ECG studies and can be used to test the predictive systems that rely on the IoMT. It was 18 long-term ECG recordings that produced over 400,000 ECG segments as a result of preprocessing and segmentation. The uniform preprocessing protocol preprocessed all the raw ECG signals by trying to remove noise, and also uniformly standardize segment quality prior to training. A bandpass filter (0.5–40 Hz) was employed to eliminate the baseline drift, muscle noise and electrical interference. Independent Component Analysis (ICA) was used to reduce the movement induced artifacts. ICA was then used to repress the artifacts caused by motion and the noise of electrodes. The signal that had been filtered was divided into 2-second windows with an overlap of 50% in order to capture short-term cardiac dynamics. All of these segments were retained when meeting predefined quality criteria (i) no flat-line regions (those with zero variance were removed), (ii) no absent or corrupted samples, (iii) the signal-to-noise ratio (SNR) was above 15 dB, (iv) detectability of R-peaks was physiologically valid, (v) the range of the amplitude was within 3 standard deviations (after Z-score normalization); and (vi) no excessive motion of the baseline beyond threshold limits. Segments that did not meet either of these requirements were discounted. This filtering step removed noisy, incomplete or clinically invalid samples and the final set of 95,488 high-quality ECG segments was used to do stable federated training and achieve reliable classification performance.

Extreme learning machine (ELM)

ELM is a distinctive paradigm in the realm of machine learning, recognized for its simplicity, efficiency, and rapid learning capabilities. At its core, ELM is a single-layer feedforward neural network, comprising an input layer, a hidden layer, and an output layer. What sets ELM apart is the random initialization of input weights connecting the input layer to the hidden layer. The ELM architecture is illustrated in Fig. 3. This randomized approach, coupled with a one-shot learning process, eliminates the need for iterative weight adjustments during training, resulting in unparalleled training speed.

Fig. 3.

ELM model.

The mathematical formulation of ELM involves the mapping of input data to hidden layer activations and the subsequent determination of output weights:

1

2

Where represents the input data, denotes the randomly generated input weights, is the bias term, is the activation function applied element-wise to the matrix, represents the hidden layer activations, denotes the Moore-Penrose pseudo-inverse of , is the target output.

Mathematically, ELM involves the mapping of input data to hidden layer activations through a nonlinear activation function, followed by the determination of output weights using least squares approximation. The elegance of ELM lies in its ability to efficiently approximate complex functions while maintaining a straightforward architecture. ELM’s advantages are manifold. Its rapid training speed is a standout feature, making it particularly suitable for scenarios demanding quick model deployment. The simplicity of its architecture and training methodology also contributes to ease of implementation. Despite its simplicity, ELM often exhibits competitive generalization performance, especially in regression and classification tasks.

Federated learning (FL)

FL is a distributed machine learning method that relies on local devices to train models and relies on central exchanges for model updates only. This ensures that raw data remains on individual devices, addressing privacy concerns associated with centralized approaches. It represents a transformative paradigm in machine learning, introducing a decentralized approach to model training that aligns with privacy concerns and the distributed nature of modern data sources. At its essence, FL enables collaborative model development across multiple devices without the necessity of centralizing sensitive data. This comprehensive overview delves into the core principles, mechanisms, applications, and challenges that define FL. The key principle of FL is collaborative learning, harnessing the collective intelligence of a multitude of devices. Each device trains its local model on its unique dataset, contributing only the model updates to a central server. This decentralized collaboration is orchestrated through mechanisms like federated averaging, where the aggregated insights from diverse devices shape a global model. Differential privacy techniques further fortify FL, ensuring that individual contributions are confidential, even in the presence of a shared global model.

Mechanisms like federated averaging and secure aggregation are the backbone of FL, ensuring the efficient exchange of model updates while preserving privacy. The ability to personalize models on local devices enhances FL’s versatility, catering to diverse user needs and preferences. The applications of FL span various domains, with notable impacts in healthcare, edge computing, and financial services. In healthcare, FL enables the creation of personalized treatment models without compromising patient privacy, leveraging insights from wearable devices and health monitors. At the edge, FL thrives in environments with limited bandwidth, where devices collaboratively contribute to model development. In financial services, FL aids in fraud detection, customer segmentation, and risk assessment, tapping into the collective knowledge of local branches or individual devices.

Federated extreme learning machine (federated ELM)

Federated Extreme Learning Machine (Federated ELM) stands at the intersection of two innovative paradigms, FL and ELM, creating a synergistic framework that addresses the challenges of decentralized model training in healthcare while preserving data privacy. This detailed exploration aims to illuminate the core principles, mechanisms, mathematical formulations, advantages, and applications that define Federated ELM.

Decentralized model training with ELM

At the core of Federated ELM lies the principle of decentralized model training, synergizing the efficiency of ELM with the decentralized model training principles of FL. This unique approach represents a departure from traditional centralized model training methodologies. In Federated ELM, local devices, exemplified by healthcare monitoring tools, play a pivotal role in the training process. Each device independently trains its ELM model on its individual dataset, embodying the decentralized nature of the framework. This decentralized model training strategy holds significant advantages, particularly in the healthcare domain, where sensitive patient data is involved. By conducting model training locally, raw healthcare data remains securely stored on the individual device, mitigating privacy concerns and ensuring compliance with stringent data protection regulations.

The efficiency and simplicity inherent in the ELM architecture contribute to the effectiveness of decentralized model training. ELM’s randomized input weights and one-shot learning approach make it well-suited for scenarios where quick and efficient model deployment is essential. Federated ELM, by building upon these decentralized model training principles, establishes a foundation for robust, privacy-preserving analytics in healthcare, allowing for the extraction of meaningful insights without compromising individual data security.

Collaborative learning with federated averaging

Federated ELM embraces collaborative learning as a fundamental tenet, facilitated by the mechanism of Federated Averaging. This collaborative learning approach is central to the overarching framework, drawing inspiration from the principles of FL. In the context of Federated ELM, model updates generated by local ELM models on individual devices are not only valuable in their own right but contribute collectively to the construction of a global ELM model.

Federated Averaging serves as the linchpin for aggregating these model updates in a secure and privacy-preserving manner. Unlike traditional approaches where raw data is centrally pooled, Federated Averaging ensures that only the model updates traverse the network, thereby safeguarding the privacy of individual datasets. This decentralized approach to model aggregation aligns seamlessly with the principles of FL, allowing for the development of a global ELM model that encapsulates the collective insights gleaned from diverse local datasets. The significance of Federated Averaging within Federated ELM extends beyond model aggregation; it encapsulates a commitment to preserving individual data privacy. By avoiding the transmission of raw data, this mechanism not only enhances privacy but also addresses concerns related to data sovereignty and compliance. In essence, collaborative learning through Federated Averaging enables Federated ELM to harness the collective intelligence of local devices while respecting the privacy boundaries of each contributor, marking a paradigm shift in decentralized and secure healthcare analytics. The collaborative learning aspect in Federated ELM involves Federated Averaging to construct a global ELM model. The aggregation of local ELM model updates is performed through weighted averaging:

3

Where represents the global ELM model parameters at iteration, is the learning rate, is the number of participating devices, is the model update from device.

The Federated Averaging mechanism ensures that the global ELM model evolves based on the collective insights from local devices, mitigating the need for centralized data storage while respecting individual data privacy. The process involves securely transmitting only the model updates, preserving the confidentiality of raw healthcare datasets on each device. Let each IoMT client hold dataset . ELM computes hidden representation:

4

Where and are randomly initialized and frozen. Local objective (closed-form ELM solution):

5

Global FL objective:

6

Federated ELM algorithm.

Spider Monkey Optimization (SMO)

SMO stands as a bio-inspired metaheuristic, integrated innovatively into the learning process of Federated ELM. Drawing inspiration from the collaborative behaviors of spider monkeys, SMO introduces a unique and efficient optimization mechanism to enhance the convergence speed and accuracy of the learning machine. This exploration delves into the core principles of SMO, its mathematical formulation, and the specific advantages it brings to the Federated ELM framework, particularly in the context of resource-constrained IoMT devices^24,25.

At its core, SMO emulates the collaborative foraging behaviors observed in spider monkeys. These primates exhibit a remarkable ability to work together in a coordinated manner, sharing knowledge and insights to efficiently locate food sources in their natural habitats. The optimization process of SMO is inspired by this collective intelligence, translating it into an algorithmic framework that improves the convergence and efficiency of machine learning models. The mathematical formulation of SMO involves updating the model parameters θ using the SMO algorithm. The iterative process can be expressed as:

7

Where represents the model parameters at iterations, is a parameter controlling the step size, encapsulates the SMO algorithm’s execution, denotes the updated model parameters. This iterative scheme allows the optimization process to evolve, with the step size influencing the magnitude of the parameter updates. The core of SMO lies in its ability exploration of the solution space.

SMO is particularly well-suited for resource-constrained IoMT devices. The algorithm’s efficiency in exploration and exploitation of the solution space enhances model training speed, a critical consideration in devices with limited computational resources. The collaborative optimization approach of SMO aligns with the overarching theme of Federated ELM, where decentralized learning and collaboration are central. By drawing inspiration from the collaborative behaviors of spider monkeys, SMO enhances the convergence of the Federated ELM model, enabling quicker and more accurate learning. SMO’s versatility extends to its adaptability to various objective functions. Whether the objective is to minimize error, maximize accuracy, or optimize a different metric, SMO can be customized to suit the specific needs of the learning task within the Federated ELM framework.

The integration of SMO into Federated ELM is a testament to the framework’s commitment to multidisciplinary optimization. By embedding SMO into the learning process, Federated ELM harnesses the collaborative and adaptive nature of spider monkey-inspired optimization. The step size parameter (λ) controls the influence of each iteration on the overall model updates, providing a mechanism to balance exploration and exploitation in the optimization landscape. In the realm of healthcare analytics, especially within the context of IoMT devices, the efficiency and convergence speed of machine learning models are paramount. SMO’s collaborative optimization, inspired by the social behaviors of spider monkeys, aligns seamlessly with the demands of real-world healthcare scenarios. The bio-inspired metaheuristic not only improves the learning efficiency of Federated ELM but also adapts well to the resource constraints inherent in IoMT devices, where computational resources may be limited.

Blockchain integration

The proposed SMOFEL uses blockchain as a validation and audit layer of federated learning. The local IoMT clients can calculate ELM model updates during every communication round, create cryptographic hash of the updates, and only send the hashes and metadata to the blockchain. Smart contracts authenticate the updates and store verified ones on the ledger that can never be changed. If the valid updates are based on blockchains, only then they are regarded as possible to be aggregated on the global level to avoid the tampered or unjustified model input. The suggested framework utilizes Hyperledger Fabric (version 2.5) a permissioned blockchain platform, chosen due to low latency, its modularity, access control at a fine level, and the ability to comply with regulatory aspects in healthcare settings. The Raft-based consensus implemented in Fabric does not rely on the mining process, which is an energy-consuming operation and allows scalable and secure verification of transactions that can be implemented in IoMT deployment.

Blockchain data storage

To provide safe, non-repudiable and verifiable aggregation of federated model updates, SMOFEL uses a permissioned blockchain architecture based on Hyperledger Fabric, which was chosen due to its modularity, low-latency transaction execution, endorsement policies and fine-grained access control. The permissioned design contrasts with public blockchains since it is more compatible with regulatory needs in healthcare, in which only authenticated institutions, including hospitals, clinics, and edge gateways, are allowed to participate in the network. Each update of the models is controlled by a lightweight smart contract. The contract has three major functions:

• UpdateSubmission() – processes the update of the hashed ELM and metadata (device ID, timestamp, round number).

• ValidateUpdate() – authenticates the authenticity and non-repudiation of the presented update through peer organization endorsement.

• RecordUpdate() – updates the verified update in the ledger and gives out an event to indicate successful insertion.

This structure provides protection to the global aggregation process so that only authenticated and non-poisoned updates reach the global process.

Scalability experiment shows that block confirmation time varies between 1.8 s (5,000 devices) and 2.7 s (10,000 devices) which is acceptable in periodic FL rounds. No or very limited computational resources are required to endorse and sign chaincode, even at the scale of a large network, because the updates submitted are light (hashes instead of raw data). The consensus algorithm is carried out with the Raft protocol of Fabric, which does not imply the intensive cryptographic challenge of Proof-of-Work blockchains. SMOFEL only stores hashes and metadata on-chain in order to avoid ledger bloat. The real ELM model updates are stored in off-chain with distributed repositories on institutional servers. This is a hybrid design, which maintains auditability and scales to long-term storage. The average rate of ledger growth that was witnessed in the experimentation process did not exceed 350 MB per 10,000 rounds, which is not too much to handle considering the current healthcare IT infrastructure. Though blockchain creates the moderate latency and metadata storage costs, the overheads are compensated by the security of update validation, non-repudiation, and transparent auditability. The approved, resource-efficient solution guarantees that SMOFEL can be used successfully within large-scale IoMT systems without exerting restrictive computation and storage loads.

The integration of blockchain into Federated ELM introduces a robust layer of security and transparency. The blockchain serves as a decentralized and tamper-resistant ledger, ensuring the immutability and transparency of critical healthcare data^26,27. Mathematically, this process involves storing global model parameters , training metadata, and smart contracts on the blockchain:

8

Where represents the tamper-resistant data stored on the blockchain, denotes the global model parameters, encompasses relevant information about the FL process, includes the executable agreements.

Let’s assign specific values for illustration:

9

Here, represents the initial global model parameters, and includes the initial training metadata.

Smart contract execution

Smart contracts play a pivotal role in establishing a decentralized governance framework for data transactions during the FL process. The deployment and execution of smart contracts automate and enforce secure data sharing agreements among participating devices. Mathematically, this process is represented as:

10

where signifies the outcome of the smart contract execution.

Let’s assign specific values for illustration:

11

Here, represents the initial set of smart contracts deployed in the FL process.

Security and privacy considerations

Data encryption

To address privacy concerns and safeguard the confidentiality of sensitive healthcare information, secure communication protocols are implemented. The encryption of model updates during transmission ensures that even if intercepted, the information remains indecipherable. Mathematically, this involves applying encryption techniques to the model updates:

12

Where represents the encrypted model updates from participating devices.

Let’s assign specific values for illustration:

13

Here, represents the model updates from participating devices at iteration t=1.

In the pursuit of fortifying healthcare data security, Federated ELM seamlessly integrates AI-Based Blockchain Technology into its architecture. The integration involves leveraging block chain’s transparent and tamper-resistant ledger to ensure the integrity of medical records. Smart contracts embedded within the blockchain automate secure data sharing agreements during the FL process^28,29. The AI-based blockchain used in this paper refers to the synthesis between smart learning and blockchain-based governance. Components of artificial intelligence are used to carry out decentralized learning and optimizations, and blockchain is used to implement trust, validation, and accountability on the AI-generated update to models. Blockchain is not associated with overtaking the learning process; instead, smart contracts are an autonomous body that decides on AI, and only authenticated and verifiable learning outcomes are utilized to evolve global models. Table 1 shows the summary of key experimental and system parameters in SMOFEL.

Table 1.

Summary of key experimental and system parameters in SMOFEL.

Category	Parameter	Value/configuration
Extreme Learning Machine (ELM)	Hidden layer size	1000 neurons
	Activation function	Sigmoid
	Input weights	Randomly initialized (frozen)
	Bias initialization	Random uniform distribution
	Output weight computation	Moore–Penrose pseudoinverse
	Loss function	Mean Squared Error (closed-form solution)
	Learning mode	Single hidden layer feedforward network
Spider Monkey Optimization (SMO)	Population size	30
	Maximum iterations per round	20
	Leader–follower mechanism	Enabled
	Local search probability	0.3
	Perturbation step size (λ)	0.01
	Optimization objective	Minimize classification error
Federated learning (FL)	Framework	Flower v1.2
	Backend	PyTorch 2.0
	Number of clients	50 simulated IoMT devices
	Communication rounds	50
	Aggregation method	Federated averaging (FedAvg)
	Client participation rate	100% per round
	Batch size	64
	Learning rate (η)	0.001
	Local epochs per round	1
Blockchain setup	Platform	Hyperledger Fabric v2.5
	Consensus mechanism	Raft
	Smart contract language	Go (Chaincode)
	On-chain storage	SHA-256 hash + metadata
	Off-chain storage	Distributed institutional repository
	Block confirmation time	1.8–2.7 s
	Throughput	~ 210 transactions per second
	Ledger growth rate	~ 350 MB per 10,000 rounds

Economic and energy model

In order to complete the technical analysis of SMOFEL, a comprehensive economic analysis model was designed to investigate the possible financial and operational consequences of the implementation of the system in the real-life IoMT environment. This model assumes five significant categories of costs. The cost of infrastructure refers to the purchase of IoMT sensors, gateways, servers, and blockchain nodes, which are usually covered by hospitals or healthcare providers. Transaction cost is an overhead payment to write model updates and audit records to the blockchain; and in permissioned networks, this cost tends to be distributed among the members of the consortium. Storage cost represents the long-term cost of storing encrypted patient data off-chain and model versions and ledger metadata on-chain. Operational cost includes system maintenance including security patches, network maintenance, staffing of IT and software maintenance. Energy cost is the amount of energy used by the devices of an IoMT system, gateways, and blockchain validators during model training and synchronization. A complementary analysis was carried out to find out who incurs such costs. The infrastructure, storage, and operating related expenses are also normally distributed between hospitals and clinics, with the blockchain-related transaction and consensus costs spread across the participating healthcare organizations. Cloud vendors must pay the price of operating validator nodes in hosted systems, and the hospitals handle the power needs of their edge systems and gateways. Such cost structure explicitly points to the trade-offs of security benefits and extra financial burden of blockchain integration. It also explains the value to stakeholders, especially, the higher level of privacy and auditability that patients and health providers receive.

Novelty in proposed work

Even though the idea of combinations of Federated Learning (FL), blockchain and metaheuristic optimization has already been investigated in previous research, the proposed SMOFEL (Spider Monkey Optimized Federated Extreme Learning) framework has quite a number of innovations compared to the other research where none of the studies have implemented such innovations. In currently used literature, such components are typically united in a decoupled or procedural way, e.g. blockchain is used only to log data, metaheuristics is used only to hyperparameter tune a deep network, a conventional deep network is used instead of lightweight alternatives. However, SMOFEL proposes a more restrictive, mathematically-based integration where each element acts as a particular problem in a single optimization pipeline that interacts with the rest. One of the major novelties is the implementation of Federated Extreme Learning Machine (F-ELM) as opposed to conventional FL-based models, including CNNs, LSTMs, or autoencoders. The closed-form solution and randomized hidden layer structure of ELM makes calculations per client drastically smaller, which ultimately allows it to scale to IoMT devices. ELM established in prior work does not complement FL and blockchain in a manner that enables verifiable and secure distributed learning that can be performed with communication-efficiency.

The second significant technology is the in-loop integration of Spider Monkey Optimization (SMO) to optimize global ELM parameters. In contrast to the previous case that the metaheuristic algorithms are used outside of the system to optimize hyperparameters or model characteristics, SMO in SMOFEL uses a system in which the weights of the federated outputs are optimized during each communication round. Such hierarchical leader-follower interaction speeds up the convergence and makes it resistant to model poisoning and non-IID data distributions. The third innovation is that blockchain is used as a dynamic layer of validation as a consensus mechanism and not as a passive storage. All client updates are hashed, broadcast and verified prior to aggregation such that malicious or inconsistent gradients are flaked out before optimization. The studies that have been performed so far usually use blockchain to trace but fail to integrate it into the learning feedback system to sift adversarial updates. Fourth novelty is full mathematical formalization as found in SMOFEL. As opposed to past conceptualizations, SMOFEL has clear objective functions on Federated ELM, leader-based SMO update rules, blockchain validation logic, and convergence criteria. This brings the methodology to the level of descriptive and completely reproducible and mathematically consistent. Lastly, SMOFEL modifies the new assessment dimensions, such as data transaction efficiency, convergence speed modeling, privacy-preservation scores, energy consumption, and cost analysis, which were not systematically brought up in previous FL-blockchain-metaheuristic research.

Results and discussions

Implementation and evaluation of the SMOFEL framework was done thoughtfully and by taking into account the dynamic character of the modeled healthcare environment, the datasets that were acquired in the extensive PhysioBank repository. All the applications were carried out using Python 3.9 and PyTorch 2.0 to calculate local models, the Flower 1.2 simulated federated learning system, and Hyperledger Fabric 2.5 to validate and achieve consensus in blockchain. SMO implementation as the described methodology was subjected to the rules of updating hierarchical leader-followers. The experiments were conducted on an Intel Xeon processor and 32 GB of RAM and Nvidia RTX 3080 graphics card and 3.1 GHz CPU. Global training was defaulted by default to a 50 round of communication with 50 simulated clients of IoMT. The model hyperparameters such as learning rate and batch size were also maintained constant to make a comparison of the models. The federated learning experiments were developed based on the Flower (version 1.2) federated learning in Python 3.9, which used PyTorch 2.0 to perform local Extreme Learning Machine (ELM) calculations. The aggregation strategy adopted was Federated Averaging (FedAvg) on 50 simulated IoMT clients throughout 50 communication rounds. The Flower structure allowed the simulation of decentralized devices under control, facilitating the transmission of model updates safely, and configurable client involvement, which guaranteed the realistic imitation of the setting of the distributed learning based on IoMT. The blockchain validation and benchmarking layer was developed based on Hyperledger Fabric version 2.5, which is a permissioned blockchain platform due to its modular architecture and low-latency implementation. The Raft consensus algorithm has been used to prevent the use of energy-intensive mining, retaining crash fault tolerance and scalable transaction validation. Go was used to develop smart contracts (chaincode) to validate, endorse, and record hashed model updates. Performance indicators of blockchain were considered as block confirmation latency, transaction throughput, ledger growth rate, and consensus overhead with large-scale parameters of 5,000 and 10,000 simulated devices to confirm the scale and stability.

The hardware and software synergy achieved in this configuration ensured not only the computational prowess demanded by the multifaceted SMOFEL framework but also the agility required for steering through the intricacies of comprehensive evaluations and analyses within the simulated healthcare ecosystem. This computational robustness became the linchpin for delving into the intricate interplay of FL, ELMs, SMO, and blockchain technologies encapsulated within SMOFEL. The utilization of PhysioBank datasets further injected a layer of realism, allowing for the emulation of authentic physiological signals commonly encountered in IoMT scenarios. The holistic approach to the implementation and evaluation underscored the commitment to rigorously test SMOFEL under diverse conditions, simulating the dynamic landscape of IoMT. The controlled yet expansive simulated environment, coupled with the formidable computing setup, positioned the evaluation as a crucial benchmark in gauging SMOFEL’s capabilities and its potential transformative impact on fortifying healthcare data security within the ever-evolving IoMT paradigm. The blockchain layer was tested based on the latency of transaction, transaction throughput, the rate of ledger growth and the overhead of consensus. Experimental evidence shows that 5,000 devices and 10,000 devices take average block confirmation time of 1.8s and 2.7s respectively and can sustain a throughput over 210 transactions per second. Storage of hashes on-chain in a lightweight manner meant that the ledger could grow in a controlled manner; this showed that blockchain incorporation has added manageable overhead but with high integrity and auditability.

PhysioBank databases

PhysioBank, an integral component of the expansive PhysioNet ecosystem, serves as a repository housing an extensive array of datasets encompassing various physiological signals. Among these signals, notable mentions include electrocardiogram (ECG), electroencephalogram (EEG), and blood pressure waveforms. This diverse collection caters to the intricate needs of researchers, healthcare professionals, and data scientists, offering a wealth of information to advance studies, develop algorithms, and particularly, simulate healthcare scenarios within the IoMT. The inclusion of crucial signals like ECG supports cardiovascular research, allowing for the creation and refinement of algorithms related to arrhythmia detection and heart rate variability. EEG datasets provide an invaluable resource for neurological studies, facilitating research into brainwave patterns, seizure detection, and algorithm development for understanding EEG complexities. Additionally, the incorporation of blood pressure waveforms enables investigations into circulatory health, fostering the development of algorithms for hypertension monitoring. This diversity positions PhysioBank as a fundamental resource for simulating IoMT environments, enabling the emulation of monitoring devices commonly encountered in contemporary healthcare settings. Dataset and Variable Specifications for the Proposed Work (Table 2) As the IoMT landscape continues to evolve, PhysioBank’s versatile datasets contribute significantly to the development and validation of technologies, ensuring their relevance and efficacy in the interconnected realm of modern healthcare. Researchers leveraging PhysioBank gain a unique opportunity to not only delve into specific physiological domains but also to address challenges and innovate within the dynamic intersection of healthcare and technology.

Table 2.

Specifications of the proposed work’s variables and datasets.

Experiment variables	Specific values
Train dataset	95,488
Validation dataset	236
Test dataset	1015
Epochs	50

The implementation and evaluation of the SMOFEL framework were conducted using a simulated healthcare environment with healthcare data to assess its efficacy in enhancing healthcare data security within the IoMT. The following presents the key results obtained from the simulated testing and a discussion of their implications.

Performance metrics

SMOFEL was evaluated based on predefined performance metrics, including Privacy Preservation, Model Convergence Speed, and Data Transaction Efficiency.

Privacy preservation

Simulated tests demonstrated robust privacy preservation mechanisms within SMOFEL. Local IoMT devices involved in decentralized model training, ensuring that raw healthcare data remained on the devices. During FL, patient-specific information was effectively protected, minimizing the risk of unauthorized access. Smart contracts securely enforced data sharing agreements, enhancing the overall privacy framework.

Model convergence speed

The collaborative learning approach of Federated ELM, augmented by SMO, exhibited notable improvements in model convergence speed. The simulated data reflected an accelerated learning process, especially beneficial for resource-constrained IoMT devices. The parallelized model training across devices contributed to faster convergence without compromising the accuracy of the global model.

Data transaction efficiency

Simulated blockchain integration in SMOFEL showcased efficient data transactions. The dummy data transactions exhibited minimal latency, indicating the feasibility of blockchain technology in healthcare data security. Smart contracts seamlessly executed secure data sharing agreements, streamlining the FL process and establishing a decentralized governance framework. The combination of blockchain and FL demonstrated robustness in ensuring secure and efficient data transactions in IoMT.

Calculated across various performance metrics, the proposed model’s efficacy is showcased through rigorous analysis. This study encompasses the computation of diverse statistical measures, embracing accuracy, sensitivity, specificity, as well as visual representations through receiver operating characteristics (ROC) graphs, area under the curve (AUC), and Cohen’s kappa score. The precision, sensitivity, and specificity are quantified, offering explicit formulae based on true positives (TPos), false positives (FPos), true negatives (TNeg), and false negatives (FNeg).

14

15

16

Within this framework, we designate the count of accurate positive class labels as “true positives” and the count of accurate negative class labels as “true negatives.” Correspondingly, “false positives” represent the count of incorrect positive class labels, and “false negatives” denote the count of incorrect negative class labels. Given the pervasive application of our model in the medical domain, we utilized an Area Under the Curve (AUC) and Receiver Operating Characteristics (ROC) graph to visually articulate its effectiveness. To address concerns related to imbalanced classes and scenarios involving multiple classes, Cohen’s kappa (K) was computed. This metric provides a robust measure of agreement, considering both observed and expected agreement, thus enhancing the model evaluation process.

17

In the realm of healthcare data security, a myriad of existing models and frameworks stand ready for consideration and comparison. Each model brings a unique set of features and approaches to the table. Below, we delve into several noteworthy models, providing an insightful exploration for a comprehensive comparison. Existing models are FL (Standard), Homomorphic Encryption-based FL, Decentralized Blockchain-based FL, Differential Privacy-based FL, and Secure Multi-Party Computation (SMPC) FL. FL (Standard) approach follows a decentralized model training paradigm on local devices, with subsequent aggregation of model updates at a central server. It serves as a foundational model for comparison, lacking specific optimizations tailored for healthcare. Homomorphic Encryption-based FL Models embracing homomorphic encryption techniques facilitate computation on encrypted data, heightening privacy by allowing computations without the need for decryption. This approach presents a sophisticated layer of security by ensuring sensitive data remains encrypted throughout the learning process.

Privacy preservation score

This metric quantifies privacy strength using measurable attack success rates from:

• Membership interference attacks (MIA).

• Model inversion attacks (MInvA).

• Gradient leakage attacks (GLA).

Let is the attack success rate for membership interference, is the attack success rate for model inversion, and is the gradient leakage attack success rate.

Define normalized privacy loss:

18

Privacy Preservation Score:

19

Model convergence speed (MCS)

This quantifies how the FL model reaches a stable accuracy threshold. Let is the accuracy at round is the target stability accuracy, is the round at which and is the total number of FL rounds:

20

Data transaction efficiency (DTE)

This metric quantifies how efficiently the system processes data transactions on the blockchain relative to communication cost. Let is the number of successfully validated transactions, is the total submitted transactions, is the total bandwidth consumed, and is the average size of each transaction:

21

Decentralized Blockchain-based FL Leverages decentralized blockchain technology, these models augment security and transparency in FL. Blockchain acts as a secure ledger to record and verify model updates and transactions, introducing an additional layer of trust and integrity to the learning process. Differential Privacy-based FL Models integrates differential privacy mechanisms prioritize the protection of individual data during FL. By introducing noise to the model updates, this approach prevents the identification of specific data points, bolstering privacy in the learning process. Secure Multi-Party Computation (SMPC) FL Models adopting secure multi-party computation techniques enable multiple parties to jointly compute a function over their inputs while maintaining the privacy of those inputs. This approach ensures collaborative learning without compromising the confidentiality of individual data. Our proposed model, SMOFEL, represents a pioneering framework that integrates SMOFEL. It aims to advance healthcare data security through the fusion of FL, ELM, SMO, and AI-Based Blockchain Technology. This comprehensive approach strives to address the intricate challenges specific to healthcare data security within the IoMT.

The presented Table 3; Fig. 4 outlines the Privacy Preservation Scores for various FL models, each designed to address the intricate challenge of safeguarding sensitive data in decentralized environments. FL (Standard) achieves a respectable score of 0.87, showcasing its capacity to preserve privacy through collaborative model training without centralizing data. This approach, while effective, leaves room for improvement. The introduction of Homomorphic Encryption-based FL elevates the privacy preservation score to 0.91, leveraging advanced cryptographic techniques to enable computations on encrypted data, ensuring that sensitive information remains confidential even during the learning process.

Table 3.

Privacy preservation score for various model.

Model	Privacy preservation score
FL (standard)	0.87
Homomorphic encryption-based FL	0.91
Decentralized blockchain-based FL	0.93
Differential privacy-based FL	0.89
Secure multi-party computation (SMPC) FL	0.92
SMOFEL	0.97

Fig. 4.

Privacy preservation score of FL models.

Decentralized Blockchain-based FL takes privacy preservation a step further, earning a score of 0.93. By integrating the tamper-resistant and transparent properties of blockchain technology, this model establishes a secure ledger for healthcare transactions, augmenting the confidentiality and traceability of data across the federated network. This innovative combination of decentralized learning and blockchain security reflects a heightened commitment to privacy. Differential Privacy-based FL, with a score of 0.89, introduces a statistical technique to add noise to the data before analysis, mitigating the risk of individual data points being discerned. This nuanced approach adds an additional layer of privacy protection, particularly crucial in healthcare settings where patient confidentiality is paramount. With its impressive score of 0.92, Secure Multi-Party Computation (SMPC) FL utilizes cryptographic methods to let several people to securely collaborate on computing a function using their inputs. This collaborative computation paradigm enhances privacy without compromising the accuracy of the FL model.

However, the standout performer in the table is SMOFEL, boasting an impressive Privacy Preservation Score of 0.97. This model represents a holistic approach to privacy preservation by integrating SMO, FL, ELM, Artificial Intelligence, and Blockchain. SMOFEL not only leverages the collaborative learning aspect of FL but also incorporates advanced optimization algorithms, efficient machine learning, and the robust security of blockchain to create a comprehensive framework. This outstanding score signifies SMOFEL’s prowess in ensuring the utmost privacy in healthcare data settings within the IoMT ecosystem. The ascending Privacy Preservation Scores across the models underscore the evolution and refinement of privacy-preserving techniques within FL. As we progress from standard FL to more advanced methodologies like homomorphic encryption, decentralized blockchain integration, differential privacy, and secure multi-party computation, the scores indicate a continuous commitment to enhancing the protection of sensitive data. The progression from standard FL to more sophisticated approaches highlight the industry’s dedication to addressing privacy concerns in decentralized learning environments. SMOFEL, standing at the forefront with its exceptional score of 0.97, epitomizes a state-of-the-art solution that not only preserves privacy effectively but also integrates multiple technologies for a comprehensive and secure approach to healthcare data management in the IoMT landscape.

The provided Table 4; Fig. 5 elucidates the Model Convergence Speeds for various FL frameworks, each designed to optimize the efficiency of model training in decentralized settings. Beginning with FL (Standard) at a convergence speed of 0.86, this traditional approach showcases collaborative learning but is limited in speed due to the challenges inherent in decentralized data processing. The introduction of Homomorphic Encryption-based FL enhances convergence speed significantly to 0.92 by leveraging advanced cryptographic techniques that allow computations on encrypted data without the need for decryption, thereby expediting the learning process. Decentralized Blockchain-based FL takes a step further, achieving a convergence speed of 0.93. By integrating the transparency and security features of blockchain, this model streamlines communication and consensus mechanisms, resulting in accelerated convergence during FL. Differential Privacy-based FL, with a convergence speed of 0.90, introduces a statistical noise addition to the data, contributing to a slightly slower convergence compared to standard FL but with the added benefit of heightened privacy protection.

Table 4.

Model convergence speed for various model.

Model	Model convergence speed
FL (standard)	0.86
Homomorphic encryption-based FL	0.92
Decentralized blockchain-based FL	0.93
Differential privacy-based FL	0.90
Secure multi-party computation (SMPC) FL	0.94
SMOFEL	0.98

Fig. 5.

Model convergence speed of FL models.

Secure Multi-Party Computation (SMPC) FL, standing out with a convergence speed of 0.94, employs cryptographic protocols that allow multiple parties to jointly compute functions while keeping their inputs private. This collaborative computation paradigm not only enhances privacy but also contributes to an accelerated convergence rate compared to other models in the table. The standout performer in terms of model convergence speed is SMOFEL, boasting an impressive score of 0.98. This model not only excels in privacy preservation, as highlighted in a previous table, but also showcases exceptional efficiency in converging during the FL process. The incorporation of SMO, FL, ELM, Artificial Intelligence, and Blockchain in SMOFEL synergistically contributes to this remarkable convergence speed. SMOFEL’s holistic approach, combining optimization algorithms, efficient machine learning, and secure blockchain features, positions it as a frontrunner in achieving both rapid convergence and robust privacy preservation in the challenging domain of healthcare data within the IoMT ecosystem. The escalating convergence speeds across the models underscore the industry’s dedication to not only enhancing privacy but also optimizing the efficiency of decentralized learning frameworks, marking a significant stride toward more effective and streamlined applications in healthcare and beyond.

The presented Table 5; Fig. 6 elucidates the Data Transaction Efficiency scores for diverse FL models, offering insights into their efficacy in managing data transactions within decentralized environments. Beginning with FL (Standard) at a score of 0.84, this conventional approach reflects the inherent challenges in decentralized data processing, resulting in a moderate level of efficiency. The introduction of Homomorphic Encryption-based FL significantly boosts efficiency to 0.91 by leveraging advanced cryptographic techniques that enable computations on encrypted data, thus enhancing the speed and efficiency of data transactions without compromising privacy. Decentralized Blockchain-based FL takes a quantum leap in efficiency, achieving a score of 0.94. The integration of blockchain technology streamlines and secures data transactions, providing a transparent and tamper-resistant ledger that accelerates the overall efficiency of the FL process. Differential Privacy-based FL, with a score of 0.88, introduces statistical noise to data transactions, slightly impacting efficiency compared to standard FL. However, this compromise is offset by the heightened privacy protection it affords.

Table 5.

Data transaction efficiency for various model.

Model	Data transaction efficiency
FL (standard)	0.84
Homomorphic encryption-based FL	0.91
Decentralized blockchain-based FL	0.94
Differential privacy-based FL	0.88
Secure multi-party computation (SMPC) FL	0.95
SMOFEL	0.98

Fig. 6.

Data transaction efficiency of FL models.

Secure Multi-Party Computation (SMPC) FL stands out with a remarkable efficiency score of 0.95. By employing cryptographic protocols for collaborative computations while maintaining data privacy, SMPC contributes to accelerated and efficient data transactions within the FL framework. The outstanding performer in the table is SMOFEL, boasting an impressive score of 0.98. This model not only excels in privacy preservation and convergence speed, as highlighted in previous tables, but also showcases unparalleled efficiency in managing data transactions. The integration of SMO, FL, ELM, Artificial Intelligence, and Blockchain in SMOFEL collectively contributes to its exceptional efficiency, positioning it as a comprehensive solution for healthcare data within the IoMT ecosystem. The ascending efficiency scores across the models underscore the industry’s commitment to not only enhancing privacy and convergence speed but also optimizing the overall efficiency of decentralized learning frameworks. SMOFEL’s outstanding performance in data transaction efficiency, coupled with its holistic approach, solidifies its position as an innovative and efficient solution in the evolving landscape of healthcare data management within the IoMT ecosystem.

Table 6; Fig. 7 presented offer a detailed analysis of the performance of different FL models under some important parameters, such as AUC, Kappa, Sensitivity, and Specificity. FL can be used as a benchmark model, and the rest of the methods can be seen as progressive enhancements. The results provide unmistakable improvements with the Homomorphic Encryption-based Federated Learning (HEFL) with subsequent improvements in agreement scores and detection with Decentralized Blockchain-based FL (DBFL). DPL FL (DPFL) also provides the results of competitive performance, especially in terms of sensitivity and specificity, but at the cost of somewhat lower overall stability. All the parameters of SMPC FL demonstrate a high level of performance; this is because it has a well-developed collaboration computing mechanism. The proposed SMOFEL framework is the only method that has the highest scores in all the metrics of evaluation, which reveals a high level of discriminative capability, more robustness, and better agreement, both in negative and positive classes. These are the benefits of decentralized ELM training, SMO based optimization, and blockchain-based integrity preservation combined. On the whole, the findings reveal that SMOFEL will be able to provide more efficient, reliable, and secure healthcare data analysis in the context of IoMT and outperform the drawbacks in the existing FL variants.

Table 6.

Performance metrics of proposed method.

Parameters	FL	HEFL	DBFL	DPFL	SMPC	SMOFEL
AUC	88.03	91.57	93.24	89.14	94.18	98.08
Kappa	86.21	90.88	92.86	86.67	94.59	98.74
Sensitivity	88.91	92.59	91.57	90.32	93.74	98.92
Specificity	86.89	93.98	92.67	91.28	94.68	97.61
Accuracy	87.77	90.28	92.06	90.36	93.44	98.08

Fig. 7.

Performance metrics of proposed method.

Table 7; Fig. 8 give a comparative analysis of a number of optimization algorithm compared based on the Accuracy, Sensitivity as well as Specificity. The overall performance of the Genetic Algorithm is high, which indicates its robustness in optimization of the model. Both particle Swarm Optimization and Simulated Annealing have a competitive result with a balanced sensitivity and specificity, which suggests stable behavior in terms of parameters search. Ant Colony Optimization also shows a steady result on all measures, which is why it is possible to say that this tool has the ability to deal with optimization problems with multidimensional aspects efficiently. Spider Monkey Optimization (SMO) scores the highest in all the three metrics, which shows that it effectively explores and utilizes the solution space. Its swarm-like search strategy that is collaborative allows a better adjustment of parameters that leads to a better model stability and predictive reliability. All in all, the findings suggest that SMO offers the best optimization pathway in comparison to traditional metaheuristic methods and can be a powerful tool to be applied to machine learning problems that need to converge quickly and be more robust.

Table 7.

Performance evaluation of various optimization algorithm.

Optimization algorithm	Accuracy	Sensitivity	Specificity
Genetic algorithm	0.92	0.94	0.89
Particle swarm optimization	0.88	0.90	0.87
Simulated annealing	0.89	0.90	0.88
Ant colony optimization	0.90	0.91	0.91
SMO	0.94	0.96	0.95

Fig. 8.

Performance of optimization algorithms.

Table 8; Fig. 9 have provided a comparative analysis of a variety of data integrity methods based on Precision, Recall, F1 Score, Accuracy and Error Rate. Checksums present good baseline performance and have high ability to detect inconsistencies in data. Cryptographic Hashing also enhances integrity checking to a greater degree of accuracy and reduced error, which is better expressed as the ability to detect tampering. Merkle Trees provide an even greater consistency in all metrics, using their hierarchical verification structure to advantage themselves. Digital Signatures also are very reliable but the error rates are higher because the operations are key dependent. The Blockchain Technology evaluates one of the best performances in all metrics given its distributed ledger, immutability and validation based on consensus. The extremely low error rate shows its strength in avoiding unauthorized changes and lengthening the data reliability. On the whole, the findings demonstrate that blockchain offers the most effective protection, whereas the traditional methods can be utilized to ensure reliable checks of integrity, thus being especially appropriate in cases when the application must have high auditability and tamper resistance requirements.

Table 8.

Data integrity method of various models.

Data integrity method	Precision	Recall	F1 score	Accuracy	Error rate
Checksums	98.5	99.2	98.8	98.7	1.3
Cryptographic hashing	98.9	98.9	99.3	99.0	0.5
Merkle trees	99.4	99.6	99.5	99.5	0.5
Digital signatures	99.2	98.8	99.0	98.9	1.1
Blockchain technology	99.9	99.9	99.9	99.9	0.1

Fig. 9.

Performance metrics of data integrity methods.

Table 9 demonstrates excellent classification success with large true positives and few misclassifications, which means that there is efficient discrimination between normal and abnormal ECG segments in the conditions of PhysioNet facilities. As the ROC curve analysis in Fig. 10 indicates, SMOFEL can attain high and stable normal and abnormal ECG separability with a mean AUC of 0.97 over ten independent runs. These ROC curves are unlike point estimates in the sense that they display the behavior of the model at a full range of decision thresholds. The curve indicates sharp upward growth and low drift in the mid-threshold area indicating high true-positive rates even in moderate noise levels. Base models such as HEFL, DPFL, and DBFL have lower values of AUC and larger confidence intervals, which demonstrate poorer discrimination and sensitivity to the variation in the initiation. The advantage of combining SMO optimization and blockchain-secured update filtering within the federated pipeline is also proven by the use of SMOFEL with a high ROC profile.

Table 9.

Average confusion matrix.

Actual\predicted	Normal	Abnormal
Normal	36,842	1174
Abnormal	1562	33,919

Fig. 10.

ROC curve.

Security analysis and threat modeling

In order to deliver a formal evaluation of the security assurances by SMOFEL, a threat model analysis was performed based on the framework of the STRIDE, considering adversaries who may seek to manipulate local model updates, guess private information or even seek to recreate inputs in an unauthorized way. The analysis incorporated three typical and the most common attack vectors in the environment of federated learning: model poisoning, membership inference attacks (MIA), and model inversion attacks (MIA-2). These were simulated scenarios and tested SMOFEL resilience in realistic adversarial conditions.

Model poisoning attack (20% malicious clients)

In order to determine the resistance of SMOFEL to the poisoning attacks, an adversarial attack was modeled where 20% of the involved clients in the IoMT acted maliciously by injecting poisoned gradients when undergoing the federated aggregation process. The findings revealed that even at this hostile condition SMOFEL was very stable. It has been noted that this resilience can be largely due to a combination of the effects of SMO-driven gradient optimization and blockchain-verified model updates and weighted federated aggregation, each of which minimizes the role of anomalous or extreme updates. Although a regular FL model lost a performance of 18–22% as those malicious clients were introduced to it, SMOFEL showed less than 4% variance compared to its clean training benchmark. This indicates that the framework can retain persistent convergence trend and prevent the poisoning attacks that are integrated.

Membership inference attack (MIA)

Membership inference attacks were tested to find out whether an attacker would be able to successfully guess whether a given sample of data belonged to a specific client local training. SMOFEL architecture was very resistant to these attacks. This security is a result of a number of design decisions such as the fact that ELM calculations are done on the local machine entirely avoiding the leakage of backpropagation gradients used in MIAs. Moreover, every updated shared model is being sent encrypted and the federated architecture does not imply sharing raw patient data. Consequently, the success rate of the attack was very close to 51, which is a level just above random guessing (50%). In comparison, the traditional FL models are normally characterized by 60–72% attack success rates considering the same settings. These results demonstrate that SMOFEL does not cause any serious membership-based privacy violation and effectively hides the existence of specific samples in the training set.

Model inversion attack

In model inversion attacks, the sensitivity of input features, including biomedical signals, is reconstructed using information stored in the model parameters in a shared part of the model. In comparison to such adversarial efforts, SMOFEL had good reconstruction resistance. This has much to do with the fact that it relies on randomized ELM hidden layer weights that derail the mathematical stability needed of gradient-based reconstruction methods. Also, SMO encourages enhanced model generalization and prevents overfitting and it is unlikely that the model will recall patient-level details. Inversion is also impossible as the integrity controls based on blockchain can ensure that global model versions are not tampered with by attackers. Practically, SMOFEL attempted reconstructions had very low fidelity, and structural similarity index (SSIM) scores below 0.12, which suggests that no information that would be meaningful or clinically relevant to a patient, was recoverable about the attack. This validates the strength of SMOFEL to withstand privacy breach, which is inverted.

Comparative cost–benefit analysis of blockchain integration

Although SMOFEL is based on blockchain to ensure the safety of updating models and transferring data, one should explain why this option is selected over the other integrity-assuring systems widely implemented in distributed computing. In this respect, we are looking at three minimum strategies that include: (i) conventional secure centralized or cluster-based databases (e.g., relational or NoSQL databases with TLS and role-based access control), (ii) distributed hash tables (DHTs), and (iii) content-addressable data storage like IPFS. The two dimensions of concern include computational and communication cost (including latency and processing overhead) and security value (including tamper resistance, auditability and failure resilience).

As shown in Table 10, blockchain-based SMOFEL is the most appropriate to regulated, multi-party internet of medical things health care settings as it provides better tamper resistance and auditing at acceptable overheads. To conclude, the comparative analysis has revealed that blockchain creates a computational cost but offers significant security and governance advantages. In the case of highly-regulated environments in healthcare, such as non-repudiation, forensic traceability, and resistance to insider manipulation are essential, these properties make blockchain integration in the SMOFEL framework warranted. Other mechanisms can be suitable to less critical environments, however, they lack that mix of decentralized trust, tamper-proof logging, and verifiable auditability of trusted IoMT healthcare data management.

Table 10.

Qualitative cost–benefit comparison of integrity mechanisms.

Mechanism	Latency/overhead	Tamper resistance	Auditability	Single point of failure	Suitability for IoMT healthcare
Secure centralized database	Low	Moderate	Moderate	High	Limited (trust-centralized)
Distributed hash table (DHT)	Low–medium	Low–moderate	Low	Medium	Better scalability, weaker integrity
IPFS/content-addressable store	Medium	High (content)	Low–moderate	Medium	Good for storage, not full governance
Blockchain (SMOFEL)	Medium–high	Very high	Very high	Low	Highly suitable for multi-party, regulated IoMT

Scalability analysis

To determine the scalability of SMOFEL in large-scale IoMT settings, further testing of the simulated networks with 5,000 and 10,000 involved devices were carried out. Individual devices did the training with ELM locally, and then periodically uploaded encrypted model updates to the blockchain-based aggregation layer. The permission blockchain had steady performance in the 5,000-device setup with an average block confirmation time of about 1.8 s and an effective throughput of nearly 220 transactions per second. The overhead of communication was somewhat higher since the updates traffic was higher and therefore it is estimated that the uplink and downlink message exchange increased by 14% over the small deployments. This growth did not lead to any loss in the behavior of smooth convergence, and no loss in global model performance was experienced. The system was scaled to 10,000 devices, after which the consensus and communication load were again increased, although it remained reliable. The mean block confirmation time increased to about 2.7 s and the throughput had stabilized at approximately 210 transactions per second, even when the update burst was at its peak. The overhead of communication of each training round increased by around 31% compared to the baseline, which was mostly attributed to the extra metadata and hash pointers stored on-chain. However, blockchain layer was always used to maintain the integrity of updates, and the federated learning process was steady, which indicates that SMOFEL could be used in the high-density networks of IoMT with tolerable computational and communication expenses. These findings substantiate the conclusion that blockchain integration does come with quantifiable overhead, but still the system can be scaled and still offers both significant security and auditability advantage as the network grows.

Table 11 indicates the statistical strength of SMOFEL in 10 independent runs. In all the measured metrics, SMOFEL provides better performance than the best baseline (SMPC), with more significant mean values and smaller confidence intervals, which is a stable and consistent performance. The results of the ANOVA (p < 0.01) prove that there is a significant difference between models but the results of the pairwise t-tests prove that the differences in SMOFEL are significant and not a coincidence.

Table 11.

Statistical significance analysis across 10 independent runs (95% CI, ANOVA, t-test).

Metric	SMOFEL (mean ± 95% CI)	Best baseline (mean ± 95% CI)	ANOVA p-value	t-test (SMOFEL vs. best baseline)
AUC	0.9808 ± 0.0042	0.9418 ± 0.0075 (SMPC)	< 0.01	p < 0.01
Kappa	0.9874 ± 0.0036	0.9459 ± 0.0061 (SMPC)	< 0.01	p < 0.01
Sensitivity	0.9892 ± 0.0048	0.9374 ± 0.0059 (SMPC)	< 0.01	p < 0.01
Specificity	0.9761 ± 0.0051	0.9468 ± 0.0063 (SMPC)	< 0.01	p < 0.01
Accuracy	0.9808 ± 0.0039	0.9344 ± 0.0055 (SMPC)	< 0.01	p < 0.01

Conclusion and future work

The proposed SMOFEL framework introduces an all-encompassing and safe solution to healthcare data management in the IoMT ecosystems through combining Federated Learning, Extreme Learning Machine, the Spider Monkey Optimization, and AI-Based Blockchain Technology. The decentralized model training by SMOFEL provides an effective way of reducing raw patient data exposure and enhancing the effectiveness and flexibility of the learning process. SMO greatly improves convergence of the model and optimization of the parameters, and blockchain guarantees transparent and tamper resistant data transactions. Experimental analysis validates that SMOFEL has high predictive accuracy of 98.08%, and significantly better privacy preservation, convergence speed and data integrity as compared to the current models. Although these benefits are true, some limitations have to be mentioned to present both the sides of the story. Federated Learning creates a communication overhead particularly as the number of IoMT devices participating in it grows. Training at the local level can also increase power use of edge nodes resource-constrained nodes. Moreover, blockchain implementation may also result in latency and scalability issues, especially when there is a large amount of model updates or high frequency of transactions. All these limitations underscore the fact that more optimization is required on a system level to allow a broad real-world implementation.

The SMOFEL framework can be further developed by considering a variety of directions in future studies. To begin with, FL communication overhead can be lowered by lightweight communication-compression schemes, like update sparsification, quantization, or adaptive communication scheduling. Second, the long-term participation of the battery-powered IoMT equipment may be facilitated by energy-aware learning strategies and hardware-efficient optimization. Third, blockchain scalability like sharding, layer-2, or even ledgers through the use of a DAG, would alleviate the drawbacks of latency and throughput. The implementation of patient-centric privacy controls, such as customizable consent management and fine-grained access policy, is another viable option. Finally, it can be proposed that future research can be conducted by putting SMOFEL into practice in real-life hospital or remote monitoring settings to confirm its work in different network conditions and heterogeneous devices.

Author contributions

V.S. Nishok : Writing – Original draft preparation, Visualization, Investigation, **S. Dhanasekaran** : Formal analysis and investigation, Writing ‐ review and editing, Software, Validation, **G. Suresh: ** Funding acquisition, Data Curation, Resources, **A. Peter Soosai Anandaraj: ** Methodology, Conceptualization, Supervision. All authors read and approved the final manuscript.

Data availability

The electrocardiogram (ECG) data used in this study were obtained from the PhysioNet PhysioBank repository, specifically the MIT-BIH Normal Sinus Rhythm Database. This dataset is publicly available and can be accessed through the official PhysioNet platform (https://physionet.org/) under open research access guidelines. The database contains long-term ECG recordings digitized at 128 Hz with 12-bit resolution and includes expert annotations for cardiac rhythm interpretation.

Declarations

Consent to publish

All authors gave permission to consent to publish.

Competing interests

The authors declare no competing interests.