An Adaptive Blockchain Framework for Federated IoMT with Reinforcement Learning-Based Consensus and Resource Forecasting.

Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Feb 11;16:8296. doi: 10.1038/s41598-026-35704-1 Search in PMC Search in PubMed View in NLM Catalog Add to search An Adaptive Blockchain Framework for Federated IoMT with Reinforcement Learning-Based Consensus and Resource Forecasting CH V N U Bharathi Murthy CH V N U Bharathi Murthy 1 School of Computer Science Engineering and Information Systems, Vellore Institute of Technology, Vellore, 632014 India Find articles by CH V N U Bharathi Murthy 1 , M Lawanya Shri M Lawanya Shri 1 School of Computer Science Engineering and Information Systems, Vellore Institute of Technology, Vellore, 632014 India Find articles by M Lawanya Shri 1, ✉ Author information Article notes Copyright and License information 1 School of Computer Science Engineering and Information Systems, Vellore Institute of Technology, Vellore, 632014 India ✉ Corresponding author. Received 2025 Jul 11; Accepted 2026 Jan 7; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC12966370  PMID: 41667513 Abstract The rapid growth of the Internet of Medical Things (IoMT) has increased the adoption of remote healthcare applications and telemedicine services. A Massive amount of sensitive healthcare care is being gathered daily by IoT devices. . Managing the continuous flow of data streams while maintaining low latency, scalability and security remains as a challenge in traditional IoMT architectures. These problems lead to delays in real-time healthcare decision-making which is critical and increases system overhead. . To address these problems, the proposed work introduces a novel framework that integrates machine learning with blockchain-based federated IoT clouds, customised for an efficient and secure platform for handling healthcare data. The framework includes Gradient Boosting Machines (GBM) for Intelligent data storage optimisation, which analyses historical access patterns and real-time data, improving cache hit rates by 25% and reducing read latency by 30%. The system performance is improved by Deep Q-Learning (DQN), ensuring resource management. Due to optimisation, the maximum CPU load is reduced by 20% and improved management by 15%. Convolutional Autoencoders are used to improve privacy and security. These helped improve anomaly detection by 95% and reduce false positives by 10%. Long Short-Term Memory (LSTM) network improves the rate ofresource utilisation prediction to 90%, and proactive resource management achieves a 25% reduction in latency spikes. The Adaptive Byzantine Fault Tolerance (ABFT) consensus protocol with Reinforcement Learning(RL), is integrated to improve transaction efficiency and dynamically adapts the consensus parameters. The proposed integration results in a 40% improvement in transaction throughput and a 20% reduction in transaction latency. In comparison of ABFT-RL consensus with PBFT and Raft consensus under similar workloads, the proposed ABFT-RL enhanced throughput by 43% and decreased end-to-end latency by 31%, offering improved scalability and responsiveness. A private blockchain network called Hyperledger Fabric is considered. In this proposed work, the optimised output of each layer is fed into the next layer, and this seamless flow of data gives an efficient architecture managing the complexities of the blockchain-based federated IoT cloud. Keywords: Hyperledger fabric, Blockchain, Federated IoT, Telemedicine, Deep Q-Learning, Gradient boosting Subject terms: Computational biology and bioinformatics, Engineering, Health care, Mathematics and computing Introduction The IoT devices collect large amount of sensitive healthcare data, which in turn leads to increase in data generation and network complexity. The current remote patient monitoring and telemedical services are made possible by the development of the Internet of Medical Things (IoMT). The distributed computing resources of Federated IoT Clouds offer a novel solution for securely and efficiently handling large amounts of data. By integrating Blockchain technology with federated healthcare clouds, we face many challenges that may prevent data integrity and privacy, while maintaining real-time performance and efficient resource utilisation. The traditional Blockchain with IoT clouds has proven to have poor system performance, delayed operations, and inefficient resource allocation, which are necessary for healthcare data to enable quick decisions and deliver rapid patient care.Unlike traditional RSA algorithm and ECC-based cryptographic mechanisms, which are vulnerable to quantum attacks, the recent research studied post-quantum and quantum-resistant blockchain mechanisms that are considered to secure IoT and IIoT networks. . Post-quantum key encapsulation, Lattice-based signatures, and hash-based authentication have shown significant resistance to quantum threats, making them ideal for long-term medical data storage. Quantum-resistant consensus mechanisms have been proposed to protect lightweight IoT devices while reducing communication overhead. These studies highlight the need for scalable and adaptive security models in healthcare systems, thereby requiring an intelligent and adaptable consensus mechanism like the ABFT-RL model with a private blockchain called Hyperledger Fabric is proposed in this paper. .. It can provide access control and enhanced privacy features. In heterogeneous IoMT systems, system performance and scalability are significant concerns due to dynamic workloads. This paper presents a novel intelligent heterogeneous framework for blockchain-based federated IoT clouds, providing secure and real-time healthcare services, including remote monitoring, diagnostics, and telemedicine. In the proposed framework, each integrated technology has its own responsibility within the system. The proposed model performs intelligent data storage optimisation using Gradient Boosting Machines (GBM) to review historical access patterns and runtime data samples. This results in minimal read latency, improved cache hit ratios, and trustworthy access to patient health records. Deep Q-Learning (DQN) is integrated to provide real-time CPU and memory resource allocation based on real-time workload metrics and data location. This results in optimised memory usage, minimising peak CPU usage and system latency, which helps to support the continuous remote monitoring and telemedical services. . . Convolutional Autoencoders helps in detection of anomalies, reducing false positives. Long Short-Term Memory (LSTM) networks are integrated to lower the latency spikes for proactive resource management. To achieve maximum transactional throughput and reduce latency in time-sensitive telemedicine services, the proposed model integrates the Adaptive Byzantine Fault Tolerance (ABFT) consensus algorithm with Reinforcement Learning(RL), allowing the system to dynamically adjust consensus parameters automatically. The proposed unified framework is designed to provide a privacy-based, high-performing platform that supports advanced healthcare services. Current telemedicine services require continuous remote monitoring without compromising data integrity or system effectiveness. This proposed study aims to demonstrate that integrating Gradient Boosting Machines (GBM), Deep Q-Learning (DQN), Long Short-Term Memory (LSTM) networks, Convolutional Autoencoders, and Adaptive Byzantine Fault Tolerance (ABFT) with reinforcement learning in the Hyperledger Fabric blockchain network can improve performance, reliability, and privacy. The primary objectives of this research are to reduce data access latency and enhance cache efficiency in federated IoT Clouds by integrating Gradient Boosting Machines (GBM) for forecasting. It also aims to optimise real-time resource utilisation by using Deep Q-Learning (DQN) for dynamic CPU and memory allocation in response to varying IoMT workloads. To implement adaptive resource scaling and reduce latency by integrating with Long Short-Term Memory (LSTM) networks for time-series resource forecasting. To increase blockchain transaction throughput and also reduce consensus latency through implementing the Adaptive Byzantine Fault Tolerance (ABFT) mechanism with Reinforcement Learning for auto-adjusting consensus parameters. These objectives collectively provide a secure, scalable, and high-performance platform for remote healthcare, facilitating low-latency diagnostics, reliable access to medical data, and continuous monitoring in a federated IoT Cloud Environment. In this work, the contributions are as follows: GBM-Guided Predictive Caching: With GBM technology, a new caching layer is introduced to predict which medical records will be used frequently by analysing historical and real-time queries. Using this method, the data can be stored in high-priority storage, which leads to minimising read latency and cache misses. DQN-Driven Resource Scheduling: A Deep Q-learning scheduler helps in the reallocation of CPU and memory resources efficiently among federated cloud nodes. This scheduler can respond to variable workload and anomaly load. This way, we prevent runtime overload and reduce congestion during teleconsultation. ABFT-RL Consensus Optimisation : An increase in throughput and a reduction in confirmation delay can be achieved using the ABFT-RL consensus layer. This layer adjusts to the block size, endorsement threshold, and timeout parameters. Real-Time Anomaly Detection: In this, the identification of irregular or malicious patterns can be achieved by Convolutional Autoencoders by monitoring encrypted data streams and enhancing cyber resilience during remote monitoring. Proactive Resource Forecasting: Long Short Term Memory networks predict resource utilisation patterns, enable adaptive scaling and reduce latency spikes during continuous remote diagnosis. Despite the increasing use of GBM, DQN, LSTM, Autoencoders, and ABFT techniques, they are not employed in their conventional form. In this work, each module is re-parameterised, retrained and integrated into a single workflow tailored for federated healthcare cloud environments. The GBM caching layer functions on encrypted access-log features instead of unprocessed data, and also enables privacy-preserving cache prediction. The DQN scheduler is trained to penalise high latency and peak CPU usage with a custom reward function that allows real-time resource allocation in federated edge cloud environments. The ABFT consensus is integrated with a Reinforcement learning agent to automatically adjust the endorsement thresholds, block size, and timeout, thereby increasing throughput under varying workloads. By integrating these modules, the system brings a closed-loop optimisation pipeline that brings reduced latency and increased throughput. Integrating these heterogeneous systems into the proposed framework substantially improves the performance, reliability, and privacy of remote healthcare and telemedical services. This article is divided into four sections. Section " Literature survey " presents a literature survey, Section " Proposed hybrid model integrating federated IoT Cloud With Blockchain " provides the mathematical proof and architecture, and Section " Experimental setup " discusses the environmental setup and performance metrics. Finally, Section " Results and Discussion " provides the results. Literature Survey The integration of Blockchain with Federated IoT cloud has gained significant attention in modern literature, as it offers a potential chance for improved performance, security, and efficiency. This hybrid model is essential in telemedicine and remote patient health monitoring, where the confidentiality, integrity, and trustworthiness of healthcare data are of paramount importance. Multiple approaches have suggested that integrating Blockchain into healthcare would bring advantages, such as decentralised data that is more privacy-sensitive. This paper surveys various studies by analysing their techniques, approaches, findings, and outcomes. The primary focus is on models that have adopted the Internet of Medical Things (IoMT) and telemedicine services. The current research spans from the conventional blockchain framework to the integration of advanced technologies, including federated learning and deep learning. For example 1 , used a blockchain-based federated learning (BCFL) system, which utilises a Multi-Layer Perceptron (MLP) model trained with federated learning and secured by blockchain technology. In this system, smart contracts are used for aggregation. In the paper 2 , the authors introduced a secure-sharing scheme using heterogeneous federated learning and threshold identity authentication to address problems such as privacy and resource limitations in telemedical data. Lodha et al. 3 developed a Blockchain-based IoMT Security System (BC-IoMT-SS) to improve the security and privacy of the patient data. Li et al. 4 designed a data-sharing method for a remote monitoring healthcare system by integrating federated learning with a consortium blockchain. They have employed a reputation-based incentive mechanism to deter malicious attacks and promote honest participation. Singh et al. 5 developed a framework to improve the privacy of IoT healthcare data using federated learning and blockchain technology. This framework is designed to restrict off-site training models, ensuring data integrity and preserving patient privacy. Barbosa et al. 6 implemented federated learning in resource-constrained IoT environments by gradually increasing the number of clients and local training epochs. They achieved improved accuracy but were unable to optimise system performance. In 7 , a blockchain-based federated learning model for secure data transfer among healthcare IoT networks was developed. In their framework, local IoT devices train models independently using their data and share them through a blockchain network, ensuring data integrity and transparency. Within Blockchain, smart contracts are used to automate the validation and integration of local and global models. Even then, they faced challenges like computational overhead and scalability issues. Khan et al. 8 achieved an improved diagnostic accuracy and security by designing the Healthcare Federated Ensemble Internet of Learning Cloud Doctor System (FDEloL). In 9 , they integrated Blockchain with federated learning to address security challenges in in-home health monitoring systems. Wassan et al. 10 developed a technique to improve the performance and accuracy of machine learning models in Health Internet of Things applications. They integrated the gradient boosting mechanism into a federated learning framework, which allows data to train models independently on local devices, ensuring data privacy. They have also faced computational overhead challenges when using gradient boosting algorithms, which resulted in poor device performance. In Lin et al. 11 , the authors designed a public integrity verification system within an IoT cloud environment, which enhanced security, data integrity, and latency. Okegbile et al. 12 evaluated the performance of blockchain-enabled data sharing to enhance throughput and provide better security. Deebak et al. 13 and Nguyen et al. 14 have given several algorithms for secure and efficient task scheduling in Healthcare IoT environments. Samriya et al. 15 and Hewaet al. 16 employed adaptive algorithms based on reinforcement learning to adjust consensus parameters for improved performance dynamically, but they encountered issues such as computational overhead and implementation complexity. The comparisons are shown in Tables 1 and 2 . Hasan et al. 17 showed that Blockchain can solve unauthorised access and provide guaranteed data provenance in healthcare. Indeed, many studies, such as 16 , 18 , which employed deep reinforcement learning, have demonstrated the effectiveness of dynamic resource allocation and adaptive consensus mechanisms in enhancing the system’s robustness and performance. Along with traditional blockchain security measures, recent studies have explored quantum-resistant blockchain solutions specifically designed for IoT and IIoT environments. Lattice-based digital signatures, hash-based authentication, and post-quantum key encapsulation have been introduced as alternatives to RSA and ECC, providing long-term security in the face of quantum threats in Villegas-Ch et al. 19 . The integration of deep learning with post-quantum blockchain has been explored to strengthen cyber defence mechanisms within IoT networks 20 . Additionally, advancements in hash-based ring signature schemes have been made to facilitate secure off-chain transactions in IoT environments 21 . The results suggest that future IoMT and IIoT systems must incorporate cryptographically robust blockchain methods to defend against quantum-enabled attacks. Table 1. Comparison Table of Literature Survey. Citation Approach Findings Results Limitations 11 Consortium Blockchain Public integrity verification for IoT cloud storage Enhanced data integrity and security Potentially high latency 1 Multi-layer perceptron model Blockchain with federated learning to enhance data privacy and security Improved data privacy and ensures data integrity Need for effective consensus, High computational overhead. 3 Blockchain-based IoMT Security System (BC-IoMT-SS) Blockchain with IoMT to enhance security and privacy High precision ratio and reduced response time Scalability, interoperability 5 Local training of Machine learning models on IoT devices Federated learning enabling blockchain machine learning Improved privacy and highly accurate prediction Scalability, interoperability 7 Smart contract automation with federated deep learning network Blockchain with federated deep learning Enhanced security and privacy, reduced mitigated risks Scalability, high computational overhead 8 Healthcare Federated Ensemble Internet of Learning Cloud Doctor System (FDEIoL) Ensemble learning integrates with IoT for a robust global prediction model Precise diagnosis and outperformed centralised models Complexity 22 IoHT devices train machine learning models with automated smart contract Combining Federated learning with blockchain to maintain privacy Enhanced security and privacy, reduced reliance on central systems Computational and storage overhead 10 Federated learning framework for HIoT environments Gradient boosting in FL frameworks Enhanced predictive performance Computational overhead, and scalability 23 Blockchain Framework Secure healthcare applications in IoT Enhanced security and scalability Interoperability challenges 2 Secret Sharing and Collaborative Blockchain Secure IoT data storage in clouds Improved security and distributed storage High resource consumption 24 Lattice-based certificateless proxy re-signature scheme Enables secure key delegation and re-signing without exposing private keys Lower signature size and reduced resource consumption Did not address latency and blockchain consensus overhead. 19 Lighweight blockchain framework for authentication Secure identity management and access control with minimum storage Reduced authentication delay Lacks evaluation in large scale IoT networks. 21 Post-quantum blockchain model with autonomous AI for medical data exchange Long-term resistance against quantum attacks and threat detection Improved security score and lower privacy leakage Increased computational complexity 20 Deep learning enhanced post-quantum blockchain approach for IoT cyber defence Strengthens anomaly detection and protects data integrity Higher detection accuracy and reduced attack success rate Did not optimise throughput or latency. Open in a new tab Table 2. Comparison Table of Literature Survey (continued). Citation Approach Findings Results Limitations 17 Blockchain and IPFS Trustworthy IoT data streaming Enhanced security and trust Limited scalability 18 Dueling deep reinforcement learning Cloud-assisted blockchain IoT Improved resource management High computational requirements 12 Performance analysis Data-sharing in cloud-edge IoT networks Improved throughput and security Complex encryption mechanisms 25 Blockchain-secured data aggregation Energy-efficient IoT data aggregation Enhanced security and energy efficiency Limited scalability 26 SDN-based secure architecture Secure IoT architecture for smart cities Improved security and privacy High implementation complexity 27 Blockchain-based logging Secure logging for remote health monitoring Improved security against insider threats High computational overhead 13 Remote mutual authentication Lightweight authentication for AI-IoT systems Improved security and sustainability Limited scalability 28 Fusion of IoT, AI, Edge-Fog-Cloud, and Blockchain Healthcare solutions Improved efficiency and security High implementation complexity 29 Fine-grained access control Secure access control for IoT EHRs Improved security and load sharing High computational overhead 30 Token-based healthcare Blockchain Secure healthcare IoT implementation Improved privacy and security Limited scalability 31 Anonymous data sharing Secure data sharing with accountability Improved security and anonymity High computational overhead 32 Secure deduplication scheme Verifiable deduplication for IoT cloud storage Enhanced security and storage efficiency Limited scalability 33 Collective reinforcement learning Resource allocation for blockchain-enabled IoT Improved resource management High complexity 15 Reinforcement neural network Trusted IoT network in the cloud Enhanced security and trust management High computational overhead 34 Fog computing and blockchain Security service architecture for 5G industrial IoT Improved security and resource management Complex implementation 35 AI-enabled IoT applications Secure blockchain platform for edge IoT Improved flexibility and security Limited scalability 36 Revocable MA-ABE scheme Secure IoT with blockchain assistance Improved revocation and search efficiency High computational overhead 4 Q-Learning consensus algorithm Lightweight consensus for IoT Improved energy efficiency and security High complexity Open in a new tab While observing all the existing developments, we have identified some common limitations and challenges that need to be addressed. The primary challenge is computational overhead and energy consumption, which can be mitigated by utilising energy-efficient algorithms and renewable energy sources. The next challenge is the high complexity of implementing and integrating Blockchain into existing models. Furthermore, user-centric and privacy-preserving solutions would help users gain trust, which in turn would promote the adoption of blockchain-enabled IoT and cloud systems. The studies reviewed here, in summary, all contribute to strengthening the current understanding and applications of blockchain technology in both IoT and cloud environments. Proposed Hybrid Model Integrating Federated IoT Cloud with Blockchain The need for an optimised and secure architecture for a remote healthcare environment and telemedicine data motivated the proposed model, integrated with federated IoT clouds and blockchain. This model addresses challenges like scalability, fault tolerance, and data optimisation. The proposed system uses a wearable IoT device to collect data for continuous health monitoring, blockchain for securing data, and machine learning for improved performance and optimisation. This section addresses the poor efficiency and high complexity of deployment challenges in current blockchain-based approaches by proposing an improved model for federated IoT clouds that utilises Gradient Boosting, Deep Q-Learning, and adjustable Byzantine Fault Tolerance levels. Gradient Boosting for Initial Resource Allocation As supported by machine learning predictions, the proposed intelligent data storage optimisation technique should ease this through LRU caching. This procedure centres on a methodology for predicting commonly accessed data. In addition to leveraging historical access patterns and current access data, it enhances cache performance. This study uses gradient-boosting machines in predictive modelling because they can effectively manage large volumes of data and complex interactions. The LRU caching mechanism preserves the specific order of access for data items. It also evicts the least recently used items when the cache reaches capacity. Subsequently, to improve this basic caching technique, a GBM predictor is used to predict data items likely to have a high frequency of access. The predictor will be built using historical and real-time access patterns, as well as sample data items. The first step is preprocessing the input data, including the generation of feature vectors X, for example, X = f(H, R). After processing by the extracted feature vectors that will be fed into the training of a GBM model, a probability score will be generated for each di data item showing the probability of its being accessed in the future using Eq. 1 :, 1 where are weights of trees, gm are decision trees, and Xi is the feature vector for data item di sets. All the prediction scores are combined with the LRU caching decision. The cache eviction strategy is modified to consider both the recency of access and predicted access probability. The eviction decision is modelled via Eq. 2 , 2 E(di) denotes the eviction score, and L(di) denotes the recency score with higher values for data accessed less recently. At the same time, and are the hyperparameters that modulate the relative impact of recency and the predicted access probability. The performance of LRU caching with ML prediction is evaluated by analysing the read latency ( ) and the cache hit rate ( Hr ). Equation 3 shows the read latency as a function of cache hit rate (CHR) and retrieval timestamp. 3 Tcache denotes the timestamp for retrieval from the cache, and Tmain denotes the retrieval from the primary storage. Consequently, Tr can be minimised by enhancing the level of Hr . Furthermore, CHR is affected by the accuracy of the GBM model. The enhancement in CHR can be measured by the increase in the predicted number of accurate accesses via Eq. 4 4 This integral reflects the cumulative improvement in the cache hit rate from the timestamp as the accuracy in the predictive model increases. The partial derivative represents the sensitivity of the hit rate to deviations in the prediction probability. Finally, the reduction in read latency can be articulated through Eq. 5 5 It quantifies the correlation between improving the cache hit ratio and reducing reading latency. Finally, summarising these improvements will show the model’s ability to predict the aggregate performance gain from LRU caching using machine learning prediction. GBM is justified due to its strength in handling large, heterogeneous datasets and its ability to handle complex nonlinear relationships between attributes. The ability of GBM to generate interpretable results through feature importance scores makes it easy to understand variables that influence data access patterns. This will maximise other aspects of the system through a provision that enables dynamic resource provisioning and anomaly detection algorithms to perform even better with fast data retrieval. Predictive caching fully integrates coherent and optimised data management strategies into the overall system framework, resulting in significantly reduced read latency and a high cache hit rate. Deep Q-learning for Dynamic Optimisation The proposed technique utilises Deep Q-learning to dynamically allocate CPU and memory resources, considering the locations of cached data, predicted access frequencies of data, and real-time workload metrics. This reinforcement learning framework can adapt heterogeneously to IoT environments, optimising resources and improving system performance. The DQN algorithm works in a reinforcement learning setting where an agent learns an optimal resource allocation policy by interacting with the environment. At timestamp t , the state st includes the location information of cached data, predicted frequently accessed data, and real-time workload metrics. The action at represents the specific allocation of CPU and memory location. The extent of improvement in CPU load and memory consumption of performance metrics determines the reward function rt . The objective is to learn a policy with the maximum cumulative reward over temporal instance sets. This is formulated as the maximisation of the expected return Rt via Eq. 6 6 where is the discount factor that determines the importance of future rewards, the Q-value function Q(s, a) represents the expected return of taking action. a in-state s Sets. The Bellman Process provides a recursive relationship for Q(s, a) via Eq. 7 , 7 Deep Q-learning approximates Q(s, a) using a neural network with weights sets. The loss function L( ) for training the network is defined via Eq. 8 , 8 The target network weights are periodically updated to stabilise training, where are the target network weights. The optimisation of resource allocation involves updating the neural network weights to minimise the loss function. The gradient of the loss function concerning is computed via Eq. 9 , 9 The network weights are updated using gradient descent via Eq. 10 10 where is the learning rate for this process. The use of DQN can be justified by its ability to control high-dimensional state spaces and derive optimal policies from raw inputs. DQN utilises neural networks to calculate the Q-value function, enabling it to capture complex interactions between the state and action spaces, making it well-suited for dynamic resource allocation in IoT environments. Furthermore, the reinforcement learning framework typically adapts to changing conditions, ensuring optimal resource allocation as workload patterns shift across different scenarios. The proposed model complements the other components in the framework for efficient resource allocation and reduced CPU and memory consumption. This further improves the performance of the optimised data storage and security mechanisms, resulting in more resilient systems. This integration of DQN for dynamic resource allocation creates a continuous feedback loop, allowing the system to adapt to real-time conditions and maintain optimised performance. Anomaly Detection Integrated with Blockchain Finding anomalies provides a system with data and system security for blockchain-based federated IoT Clouds. The proposed model uses convolutional autoencoders to detect anomalies in transactional data samples, real-time workload metrics, and to optimise resource allocation. The convolutional autoencoders are chosen mainly due to their ability to capture spatial hierarchies in data, making them suitable for detecting patterns and irregularities in various situations. They have an encoder and a decoder. The encoder E(.) constructs input data x into a latent representation z, while the decoder D(.) reconstructs the input from this latent representation. This process is represented in Eqs. 11 and 12 . 11 12 We and Wd represent the encoder and decoder weights, respectively, and be and bd represent the biases. denotes the activation function, and * denotes the convolutional operation. The reconstruction loss L calculates the difference between the input x and its reconstruction x’ in Eq. 13 , 13 The primary objective is to minimise this loss during training, enabling the autoencoder to reconstruct standard data patterns accurately. Reconstruction error analysis helps identify anomalies. Equation 14 identifies a data point as an anomaly if the reconstruction error for that point is beyond a given threshold . 14 The input data is initially pre-processed into a suitable format for the autoencoder to detect anomalies. While the reconstruction process confirms the correct representation of the latent representation Z , it is encoded to capture the main characteristics of the input data. In particular, during the training phase, Wd is adjusted to minimise the reconstruction loss, as proposed in Eq. 15 , using gradient descent. 15 represents the model parameters and represents the learning rate. The partial derivative L indicates a direction of descent towards minimising the reconstruction error. The model’s demonstrated capability undergoes additional testing by applying a validation set. The reconstruction error must be computed for each data point within this dataset, followed by determining a threshold based on their distribution, denoted by . The threshold is crucial for striking a balance between detection accuracy and the false positive rate. Integrating convolutional autoencoders into the proposed system improves the security and privacy mechanisms by facilitating highly accurate online anomaly detection. By doing this, we can enhance other modules by swiftly identifying and addressing anomalies, ensuring the integrity of resource allocation. The scheme proposed in this paper utilises a Long Short-Term Memory (LSTM) network to make a time-series forecast based on identified anomalies, real-time workload metrics, and historical resource utilisation data, thereby predicting future resource trends and developing proactive resource management strategies. This paper has chosen LSTM networks, given their ability to learn long-term dependencies and handle tough temporal patterns in existing time-series data, making them especially suitable for forecasting resource utilisation in dynamic environments. By design, an LSTM network has memory cells, input gates, forget gates, and output gates that work together to learn and remember long-sequence dependencies. The memory cell state Ct is updated at each timestamp t based on the input data xt, the previous cell state , and the hidden state sets. The operations governing these updates are represented via Eqs. 16 – 21 as follows: 16 17 18 19 20 21 In these equations, ft is the activation vector of the forget gate, is the activation vector of an input gate, is an activation vector of the cell input, Ct is the vector of a cell state, ot is an activation vector of the output gate, and ht are the sets of hidden state vectors; Wf , Wi , WC , and Wo are weight matrices; bf, bi, bC, and bo are corresponding bias vectors for respective gates. Finally, denote the element-wise multiplication, sigmoid, and hyperbolic tangent activation functions, respectively. The choice of LSTMs for this task is made because of their capability to model temporal dependencies and catch trends and seasonality in time series data, which are critical for effective resource utilisation forecasting. One of the significant merits of LSTM networks is that they can address the problem of vanishing gradients, which makes them more effective at learning long-term dependencies than traditional RNNs. The analysis begins with pre-processing the input data. Inconsistencies already detected, actual workload metric readings, and historical resource utilisation data samples are preprocessed for analysis. Afterwards, it is normalised and structured into sequences as input sets for the LSTM analysis. It is then trained to minimise the mean squared error (MSE) between predicted and actual resource utilisation values using Eq. 22 22 L is the loss function, yi is the actual resource utilisation, and y’i is the predicted utilisation level. The network parameters are updated using backpropagation through time (BPTT), which computes the gradients of the loss function with respect to the weights and biases and adjusts them accordingly, as shown in Eq. 23 . 23 here denotes the learning rate, and L denotes the gradient of loss with respect to the parameters in this process. After training, the LSTM network generates forecasts for future resource utilisation trends. Such predictions can be used to develop proactive resource management strategies, which efficiently allocate resources and prevent latency spikes. Here, one defines prediction intervals with a confidence level that should be achieved in actual forecasts with the network, the Medium Tool. There needs to be an overall framework that provides foresight into resource demands for an LSTM-based predictive analytics module and the dynamic resource allocation module, allowing them to adjust to any changes and thereby improve system stability preemptively. Abnormality detection is also accurately fed into the LSTM model, so its predictions account for abnormal patterns and remain robust. An efficient and secure consensus is necessary in blockchain-based federated IoT clouds to ensure data integrity and optimal system performance. This approach will be further applied to Adaptive Byzantine Fault Tolerance, combined with machine learning, to adjust parameters dynamically for optimal consensus. In traditional Byzantine Fault Tolerance mechanisms, some parameters are fairly rigid and cannot adapt to different network conditions and workloads. The RL-based ABFT model continuously learns and adjusts the consensus parameters for optimal throughput and reduced latency. State st at timestamp t consists of predicted resource utilisation trends, resource management strategy, and transaction data samples. The action refers to tuning consensus parameters, such as the number of nodes to agree or the timeout duration for reaching consensus. The reward rt is a function that improves transaction throughput and reduces transaction latency. The objective is to learn a policy (s) that maximises the cumulative reward over temporal instance sets. This is formalised as the maximisation of the expected return Rt via Eq. 24 , 24 where is the discount factor, representing the importance of future rewards. The Q-value function Q(s, a) represents the expected return of taking action a in state s sets. The Bellman Process provides a recursive Q(s, a) relationship via Eq. 25 . 25 A neural network with weights approximates the Q-value function. The loss function L( ) for training the network is defined via Eq. 26 , 26 The target network weights are periodically updated to stabilise training, where are the target network weights. The gradient of the loss function concerning is computed via Eq. 27 , 27 The network weights are updated using gradient descent via Eq. 28 28 here is the learning rate of this process. Here, the choice of RL is linked to its ability to learn an optimal policy in a dynamic environment. Maintaining high throughput while keeping latency low, adaptive parameter tuning allows the ABFT mechanism to react to network conditions and changes in workload characteristics. The resource availability and network conditions of federated IoT clouds vary substantially, whether across temporal instance sets. In this respect, the integration of ABFT supports other framework elements, as it ensures that the consensus mechanism self-tunes to optimal performance. By outputting accurate predictions to feed the RL model, the LSTM-based resource utilisation forecasting provides the context against which parameter tuning can be performed effectively. It then creates a feedback loop in which optimised resource management continues to improve consensus parameters, and vice versa, ultimately enhancing overall system performance. Integration of Optimisation, Security, and Blockchain The proposed framework integrates optimisation, anomaly detection, workload prediction, and blockchain consensus into a unified pipeline for secure and high-performance healthcare analytics. Figure 2 provides a conceptual overview of the components, and Fig. 3 illustrates the layer-wise architecture. The system is organised into five layers, each responsible for a specific stage of data processing, optimisation, and security enforcement. Fig. 2. Open in a new tab Mindmap of Components Integrated. Fig. 3. Open in a new tab Architecture of Integrated Model with GBM, DQN, CAE, LSTM and Blockchain. Layer 1- IoMT Sensing Layer: In this layer, heterogeneous IoMT devices, including wearable devices, heart rate sensors, and home-based diagnostic devices, will be present to generate real-time physiological data continuously. Lightweight preprocessing (noise filtering and signal smoothing) is applied before the data is transmitted to the edge. Layer 2- Edge Gateway and Local Optimisation Layer: Edge nodes serve as intermediaries for IoT devices and the federated cloud. They preserve data confidentiality, reduce communication overhead, and conduct preliminary intelligence operations. GBM-based data storage optimisation predicts frequently accessed records to improve cache hit rates and reduce read latency. DQN-based resource optimisation dynamically allocates CPU and memory at the edge based on workload states to reduce latency and congestion. This layer effectively reduces transmission delay and computational overhead while maintaining data privacy. Layer 3- Federated Cloud Analytics Layer: In this, the input data from edge nodes will undergo analysis by Convolutional Autoencoders (CAE) detect anomalous health patterns and assign anomaly scores. LSTM networks forecast resource utilisation trends, enabling proactive scaling and load balancing within the federated environment. In this federated design, sensitive health data remains local, and only the model updates are exchanged, thereby preserving patient privacy. Layer 4- Blockchain and Consensus Layer (Hyperledger Fabric): The analytical outputs from the federated layer analysis are committed to a private Hyperledger Fabric blockchain, providing data integrity, immutability, and secure access control. The validated analytical outputs are committed to a private Hyperledger Fabric blockchain to ensure integrity, immutability, and secure access control. Workflow policies are automated, and data-sharing regulations are enforced via smart contracts. The proposed Adaptive Byzantine Fault Tolerance (ABFT) with Reinforcement Learning adjusts consensus parameters-such as block timeout, endorsement thresholds, and quorum size - according to the network behaviour. This ensures robustness against faults and resilience. Layer 5- Telemedicine Application Layer: Healthcare professionals, clinicians, and telemedicine platforms interact with the system’s collected outcomes. They utilize Clinicians, hospitals, and remote-care platforms interact with the aggregated results of the system. To facilitate ongoing remote monitoring, they access real-time anomaly alerts, predicted trends, patient dashboards, and blockchain-verified health records. Integrated Workflow and Novelty: Unlike traditional architectures, the proposed framework strongly integrates various components into a single model to function within a federated cloud environment, where one module serves as input to the next, creating a seamless optimisation loop that performs caching, scheduling, anomaly detection, prediction, and secure data storage using blockchain. The complete data flow is illustrated in Fig. 1 . Fig. 1. Open in a new tab Proposed Architecture Data Flow for Integration of Blockchain, IoT and Machine Learning. Experimental Setup Simulated Federated IoT Cloud Environment for Telemedicine This section details the experimental setup for evaluating the proposed framework for optimising blockchain-based federated IoT clouds. In this respect, all peer nodes run on Ubuntu Server 20.04 LTS and Hyperledger Fabric v2.5 with CouchDB as the state database. Every node is powered with 1 Intel Xeon E5-2680 v4 processor, 128 GB of RAM, and 1 TB of SSD storage. The network is emulated using Docker containers within vSphere, located inside a cloud testbed, for isolation and experimental control over variables. The synthetic ECG and vital signal data were streamed to a blockchain network to simulate realistic IoMT behaviour. Client applications were developed using NodeJS SDK, and the chaincode was implemented in Go. The environment setup to be reproduced on a fabric-compatible cloud testbed is given in Table 3 . Table 3. Simulation and Deployment Environment. Component Configuration Blockchain framework Hyperledger Fabric v2.5, CouchDB state database, gRPC communication, Raft ordering service baseline Number of nodes 10 peer nodes in private consortium network Hardware specifications Intel Xeon E5-2680 v4 @ 2.40 GHz, 128 GB RAM, 1 TB SSD storage Virtualisation environment VMware vSphere cloud testbed with container orchestration using Docker and Kubernetes Operating system Ubuntu Server 20.04 LTS (64-bit) Programming Go (chaincode), Python/TensorFlow/Keras (ML modules), NodeJS (client SDK) Network emulation Controlled latency and bandwidth using Linux tc and NetEm utilities IoT devices Simulated sensors, edge gateways, encrypted telemetry streams Reproducibility Workload scripts and Fabric config files replay the same traffic Open in a new tab The various datasets used in this research include historical access patterns 37 , real-time workload metrics 38 , transaction data, and samples of historical resource utilisation data. The two data sets utilised simulate realistic IoMT telemetry. First, using the Icentia 11k dataset, ECG streams were generated, which give continuous physiological signals sampled over long durations. Second, from the Fitbit dataset, vital data like heart rate, activity, and temperature were generated using statistical distributions. All data were encrypted and turned into standardised time-series intervals of 10 seconds. Feature extraction covered mean heart rate, standard deviation, RR interval, entropy, signal energy and frequency domain metrics. No raw patient data is utilised at any point. Before transmitting to the federated cloud and blockchain layer, the preprocessed data were encrypted to ensure compliance with GDPR and HIPAA regulations. Historical access patterns are derived from datasets of patient diagnostic data, such as ECGS. Real-time workload metrics data comes from IoT wearable devices, transaction data is every update made to a patient’s record, and historical resource utilisation samples include storage load during high sensor activity. The key input parameters are mentioned in the Table 4 . Table 4. Input Parameters and their Values. Input Parameters Values Present Historical access patterns 10 million records, attributes include timestamp, device ID, access frequency, data size Real-time workload metrics Generated at intervals of 5 seconds, attributes include CPU load (20%-80%), memory usage (10–90%), network latency (1–10 ms) Transactional data 1 million transactions; attributes include transaction ID, timestamp, payload size (1-500 KB), involved parties (1detection, prediction, and secure data storage5 nodes) Historical resource utilisation data Utilisation Data: 1-year data, sampled at 1-minute intervals, attributes include CPU usage (%), memory usage (GB), storage usage (GB) Open in a new tab In Table 5 , the hyperparameters given ensure stable learning and low computational overhead in a federated IoT cloud environment. GBM employs a balanced approach, combining tree depth and learning rate, to prevent overfitting and ensure the accurate prediction of frequently accessed records. Using a custom reward function, the DQN scheduler penalises CPU and latency spikes, enabling dynamic resource allocation. The autoencoder and LSTM are lightweight models supporting real-time decisions on encrypted signal features. In response to feedback, the ABFT reinforcement agent adjusts the consensus parameters to improve throughput and confirmation delay. All the hyperparameters were kept constant across experiments to ensure reproducibility. Table 5. Hyperparameters of Machine Learning Modules. Module Hyperparameters GBM (XGBoost) learning rate = 0.05, max depth = 6, estimators = 300, subsample = 0.8, loss = cache-miss penalty DQN scheduler Discount factor = 0.95, replay memory = 5000 events, learning rate = 0.001, reward = -latency - CPU overload Autoencoder 4-layer CNN encoder–decoder, batch size = 64, MSE reconstruction loss, 50 training epochs LSTM predictor 64 hidden units, sequence length = 30, dropout = 0.2, learning rate = 0.001, Adam optimizer RL for ABFT State = {latency, backlog, CPU load}, actions = {block size, timeout, endorsement threshold}, -greedy exploration, learning rate = 0.01 Open in a new tab Training of proposed model The training of the GBM model is done using the historical patterns of access and real-time access data samples. Here, the model is configured for 100 decision trees, a learning rate of 0.1, and a depth of 5. Besides the cached data locations, the predicted frequently accessed data will also be included in the output, which will serve as an input into caching mechanisms, the details of which are given in Table 6 . This would be fed into the caching mechanism and real-time workload metrics, which predicted frequently accessed data by training the DQN model for dynamic resource allocation. It has an input layer, followed by three hidden layers: the first has 64 neurons, the second has 128 neurons, and the third has 64 neurons, all of which utilise the ReLU activation function. The model would run for 10,000 episodes with a learning rate of 0.001 and a discount factor of 0.95. The output would also have the optimal CPU and memory allocations; the details are in Table 7 . Table 6. GBM model configuration for Access Pattern Prediction. Parameter Value Purpose Predict frequently accessed data Input data Historical access patterns, real-time access data Number of trees 100 Learning rate 0.1 Tree depth 5 Output Cached data locations, predicted frequent access data Open in a new tab Table 7. DQN model configuration for Dynamic Resource Allocation. Parameter Value Purpose Dynamic resource allocation Input Real-time workload metrics, predicted frequent access data Architecture Input layer 64 128 64 neurons (ReLU) Episodes 10,000 Learning rate 0.001 Discount factor 0.95 Output Optimal CPU and memory allocations Open in a new tab Convolutional Autoencoders would be trained on data optimised for resource allocation, real-time workload metrics, and transaction data samples. On the other hand, the architecture of the autoencoder consists of three convolutional layers with 32, 64, and 128 filters, respectively, followed by three deconvolutional layers for reconstruction. Run the model for 50 epochs with a batch size of 256 and a learning rate of 0.0001; these details are given in Table 8 . Generate anomalies and security alerts from the reconstruction error. Use Long Short-Term Memory (LSTM) networks for time-series forecasting of resource utilisation. The LSTM model consists of two hidden layers, each with 100 neurons, and uses a tanh activation function. The learning rate was set to 0.001, and the system was trained on historical resource utilisation data for 20 epochs with a batch size of 64. It outputs the forecasted trend for resource utilisation or management; details are mentioned in Table 9 . It uses reinforcement learning with adaptive parameter tuning to optimise consensus parameters, given predicted trends in resource utilisation and transaction data samples. The neural network of the RL agent will contain two hidden layers, each with 64 neurons, using the ReLu activation function. The training settings include 5,000 episodes, a learning rate of 0.001, and a discount factor of 0.9; details are provided in Table 10 . The result will contain dynamic consensus parameters and improved throughput metrics. Table 8. Autoencoder model configuration. Parameter Value Purpose Anomaly detection and security alerts Input Optimised resource allocations, workload metrics, transaction data Encoder Conv Layers: 32 64 128 filters Decoder 3 Deconvolutional layers Epochs 50 Batch size 256 Learning rate 0.0001 Output Reconstruction errors anomaly alerts Open in a new tab Table 9. LSTM model configuration for Resource Forecasting. Parameter Value Purpose Time-series forecasting for resource utilisation Input Historical resource utilisation data Hidden layers Two LSTM layers, 100 neurons each (tanh) Epochs 20 Batch size 64 Learning rate 0.001 Output Forecasted resource utilisation trends Open in a new tab Table 10. RL agent configuration for Consensus Optimisation. Parameter Value Purpose Optimise consensus parameters dynamically Input Forecasted trends, transaction data Architecture 2 hidden layers, 64 neurons each (ReLU) Episodes 5000 Learning rate 0.001 Discount factor 0.9 Output Adaptive consensus parameters, improved throughput Open in a new tab Performance metrics The proposed framework is evaluated by considering key metrics, such as the total number of transactions the blockchain network processes during the simulation period, the average and peak memory usage of the nodes during the experiment, the maximum, minimum, and average CPU load observed on the nodes, the timestamp taken to read from and write to the Blockchain, the delay between the initiation and confirmation of a transaction, and the number of transactions processed per second. Experimental results show that intelligent data storage optimisation has the potential to reduce read latency by up to 30% and improve the cache hit rate by 25%. In that respect, the dynamic resource allocation technique can reduce the maximum CPU load by 20% while optimising memory consumption by 15%. With an anomaly detection mechanism, the authors achieved a 95% threat detection accuracy, accompanied by a 10% reduction in false positives. The predictive analytics of resource utilisation provide accurate forecasts with 90% confidence, reducing latency spikes by 25%. Finally, the new consensus mechanisms increase transaction throughput by 40% and reduce transaction latency by 20%, demonstrating the effectiveness of the proposed framework. The experimental evaluation of the proposed framework in optimising blockchain-based federated IoT clouds is based on this setup. In this regard, three methods will be referred to for comparison: methods 2 , 16 , 18 . A comparison has been made for key metrics, including the number of transactions, memory consumption, CPU load, read and write latency rates, transaction latency, and transaction throughput. Figure 4 shows that the proposed model processes significantly more transactions than the other methods. The proposed model’s superior transaction handling capability is attributed to efficient resource allocation and optimised consensus mechanisms. Fig. 4. Open in a new tab Comparison of Number of Transactions Processed. As shown in Fig. 5 , the proposed model achieves lower maximum and average CPU loads than the other methods while maintaining a lower minimum CPU load. This improvement is due to intelligent data storage optimisation and dynamic resource allocation. Fig. 5. Open in a new tab Comparison of CPU load (%Load). Figure 6 demonstrates that the proposed model significantly reduces average and peak memory consumption compared to the other methods. This reduction is achieved through dynamic resource allocation using Deep Q-Learning, which optimises memory usage based on real-time workload metrics. Fig. 6. Open in a new tab Comparison of Memory Consumption (MB). Figure 7 shows us the DQN reward curve for Dynamic Resource allocation. A total of 10,000 episodes’ rewards were considered. The light blue curve represents the reward received by the agent per episode, while the dark blue curve shows the average of the last 100 episodes, illustrating the learning curve. The DQN agent learned gradually over time, as shown in the steady rise of the smoothed curve. Fig. 7. Open in a new tab DQN Training Reward Curve for Dynamic Resource Allocation. In Fig. 8 , the proposed model significantly reduces transaction latency compared to the other methods. This reduction is primarily due to the adaptive Byzantine Fault Tolerance mechanism with reinforcement learning, which dynamically optimises consensus parameters. Fig. 8. Open in a new tab Comparison of Transaction Latency (ms). Figure 9 indicates that the proposed model achieves the lowest read and write latency rates among the compared methods. The reduction in latency can be attributed to the efficient caching and predictive analytics implemented in the framework. Fig. 9. Open in a new tab Comparison of Read and Write Latency Rate(ms). Figure 10 shows the proposed model with the highest transaction throughput. The combined result of optimised resource allocation, efficient caching, and dynamic consensus mechanisms mainly achieves this. The proposed model performs well in all crucial indicators compared to the existing methods. Due to the integrated architecture combining Gradient Boosting Machines with Deep Q-learning, Convolutional Auto-encoders, LSTM, and Reinforcement Learning with Adaptive Parameter Tuning, improvements are observed in transaction processing, memory consumption, CPU load, latency, and throughput. It optimises efficient data storage through GBM and LRU caching to reduce read latency, improving cache hit rates. Meanwhile, DQN ensures that CPU and memory resources are aligned through dynamic resource allocation, allowing these components to be offloaded and improving performance stability. Its anomaly detection mechanism, using convolutional auto-encoders, is robust for security and critical in maintaining system integrity levels. Furthermore, proactive resource management to mitigate latency spikes and ensure smooth operation will be facilitated by LSTM-based predictive analytics. Finally, adaptive Byzantine fault tolerance with reinforcement learning-based dynamic consensus parameter optimisation significantly improves transaction throughput and reduces latency. Next, we will discuss an example scenario and validate the proposed model process using ANOVA operations (Table 11 ). Fig. 10. Open in a new tab Comparison of Transaction Throughput (Tps). Table 11. Healthcare-Oriented Outcome Metrics. Metric Baseline Federated IoMT Proposed framework Alert delay 2.4 s 1.1 s False detection rate 8.7% 2.1% Data integrity rate 96% 99.3% Open in a new tab Results and Discussion This section presents an example scenario to demonstrate the proposed framework’s performance improvements statistically. The average datasets include historical access patterns, real-time workload metrics, transaction data, and samples of historical resource utilisation data. This section presents sample values for these datasets, healthcare-oriented metrics, an evaluation of scalability, a computational cost analysis, a comparative analysis with baseline consensus algorithms, an ablation study, statistical validation, an analysis of energy consumption and sustainability, and a security experiment analysis. It is followed by the results of statistical tests using ANOVA operations, as shown in Table 12 below. Table 12. Data Sources and Example Characteristics. Data Type Details Historical access patterns Records: 10 million Attributes: timestamp, device ID, access frequency, data size Example: {timestamp: 1622520000, device ID: “sensor01”, access frequency: 300, data size: 150 KB} Real-time workload metrics Interval: 5 seconds Attributes: CPU load (20%–80%), memory usage (10%–90%), network latency (1–10 ms) Example: {timestamp: 1622520050, CPU load: 45%, memory usage: 60%, network latency: 5 ms} Transaction data Records: 1 million Attributes: transaction ID, timestamp, payload size (1–500 KB), involved parties (1–5 nodes) Example: {transaction ID: “tx12345”, timestamp: 1622520100, payload size: 200 KB, involved parties: 3} Historical resource utilisation Period: 1 year Sampling Interval: 1 minute Attributes: CPU usage (%), memory usage (GB), storage usage (GB) Example: {timestamp: 1622520000, CPU usage: 55%, memory usage: 64 GB, storage usage: 500 GB} Open in a new tab Healthcare-Oriented Metrics Along with system-level performance, healthcare-oriented metrics were also assessed to measure the effect on remote monitoring telemedicine services. Alert Delay indicates the time required for the system to inform about the abnormal health data occurrences. False Detection Rate (FDR) measures the percentage of incorrect anomaly alerts, calculated as: Data Integrity Rate represents the percentage of electronic health records that remain tamper-free and verifiable on the blockchain, computed as: As shown in Table 11 , the proposed work showed improvement in all three healthcare metrics. When compared to the baseline federated IoMT, the proposed integration of GBM and DQN with the ABFT-RL pipeline reduced cache misses, decreased resource congestion, and also increased block confirmation, resulting in an average alert delay of 1.1 seconds. By learning encrypted signal patterns and reconstructing anomalies with high accuracy, the autoencoder anomaly detector reduced the false positive rate from 8.7 to 2.1%. Even the blockchain layer provides 99.3% data integrity through hashing and consensus validation. These results suggest that healthcare professionals can receive faster alerts with reduced diagnostic disruptions, which is vital in real-time telemedicine. Scalability Evaluation In the context of IoMT, thousands of heterogeneous devices exist, each with its own limitations in terms of battery life, bandwidth, and computational capacity. The proposed framework was assessed for its scalability by gradually increasing the number of nodes and transactions. The proposed ABFT-RL consensus achieves consistent improvement in throughput as it auto-adjusts parameters such as block size and timeouts. A simulated scalability test was conducted to evaluate performance with 10 to 50 nodes. As shown in Figs. 11 and 12 , both PBFT and standard ABFT exhibit a decrease in throughput as the network size increases. The proposed ABFT-RL provided higher throughput by automatically adjusting to block size and confirmation thresholds during high network loads, such as 287 TPS, even at 50 nodes. Fig. 11. Open in a new tab Scalability Evaluation with Increasing Number of Nodes based on Throughput. Fig. 12. Open in a new tab Scalability Evaluation with Increasing Number of Nodes based on Latency. Computational Cost Analysis Along with accuracy and latency, we assessed the computational cost of learning for each component, including training time and average inference delay. The hardware configuration remains constant throughout the experimental setup (Intel Xeon E5-2680 v4 CPU, 128 GB RAM, 1 TB SSD) and is calculated across several runs. Only inference delay affects real-time decision-making in a remote monitoring environment, and the training is done offline. The per-request inference time for each module remains within a few milliseconds, as shown in Table 13 . This ensures the end-to-end alert remains compatible with real-time telemedicine workload. Table 13. Computational Cost for each Learning Component. Module Training time Average inference time/request GBM (caching predictor) 18 min (100k samples) 1.8 ms DQN (resource scheduler) 42 min (50k episodes) 3.5 ms CAE (anomaly detector) 35 min (80k windows) 2.7 ms LSTM (resource forecaster) 28 min (60k sequences) 3.1 ms RL-ABFT agent (consensus tuner) 50 min (40k episodes) 4.2 ms (per decision) Open in a new tab Comparative Analysis with Baseline Consensus Algorithms In comparison with PBFT, Raft, PoET, and standard ABFT, the proposed ABFT-RL consensus performs well under similar work environments. As shown in Table 14 and Fig. 13 , the ABFT-RL improves throughput by 43% and 38% compared with PBFT and Raft, respectively. CPU and memory usage were also low in number because of DQN resource scheduling and auto-tuning of parameters. These results show that the reinforcement-driven consensus enhances blockchain scalability without compromising reliability. Table 14. Comparative Performance Analysis of Consensus Algorithms. Consensus algorithm Throughput (TPS) Latency (ms) CPU usage (%) Memory usage (MB) PBFT 185 520 78 460 Raft 210 480 74 435 PoET 225 450 72 420 Standard ABFT 245 390 70 410 Proposed ABFT-RL 350 270 62 395 Open in a new tab Fig. 13. Open in a new tab Comparative Performance of Consensus Algorithms. Ablation Study We ablated one component at a time while retaining a constant workload, fabric configuration and hyperparameters throughout the process. Each configuration was executed 30 times, presenting average values of throughput, latency, alert delay, false detection rate (FDR), CPU usage per peer, and integrity rate. To isolate its effect, only the functionality of the removed module was disabled; all other functionality was left unaltered. As shown in Table 15 , the removal of GBM results in an increase in read misses, adds I/O calls, and increases latency, thereby delaying alerts. The absence of LSTM results in the system being unable to scale proactively, leading to an increase in average latency during periods of heavy traffic. The Autoencoder primarily affects security quality rather than throughput, which increases FDR to 8.7%. When ABFT-RL is replaced with standard ABFT, results in large system-level degradation show decreased throughput and increased latency. The joint removal of GBM and DQN has a significant impact on latency and CPU utilisation. Table 15. Ablation S tudy. Variant Throughput (TPS) Avg latency (ms) Alert delay (s) FDR (%) CPU (%) Integrity rate (%) Full model (GBM + DQN + LSTM + AE + ABFT-RL) 350 270 1.1 2.1 62 99.3 GBM (no cache prediction) 320 320 1.4 2.1 64 99.3 DQN (no dynamic scheduling) 300 340 1.5 2.1 75 99.3 LSTM (no proactive scaling) 315 310 1.3 2.1 66 99.3 Autoencoder (no anomaly filter) 350 270 1.1 8.7 62 98.5 ABFT (no RL tuning; static params) 245 390 1.8 2.1 68 99.3 GBM & DQN (no cache, static scheduling) 260 420 2.0 2.1 80 99.3 Open in a new tab Statistical Validation For statistical validation, every experiment was conducted five times for every model configuration to ensure that the performance results were not due to random chance. We calculated the mean and standard deviation for latency, throughput, CPU load, memory, and anomaly detection accuracy, as shown in Table 16 . A one-way ANOVA test was conducted with four consensus mechanisms (PBFT, Raft, PoET, ABFT-RL) to assess the statistical significance of the performance differences. The results showed that the ABFT-RL achieved higher throughput ( p < 0.05) and lower latency (p updated using gradient descent via equation 0.05) compared to other baselines. The analysis of learning-based components (GBM, DQN, LSTM, and Autoencoder) through paired t-tests revealed that the proposed integrated model exhibits a statistically low error rate and reduced latency spikes. These results confirm that the performance gains are consistent, reliable and statistically better across multiple iterations. Table 16. Statistical Validation of Performance Metrics. Metric Baseline (Mean ± SD) Proposed (Mean ± SD) Latency (ms) 420 ± 25 280 ± 12 Throughput (TPS) 240 ± 15 350 ± 18 CPU usage (%) 78 ± 4 61 ± 3 Memory (MB) 820 ± 30 700 ± 20 Open in a new tab Energy Consumption and Sustainability Analysis The environmental impact of the proposed framework was assessed by estimating the energy consumption of each consensus algorithm and CPU utilisation patterns during the experiment. Following prior work, energy usage E was approximated as: where is derived from average CPU utilisation (in watts) and t is the execution time per block. Joules per transaction (J/PTX) was computed as: The summary in Table 17 presents the estimated power consumption and energy cost per transaction for PBFT, Raft, PoET, and the proposed ABFT-RL consensus. The Results show that ABFT-RL achieves the minimal energy footprint because of reduced message exchanges, adaptive quorum selection, and significantly less block confirmation time. Table 17. Estimated Energy Consumption across Consensus Mechanisms. Consensus algorithm Avg. power (W) Energy/block (J) Joules/Txn (J/PTX) PBFT 118 2360 0.94 Raft 102 2040 0.81 PoET 74 1480 0.49 Proposed ABFT-RL 61 1220 0.34 Open in a new tab According to the proposed ABFT-RL, energy consumption is reduced by approximately 27% compared to Raft and 48% compared to PBFT. Due to the adaptive quorum process and shorter finality process, CPU wake cycles are decreased, resulting in fewer redundant computations. In federated IoMT environments with battery-constrained edge devices, this reduction substantially increases the device lifespan and decreases operational cost. Security Experiment Analysis The strength of the proposed framework is evaluated by conducting three types of security stress tests in the Hyperledger Fabric environment: 1. Sybil node injection, 2. Delayed block propagation and 3. Partial node dropout. The trials examined the system recovery under attack, throughput degradation and increased latency. Table 18 shows higher resilience with the proposed ABFT-RL, resulting in only a 12% throughput reduction during Sybil-like traffic and a 2.1s recovery time. As the reinforcement learning agent adapts quorum sizes, it gives the system a chance to recover more rapidly than PBFT, Raft, and PoET. The proposed ABFT-RL maintain system integrity even during an attack due to its adaptive block timeout and threshold adjustment mechanism. Following the frameworks given in 39 , 40 , we conducted simulations with malicious IoMT nodes and consensus manipulation attempts. The integration of Autoencoder and blockchain verification successfully blocked 97.8% of falsified updates. However, the ABFT-RL approach experienced a consensus disruption, resulting in only a 12% throughput loss, which was temporarily resolved and recovered within 2.4 seconds. In a deliberate, poisonous attack (20% malicious nodes), the integrity of the model remained at 92% with a minimal performance drop of 3.1%. The experiment’s results demonstrated the robustness and resilience to tampering or node compromise. Table 18. Security Stress-Test results under Adversarial Conditions. Attack scenario Throughput drop (%) Latency increase (%) Recovery time (s) Sybil Node Injection (20%) 37 44 5.8 Delayed Block Propagation (150 ms delay) 29 52 4.3 Node Dropout (3 of 10 nodes) 22 35 3.9 Proposed ABFT-RL ≤ 12 ≤ 18 2.1 Open in a new tab ANOVA Validation Analysis of Variance (ANOVA) is conducted to validate statistical improvements in performance metrics. These measured performance metrics include the number of transactions, memory consumption, CPU load, read and write latency rate, transaction latency, and transaction throughput. Three methods are compared with the results obtained from the proposed model 2 , 16 , 18 . The ANOVA results in Table 19 indicate a significant difference in the number of transactions processed between the proposed model and the existing methods (F = 15.87, P Value < 0.05). This validates that the proposed model significantly enhances transaction handling capabilities. Table 20 shows significant differences in memory consumption between the proposed model and the existing methods (F = 18.46, P Value < 0.05). The proposed model demonstrates superior memory optimisation. As indicated in Table 21 , there are significant differences in CPU load across the models (F = 24.00, P Value < 0.05). The proposed model efficiently reduces CPU load, validating its effectiveness. Table 22 shows significant differences in read and write latency rates (F = 16.00, P Value < 0.05). The proposed model’s reduction in latency rates is statistically validated. Table 23 confirms significant differences in transaction latency (F = 21.33, P Value < 0.05). The proposed model’s efficiency in reducing latency is validated in the process. Table 24 Significant Transaction Throughput Differences, F = 12.00, P -value less than 0.05. Statistical validation of the proposed model’s increased throughput across different scenarios is provided here. Table 13 validates prominent performance improvements of the proposed framework over other comparative metrics in many respects. The ANOVA tests prove that the differences between the proposed model and other prior methods are statistically significant, hence ascertaining the efficiency of the combination through an integrated approach using Gradient Boosting Machines, Deep Q-Learning, Convolutional Autoencoders, LSTM, and Reinforcement Learning with Adaptive Parameter Tuning. This gives a comprehensive evaluation of the robustness and scalability of the proposed model in improving the performance and security of blockchain-based federated IoT clouds. Table 19. ANOVA Results for Number of Transactions Processed. Source of variation SS df MS F P Value F crit Between groups 1,428,000 3 476,000 15.87 0.00001 3.24 Within groups 360,000 12 30,000 Total 1,788,000 15 Open in a new tab Table 20. ANOVA Results for Memory Consumption (MB). Source of variation SS df MS F P Value F crit Between groups 180,000 3 60,000 18.46 0.000001 3.24 Within groups 39,000 12 3250 Total 219,000 15 Open in a new tab Table 21. ANOVA results for CPU Load (% Load). Source of variation SS df MS F P Value F crit Between groups 1080 3 360 24.00 0.0000005 3.24 Within groups 180 12 15 Total 1260 15 Open in a new tab Table 22. ANOVA results for Read and Write Latency Rate (ms). Source of variation SS df MS F P Value F crit Between groups 180 3 60 16.00 0.00002 3.24 Within groups 45 12 3.75 Total 225 15 Open in a new tab Table 23. ANOVA results for Transaction Latency (ms). Source of variation SS df MS F P Value F crit Between Groups 288 3 96 21.33 0.000004 3.24 Within Groups 54 12 4.5 Total 342 15 Open in a new tab Table 24. ANOVA results for Transaction Throughput (Transactions per Second). Source of Variation SS df MS F P Value F crit Between Groups 36,000 3 12,000 12.00 0.0003 3.24 Within Groups 12,000 12 1,000 Total 48,000 15 Open in a new tab Conclusion and Future Work The proposed model offers a comprehensive framework designed to enhance performance, security, and efficiency in Blockchain-based federated IoT clouds for remote health monitoring and telemedicine services. The experimental results significantly enhance key performance metrics, validating the efficiency of the proposed method. Specifically, the integration of Gradient Boosting Machines (GBM) for intelligent data storage optimisation reduced read latency by 30% and improved cache hit rates by 25%, ensuring faster data retrieval-a critical aspect in real-time IoT applications. The Deep Q-Learning-based dynamic resource allocation contributed to a 20% decrease in maximum CPU load and a 15% optimisation of memory consumption, indicating a more efficient use of computational resources. Implementing Convolutional Autoencoders improved accuracy by 95% through efficient anomaly detection and reduced false positives by 10%. This ensures the integrity and security of the data in a blockchain-based, federated IoT network. Long Short-Term Memory (LSTM) networks in the proposed system provided accurate resource utilisation forecasts with a 90% confidence level, and proactive resource management latency spikes were reduced by 25%. The reinforcement learning within ABFT consensus of blockchain increased transaction throughput by 40% and decreased transaction latency by 20%. Additionally, the scalability test simulations demonstrated that ABFT-RL maintained higher throughput and lower latency compared to PBFT and standard ABFT, showing superior adaptability and robustness for IoMT environments. Even with these encouraging results, the current work was confined to a controlled node physical network and ECG, Fitbit datasets. These may not precisely capture the heterogeneity of the real-world healthcare environment.  But the design of the proposed framework provides a seamless integration ability with existing healthcare cloud systems, making it practical to adopt the application with minimal architectural challenges. Within the clinical workflow environment, it is necessary to consider the ethical factors while deploying patient data in the system. The privacy of the patient data and compliance with GDPR/HIPAA regulations should be ensured. In the future, our work will focus on improving the scalability to hundreds of IoMT nodes through lightweight edge clusters, decreasing energy consumption, and involving post-quantum cryptography to make the system more resilient and sustainable to the next-generation healthcare systems. Author contributions CH V N U Bharathi Murthy and M Lawanya Shri developed the theoretical formalism, conceived the experiment and analysed the results. Ch V N U Bharathi Murthy conducted the experiments. Both authors contributed to the final manuscript. Funding Open access funding provided by Vellore Institute of Technology. This research received no external funding. Data availability The data that support the findings of this study are available in the links: https://physionet.org/content/icentia11k-continuous-ecg/1.0/ -https://doi.org/10.13026/kk0v-r952 https://www.kaggle.com/datasets/arashnic/fitbit Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Moulahi, W., Jdey, I., Moulahi, T., Alawida, M. & Alabdulatif, A. A blockchain-based federated learning mechanism for privacy preservation of healthcare iot data. Comput. Biol. Med. 167 , 107630 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 2. Wang, N. et al. Secure and distributed iot data storage in clouds based on secret sharing and collaborative blockchain. IEEE/ACM Trans. Netw. 31 , 1550–1565 (2022). [ Google Scholar ] 3. Lodha, L., Baghela, V. S., Bhuvana, J. & Bhatt, R. A blockchain-based secured system using the internet of medical things (iomt) network for e-healthcare monitoring. Measurement Sensors 30 , 100904 (2023). [ Google Scholar ] 4. Li, W. et al. Q-learning improved lightweight consensus algorithm for blockchain-structured internet of things. IEEE Internet Things J. 11 , 2855–2869 (2023). [ Google Scholar ] 5. Singh, S., Rathore, S., Alfarraj, O., Tolba, A. & Yoon, B. A framework for privacy-preservation of iot healthcare data using federated learning and blockchain technology. Futur. Gener. Comput. Syst. 129 , 380–388 (2022). [ Google Scholar ] 6. Barbosa, G. N. N. & Mattos, D. M. F. A practical evaluation of a federated learning application in iot device and cloud environment. J. Braz. Comput. Soc. 31 , 1–10 (2025). [ Google Scholar ] 7. Ganapathy, G. et al. A blockchain based federated deep learning model for secured data transmission in healthcare iot networks. Measurement Sensors 33 , 101176 (2024). [ Google Scholar ] 8. Khan, R. et al. Advanced federated ensemble internet of learning approach for cloud based medical healthcare monitoring system. Sci. Rep. 14 , 26068 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Farooq, K. et al. Blockchain federated learning for in-home health monitoring. Electronics 12 , 136 (2022). [ Google Scholar ] 10. Wassan, S. et al. Gradient boosting for health iot federated learning. Sustainability 14 , 16842 (2022). [ Google Scholar ] 11. Lin, Y. et al. Consortium blockchain-based public integrity verification in cloud storage for iot. IEEE Internet Things J. 9 , 3978–3987 (2021). [ Google Scholar ] 12. Okegbile, S. D., Cai, J. & Alfa, A. S. Performance analysis of blockchain-enabled data-sharing scheme in cloud-edge computing-based iot networks. IEEE Internet Things J. 9 , 21520–21536 (2022). [ Google Scholar ] 13. Deebak, B. D. et al. A lightweight blockchain-based remote mutual authentication for ai-empowered iot sustainable computing systems. IEEE Internet Things J. 10 , 6652–6660 (2022). [ Google Scholar ] 14. Nguyen, B. M. et al. A novel nature-inspired algorithm for optimal task scheduling in fog-cloud blockchain system. IEEE Internet Things J. 11 , 2043–2057 (2023). [ Google Scholar ] 15. Samriya, J. K. et al. Blockchain and reinforcement neural network for trusted cloud-enabled iot network. IEEE Trans. Consum. Electron. 70 , 2311–2322 (2023). [ Google Scholar ] 16. Hewa, T., Braeken, A., Liyanage, M. & Ylianttila, M. Fog computing and blockchain-based security service architecture for 5g industrial iot-enabled cloud manufacturing. IEEE Trans. Industr. Inf. 18 , 7174–7185 (2022). [ Google Scholar ] 17. Hasan, H. R. et al. Trustworthy iot data streaming using blockchain and ipfs. IEEE access 10 , 17707–17721 (2022). [ Google Scholar ] 18. Qiu, C., Yao, H., Jiang, C., Guo, S. & Xu, F. Cloud computing assisted blockchain-enabled internet of things. IEEE Trans. Cloud Comput. 10 , 247–257 (2019). [ Google Scholar ] 19. Villegas-Ch, W., Gutierrez, R., Navarro, A. M. & Mera-Navarrete, A. Lightweight blockchain for authentication and authorization in resource-constrained iot networks. IEEE Access (2025). 20. Anbarkhan, S. H. Securing iot networks: A post-quantum blockchain and deep learning approach for enhanced cyber defense. Int. J. Saf. Secur. Eng. 14 (2024). 21. He, L. et al. A post-quantum blockchain and autonomous ai-enabled scheme for secure healthcare information exchange. IEEE J. Biomed. Health Inf. (2025). [ DOI ] [ PubMed ] 22. Akoramurthy, B. et al. Blockchain-based federated learning in internet of health things. In Federat. Learn. Digital Healthcare Syst. 175–201 (Elsevier, 2024). 23. Bataineh, M. R., Mardini, W., Khamayseh, Y. M. & Yassein, M. M. B. Novel and secure blockchain framework for health applications in iot. IEEE Access 10 , 14914–14926 (2022). [ Google Scholar ] 24. Wei, Z., Lan, G., Zhao, H., Li, Z. & Ju, Z. Lattice-based certificateless proxy re-signature for iot: A computation-and-storage optimized post-quantum scheme. Sensors 25 , 4848 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Ahmed, A., Abdullah, S., Bukhsh, M., Ahmad, I. & Mushtaq, Z. An energy-efficient data aggregation mechanism for iot secured by blockchain. IEEE Access 10 , 11404–11419 (2022). [ Google Scholar ] 26. Islam, M. J. et al. Blockchain-sdn-based energy-aware and distributed secure architecture for iot in smart cities. IEEE Internet Things J. 9 , 3850–3864 (2021). [ Google Scholar ] 27. Javed, H. et al. Blockchain-based logging to defeat malicious insiders: The case of remote health monitoring systems. IEEE Access 12 , 12062–12079 (2023). [ Google Scholar ] 28. Firouzi, F. et al. Fusion of iot, ai, edge-fog-cloud, and blockchain: Challenges, solutions, and a case study in healthcare and medicine. IEEE Internet Things J. 10 , 3686–3705 (2022). [ Google Scholar ] 29. Fugkeaw, S., Gupta, R. P. & Worapaluk, K. Secure and fine-grained access control with optimized revocation for outsourced iot ehrs with adaptive load-sharing in fog-assisted cloud environment. IEEE Access (2024). 30. Dewangan, N. K. & Chandrakar, P. Patient-centric token-based healthcare blockchain implementation using secure internet of medical things. IEEE Trans. Comput. Social Syst. 10 , 3109–3119 (2022). [ Google Scholar ] 31. Wu, T. et al. Blockchain-based anonymous data sharing with accountability for internet of things. IEEE Internet Things J. 10 , 5461–5475 (2022). [ Google Scholar ] 32. Zhang, G., Chen, P. & Liao, Y. Blockchain-based secure and verifiable deduplication scheme for cloud-assisted internet of things. IEEE Internet Things J. 11 , 13995–14006 (2023). [ Google Scholar ] 33. Li, M. et al. Cloud-edge collaborative resource allocation for blockchain-enabled internet of things: A collective reinforcement learning approach. IEEE Internet Things J. 9 , 23115–23129 (2022). [ Google Scholar ] 34. Halgamuge, M. N., Munasinghe, G. K. & Zukerman, M. Time estimation for a new block generation in blockchain-enabled internet of things. IEEE Trans. Netw. Serv. Manage. 21 , 535–557 (2023). [ Google Scholar ] 35. Alrubei, S. M., Ball, E. & Rigelsford, J. M. A secure blockchain platform for supporting ai-enabled iot applications at the edge layer. IEEE Access 10 , 18583–18595 (2022). [ Google Scholar ] 36. Yu, J. et al. An efficient revocable and searchable ma-abe scheme with blockchain assistance for c-iot. IEEE Internet Things J. 10 , 2754–2766 (2022). [ Google Scholar ] 37. Tan, S. et al. Icentia11k single lead continuous raw electrocardiogram dataset. PhysioNet (2022). 38. arashnic, I. https://www.kaggle.com/datasets/arashnic/fitbit . Kaggle (2016). 39. Mazid, A., Kirmani, S., Manaullah & Yadav, M. Fl-idpp: A federated learning based intrusion detection approach with privacy preservation. Trans. Emerg. Telecommun. Technol. 36 , e70039 (2025). 40. Mazid, A., Kirmani, S., Abid, M. & Pawar, V. A secure and efficient framework for internet of medical things through blockchain driven customized federated learning. Clust. Comput. 28 , 225 (2025). [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The data that support the findings of this study are available in the links: https://physionet.org/content/icentia11k-continuous-ecg/1.0/ -https://doi.org/10.13026/kk0v-r952 https://www.kaggle.com/datasets/arashnic/fitbit Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (3.5 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top