Blockchain Meets Federated Learning to Keep Medical Records Private

Electronic health records have transformed modern medicine, allowing clinicians to track a patient’s history across visits, monitor chronic conditions in real time, and feed powerful artificial intelligence models with the data they need to detect disease earlier than any human could. Yet the same digital infrastructure that makes healthcare smarter has also made patient information one of the most attractive targets for cybercriminals. A new study published in the journal Knowledge and Information Systems proposes an ambitious answer to this dilemma: a framework that combines a novel blockchain consensus mechanism with federated learning and an optimized neural network, so that hospitals can collaborate on diagnosis without ever exposing a single patient’s raw data.

The research, carried out by M. Lavanya of the Department of Artificial Intelligence and Data Science at Adhiparasakthi College of Engineering in Tamil Nadu, India, addresses a weakness that has long troubled electronic health monitoring systems. These systems continuously collect patient health data through connected devices, but the study argues that current implementations lack robust protection against privacy violations, security breaches, and unauthorized access. Health records contain some of the most sensitive personal information that exists, and a single breach can expose diagnoses, medications, and genetic predispositions. The proposed solution, called Platform-Free Proof of Federated Learning, or PFPoFL, is designed to secure the entire lifecycle of electronic health records, from storage to sharing, while keeping the system efficient enough for real-world hospital networks.

At the heart of the framework is federated learning, a machine learning paradigm that has been gaining traction precisely because it inverts the traditional data pipeline. Instead of pooling patient records in a central server where a model is trained, federated learning sends the model to the data. Each participating hospital trains the model locally on its own encrypted datasets, and only the learned model parameters, never the underlying patient information, are shared with the network. This means a hospital in one city can contribute to a diagnostic model that benefits from the experience of dozens of other institutions, without any of them ever seeing each other’s records. In the PFPoFL scheme, this collaborative learning process is anchored by a blockchain, the distributed ledger technology that guarantees that every transaction, every model update, and every record access is logged in a tamper-evident way that no single party can alter.

The consensus mechanism itself is a key innovation. Traditional blockchain networks, including the ones underpinning major cryptocurrencies, rely on energy-hungry proof-of-work computations or on stake-based schemes that may not suit the institutional realities of healthcare. The platform-free proof of federated learning consensus approach, which builds on earlier work published in the IEEE Journal on Selected Areas in Communications, ties the right to validate transactions to the quality and reliability of federated model contributions rather than to raw computing power or financial stakes. In a healthcare setting, this aligns incentives elegantly: hospitals earn credibility on the network by contributing useful, well-trained model updates, and the blockchain records those contributions in a way that is independent of any particular hardware platform or vendor. The study describes this as providing a secure and privacy-friendly mechanism for managing electronic health records across heterogeneous systems.

Once the data pipeline is secured, the framework turns to the intelligence layer. Before any model is trained, the system must decide which features of the health data actually matter, a problem that grows exponentially as monitoring devices stream hundreds of variables per patient. Lavanya’s solution is a Hybrid Gorilla Troops Binary Light Spectrum Optimization Algorithm, a mouthful of a name that describes the marriage of two nature-inspired optimization techniques. Gorilla troops optimization mimics the social foraging strategies of gorilla groups to explore a search space, while the binary light spectrum optimizer draws on the properties of light spectra to refine candidate solutions. Hybridizing the two produces a feature selection method that trims away redundant and noisy variables, leaving a compact subset that preserves diagnostic power while reducing computational burden and bandwidth consumption.

With the features distilled, the framework employs a Generative Adversarial Network-based classifier to perform disease diagnosis. Generative adversarial networks consist of two neural networks locked in a competitive game: a generator that learns to produce synthetic data resembling real samples, and a discriminator that learns to tell the fakes from the genuine article. In medical applications, this architecture has proven valuable not only for classification but also for handling the missing and imbalanced data that plague real-world health records. The study applies the classifier to three serious conditions: pneumonia, heart failure, and tuberculosis. These diseases were chosen, in part, because public datasets for them are available, allowing the framework to be evaluated on the Synthea synthetic healthcare dataset alongside publicly accessible disease datasets, a choice that supports reproducibility while sidestepping the ethical complications of handling real patient files.

The experimental results reported in the paper offer concrete numbers rather than vague promises. The proposed framework achieved a bandwidth efficiency of 7.2 megabits per second, which the author reports as higher than conventional techniques, an important consideration when large volumes of health data must move between institutions without saturating hospital networks. Encryption and decryption processes remained stable, completing within a window of 78 to 92 seconds, a range the study presents as evidence of consistent cryptographic performance across workloads. Beyond raw throughput, the experiments indicated improvements in accuracy, security, and scalability compared with existing approaches, suggesting that the framework can grow with a healthcare network rather than buckling under its weight.

Scalability deserves particular attention because it is where many blockchain-based healthcare proposals have stumbled. A distributed ledger grows with every transaction, and a national health network could generate millions of record accesses per day. By combining the consensus mechanism with feature selection that shrinks the data footprint before training, the PFPoFL framework attacks the problem from both ends: less data moves across the network, and the ledger validation process is designed to remain sustainable as participation expands. The study situates this within a broader research landscape that includes blockchain-assisted Internet of Things healthcare systems, homomorphic encryption approaches that allow computation on encrypted data, and federated learning models for heart disease prediction, positioning PFPoFL as an integrated alternative that addresses storage, sharing, and diagnosis in a single architecture.

The implications reach beyond the technical community. Health systems worldwide are under pressure to digitize while regulators tighten privacy rules, and patients increasingly expect that their medical histories will not become commodities traded on data markets. A framework in which hospitals collaborate on diagnostic intelligence without exchanging patient data could lower the barriers to participation in large-scale medical AI, particularly for smaller institutions that lack the infrastructure to maintain secure centralized repositories. If diagnostic models for pneumonia, heart failure, and tuberculosis can be trained on diverse populations drawn from many hospitals, they may also generalize better across demographics, reducing the biases that arise when models learn from a single institution’s patient base.

Caveats remain, as they do with any laboratory-stage framework. The evaluation relied on synthetic and public datasets, and translating such systems into live hospital environments will require integration with legacy electronic health record software, agreement among competing institutions on governance, and scrutiny by regulators and security auditors. The study itself notes no specific funding and declares no competing interests, and the author states that processed data and experimental results are available upon reasonable request. Still, the work offers a detailed blueprint for what privacy-preserving, collaborative healthcare AI could look like: a blockchain that rewards genuine model contributions, federated learning that keeps patient records at home, optimization algorithms that strip away the noise, and adversarial neural networks that turn the remaining signal into diagnoses. As health data breaches continue to make headlines, that combination may prove difficult for the industry to ignore.

Subject of Research: A blockchain-based federated learning framework for privacy-preserving electronic health record storage, sharing, and disease diagnosis

Article Title: Platform-free proof of federated learning consensus blockchain with optimized neural network-based electronic health records storage and sharing scheme

Article References: Platform-free proof of federated learning consensus blockchain with optimized neural network-based electronic health records storage and sharing scheme. (n.d.). https://doi.org/10.1007/s10115-026-02894-z

Image Credits: AI Generated

DOI: 10.1007/s10115-026-02894-z

Keywords: federated learning, blockchain, electronic health records, privacy-preserving machine learning, consensus mechanism, generative adversarial networks, feature selection, healthcare security, pneumonia diagnosis, heart failure, tuberculosis, health informatics