Carbon-Budgeted Federated Learning Adds Blockchain Auditing to Healthcare AI

Researchers at Vellore Institute of Technology have built a federated learning framework that tries to solve three problems at once: keeping patient data private, holding predictive performance steady, and making the carbon cost of distributed model training visible and enforceable. The system, called Carbon-Budgeted Green Federated Learning (CB-GFL), is described in a study published in Frontiers in Digital Health. It pairs a carbon-aware client selection algorithm with a blockchain auditing layer and an ensemble of gradient-boosting models, and it was tested entirely in simulation, with every participating hospital running on a single machine.

The core idea is that federated learning, in which institutions train models locally and share only model outputs rather than raw patient records, has no built-in mechanism to track or cap the energy consumed during training. The authors argue that this gap matters because large-scale model training can generate substantial carbon emissions, and the emissions vary widely depending on the local electricity generation mix. In healthcare, where models must be retrained as clinical practice and disease patterns evolve, the authors contend that carbon budgets are rarely monitored or verified during collaborative training, and that self-reported energy figures from independent institutions cannot be substantiated without an external record.

CB-GFL addresses client selection as a constrained optimisation problem. Each communication round, the framework must decide which institutions participate while keeping total estimated emissions under a predefined carbon budget. The authors formulate this as a 0-1 knapsack problem, which is NP-hard, and solve it with a greedy approximation that sorts institutions by a carbon-efficiency ratio, the normalised sample count divided by estimated CO2 emission per round. The greedy solution carries a worst-case approximation guarantee of at least 63.2 percent of the optimal selection, a bound the authors derive from the fractional relaxation analysis of greedy knapsack algorithms. Estimated emissions for each client are updated after every round using an exponential moving average, so the selection adapts to recent measured workloads. Emissions themselves are estimated with the Green Algorithms methodology, using device power draw, training time, and regional carbon intensity; the study used an electricity grid emission factor of 0.708 kgCO2 per kWh, the 2023 figure from India’s Central Electricity Authority.

The second layer is a blockchain audit trail built from three smart contracts: a Carbon Budget Ledger that records emissions, an Energy Verification Contract that checks reported energy use against predefined criteria, and an Incentive/Penalty Contract that administers a token economy. Institutions consuming less than half their allocated budget earn ten tokens, those between 50 and 80 percent earn five, those within budget but above 80 percent earn nothing extra, and those exceeding their budget are penalised ten tokens. Every transaction is secured with SHA-256 hash links to the preceding block. Across the twelve-round training runs reported, the ledger accumulated 74 blocks, comprising 60 client transactions, 12 communication-round records, and 2 system records, and all hash links verified as intact.

On the machine learning side, the framework aggregates probability predictions rather than model parameters, because its three constituent models, LightGBM, XGBoost, and Gradient Boosting, have incompatible structures and cannot be parameter-averaged. Client predictions are weighted by local dataset size, and the final classification applies a decision threshold chosen for the intended clinical operating point. The primary experiments used the Framingham Heart Study dataset, a public cardiovascular risk benchmark with 4,240 records and 15 variables; after removing missing values, 3,658 records remained, containing 3,101 non-events and 557 coronary heart disease cases, a class imbalance of 1 to 5.57. Feature engineering expanded the 15 original variables to 84 through interaction, ratio, polynomial, logarithmic, and square-root transformations, and the training split was resampled with SMOTEENN to balance the classes. A weighted focal loss with an inverse-frequency class weight of roughly 3.3 for the minority class further shifted learning emphasis toward underrepresented positive cases.

Under the balanced operating threshold of 0.865, the framework reached 83.61 percent classification accuracy, an 82.34 percent weighted F1-score, an AUC-ROC of 0.7138, and a Matthews correlation coefficient of 0.2771 on the held-out test set, with specificity of 93.08 percent. The confusion matrix contained 578 true negatives, 43 false positives, 77 false negatives, and 34 true positives. A screening threshold of 0.375 raised sensitivity to 72.07 percent at the cost of accuracy, while a high-specificity threshold of 0.945 pushed specificity to 99.84 percent for settings where false alarms must be minimised. Five-fold cross-validation yielded a mean accuracy of 84.25 percent with a standard deviation of 1.03 percent and a mean AUC of 0.7135.

To test whether the modest Framingham numbers reflect a model weakness or a property of the data, the authors benchmarked eleven centralised algorithms on the same 84-feature dataset under identical conditions. Every algorithm, from XGBoost to naive Bayes, landed within an MCC range of 0.140 to 0.290, which the authors interpret as an intrinsic ceiling imposed by the dataset’s single-time-point clinical features rather than a deficiency of any particular method. CB-GFL’s MCC of 0.277 sits at the top of that range. Against representative baselines, the framework reported 83.61 percent accuracy versus 72.10 percent for standard FedAvg and 67.00 percent for federated learning with differential privacy, while a centralised upper-bound model reached 86.50 percent. Bonferroni-corrected Wilcoxon signed-rank tests over 20 paired runs gave p-values of 0.00001 against all baselines, with Cohen’s d effect sizes between 0.823 and 0.951.

Cross-dataset evaluation extended the framework to four additional cardiovascular datasets. On the UCI 4-site Combined dataset, which pools the Cleveland, Hungarian, Switzerland, and VA Long Beach heart disease files across four hospitals in three countries with 920 records, CB-GFL achieved an AUC of 0.9113, an MCC of 0.691, and 84.24 percent accuracy, with five-fold cross-validation confirming a mean AUC of 0.9152 and a reported PR-AUC of 0.9430. The NHANES Synthetic dataset produced an AUC of 0.859 and 86.63 percent accuracy, UCI Cleveland an AUC of 0.878, and the 70,000-record Kaggle cardiovascular dataset an AUC of 0.779. The authors note that the high AUC values often reported on UCI Cleveland, around 0.92 to 0.94, come from a near-balanced 303-sample dataset that is not directly comparable to the harder Framingham 10-year prediction task.

Two validation experiments probed the carbon machinery itself. In the main runs, institutional carbon budgets were set well above measured emissions, so every client stayed compliant and earned maximum tokens, with per-round emissions on the order of 0.00002 kgCO2. To force the selection algorithm into scarcity, the authors ran binding experiments with budgets set just below measured emissions: the algorithm excluded the Switzerland and VA Long Beach sites, the penalty contract fired automatically, and performance degraded gracefully from an AUC of 0.9113 with four clients to 0.8821 with two and 0.8412 with one. Blockchain overhead across twelve rounds measured 3.3 milliseconds, which the authors report as less than 0.001 percent of a federated round.

The authors are explicit about the limits of the demonstration. The federated environment was simulated rather than deployed across real hospitals, all clients ran on a single machine without a GPU, and real-world factors such as network latency, client dropout, and hardware heterogeneity remain untested. The cross-dataset experiments followed a lighter protocol than the primary Framingham setup, and the loose carbon budgets mean genuine exclusion under scarcity was only exercised in the dedicated binding tests. Framingham performance remains modest because of the dataset’s class imbalance and limited discriminative signal. Future work, the authors write, includes testing in real multi-hospital environments, integrating dynamic carbon-intensity forecasting and adaptive client scheduling, adding secure aggregation, and incorporating differential privacy in a way that reports minority-class sensitivity and MCC rather than accuracy alone, since privacy mechanisms can disproportionately affect imbalanced classes.

Subject of Research: A carbon-budgeted federated learning framework with blockchain auditing for privacy-preserving cardiovascular risk prediction

Article Title: CB-GFL: carbon-budgeted federated learning with blockchain auditing for healthcare applications

Article References: Bhardwaj, T., & Sumangali, K. (2026). CB-GFL: carbon-budgeted federated learning with blockchain auditing for healthcare applications. Frontiers in Digital Health, 8, Article 1961099. https://doi.org/10.3389/fdgth.2026.1961099

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1961099

Keywords: federated learning, blockchain, carbon budget, healthcare AI, cardiovascular disease, Framingham Heart Study, green AI, LightGBM, XGBoost, smart contracts, privacy-preserving machine learning, carbon emissions