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

Artificial intelligence (AI) in healthcare has made it harder to protect patient data, prediction performance, and environmental sustainability in federated learning (FL) systems. A single platform for sustainable healthcare applications, Carbon-Budgeted Green Federated Learning (CB-GFL), merges carbon-aware client selection, blockchain-based auditing, and energy-efficient collaborative learning. The framework integrates (i) a carbon-budget-aware client selection strategy formulated as a 0–1 knapsack optimisation problem with a greedy approximation guarantee of approximately 63.2%; (ii) a blockchain auditing layer consisting of three smart contracts, the Carbon Budget Ledger, Energy Verification Contract, and Incentive/Penalty Contract, secured using SHA-256 hash-based immutable records; and (iii) an ensemble learning model combining LightGBM, XGBoost, and Gradient Boosting with SMOTEENN resampling. All experiments are conducted in a simulated federated environment, with clients executed on a single machine. CB-GFL achieves AUC = 0.9113 on the UCI 4-site benchmark (4 hospitals, 3 countries, n = 920), exceeding the 0.90 threshold, and AUC = 0.7138 on Framingham, consistent with the clinical gold standard of 0.70 to 0.75. In the balanced operating point, experimental findings show 83.61% classification accuracy, 82.34% weighted F1-score, 0.7138 AUC-ROC, and 0.2771 MCC, and 72.07% sensitivity in the screening configuration. The proposed system ensures 100% carbon-budget compliance and a 74-block tamper-resistant blockchain audit trail for energy responsibility. Comparison to 11 benchmark algorithms shows that the dataset's realistic predictive qualities are 0.14–0.29. Additionally, five-fold cross-validation shows steady generalisation with an average accuracy of 84.25% ± 1.03%. The results show that CB-GFL supports sustainable AI adoption and privacy-preserving, carbon-aware, auditable federated healthcare intelligence.