Clinical Equivalence of Privacy-Preserving Federated Learning: A Consortium Blockchain Framework with Equivalence Testing

<title>Abstract</title> <p> <bold>Background:</bold> Hospitals want the benefits of federated learning—collaborative model train-ing without sharing patient records—but worry that adding encryption, noise, and blockchainconsensus might degrade the very clinical predictions they depend on. So far, the research com-munity has largely sidestepped this concern, measuring accuracy drops without asking whetherthose drops matter to clinicians. We believe the right question is not “how much accuracy islost?” but “is the secured model still clinically equivalent to the original?” <bold>Methods:</bold> We built a modular consortium blockchain–federated learning framework that bringstogether three lines of defence: subsampled-Gaussian differential privacy with Renyi accounting,BFV homomorphic aggregation, and a six-function Hyperledger Fabric smart-contract suite forreal-time anomaly detection and reputation scoring, all running under PBFT consensus. Westress-tested it on two real clinical benchmarks—MIMIC-III in-hospital mortality and COVID-19 radiography—partitioned across 20 simulated institutions under realistic non-IID conditions.Rather than simply comparing accuracy numbers, we adopted a Two One-Sided Tests (TOST)protocol against a 3% equivalence margin, with Holm–Bonferroni correction for multiple com-parisons. <bold>Results:</bold> The fully secured framework trailed standard federated learning by just 2.6 percentagepoints on COVID-19 accuracy and 2.5 points on MIMIC-III AUROC—both well inside theequivalence margin (TOST p &lt; 0.001; Cohen’s d = 0.18 and 0.21). These were not mere “nosignificant difference” findings; TOST allowed us to make the positive claim that the securedmodel is clinically equivalent. Against gradient-scaling attacks with 20% malicious clients, on-chain validation caught 94.3% of poisoned updates while falsely flagging only 2.1% of honestones. The blockchain overhead was modest—about 1.8 seconds per round—and scaled sub-linearly with client count. <bold>Conclusions:</bold> Strong privacy, security, and regulatory alignment can coexist with clinical utilityin federated healthcare learning. The larger message is methodological: equivalence testing offersa rigorous, reusable standard for evaluating whether privacy-preserving clinical AI is ready forthe bedside. </p>