Blockchain for medical collaboration: A federated learning-based approach for multi-class respiratory disease classification

The scarcity and diversity of medical data have made it challenging to build an accurate global classification model in the healthcare sector. The prime reason is privacy concerns and legal obstacles which limit data-sharing scope among institutions in healthcare. On the other hand, data from a single source is hardly sufficient to develop a universal diagnosis model. While federated learning is a potential solution to privacy and data diversity concerns (allows distributed model training), an apt aggregation process for multi-class and heterogenous medical data is still at the outset. This study aims to propose a federated learning mechanism that can effectively learn from multi-class and heterogenous respiratory medical data. The proposed system trains and aggregates the local model by leveraging blockchain technology, ensuring privacy. While aggregating the local models, we introduced the weight manipulation technique that, unlike any other studies, uses the local model test accuracy as the principal parameter. The resulting metric scores show that learning from diverse and heterogenous data, the performance of the proposed federated model is analogous to a single-source model (learning from single source data). Using the novel aggregation technique, the highest testing accuracy of 88.10% has been achieved for five classes, compared to the less complex single source model, which achieved 88.60% testing accuracy. A similar trend has been observed for models with three and four classes. For developing better synergy among organizations, this study introduces an incentive mechanism for the contributing institution while the blockchain stores the records to make the system transparent and trustworthy. The proposed mechanism has been implemented using a web system, which demonstrates how the weight manipulation technique can effectively learn from heterogeneous and multi-sourced data while preserving privacy.