BDSIHD: design of a blockchain-powered deep transfer learning-based highly secure IoMT data processing model for healthcare deployments
Pooja Mishra, Sandeep Malik · International Journal of Computers and Applications · 2025
Securing healthcare deployments has become one of the primary concerns for Internet-of-Medical-Things (IoMT) designers. This is due to the fact that IoMT deployments are under constant attacks from both internal & external adversaries. To incorporate attack detection, researchers have observed that blockchain-based deployments are highly efficient, due to their immutability, traceability, distributed computing & transparency characteristics. But single-chained deployments cannot be scaled due to an exponential increase in computational delays. Moreover, processing efficiency of secure IoMT-based data must be enhanced, which will assist in adding robustness to these deployments. This paper suggests creating a highly secure IoMT data processing model powered by deep transfer learning and blockchain that can be applied to multimodal healthcare installations in order to combine these features. The suggested framework optimizes blockchain write & read delays via incorporation of a Proof-of-Medical-Trust (PoMT) consensus, that is capable of distributed deployment and has inbuilt sidechain support for multichain use cases. The PoMT Model uses a Genetic Chain Optimization (GCO) method, that assists in segregating single chained data into sharded chains. Data is securely stored on these sharded chains, and is processed via a deep transfer learning (DTL) based model that fuses Gated Recurrent Unit (GRU) with Long-Short-Term Memory (LSTM), and Recurrent Neural Networks (RNNs) for classification of underlying data into different disease conditions. The classified data is further processed via a customized 2D Convolutional Neural Network (2D CNN), that helps with recognition of disease severity and progression levels via augmented analysis. The model was tested under Sybil, Spoofing, & Spying attack types, and was observed to be 10.4% faster in terms of mining performance, 6.5% faster in terms of reading performance, and showcased 9.3% better energy efficiency when compared w.r.t. standard storage models. Additionally, it was noted that the model had 2.9% higher disease classification performance, with 3.5% higher severity detection accuracy under clinical scenarios. Because of this, the model can be used for large-scale use cases & scenarios.