Extracting evidence for processing of medical claims can be time-consuming and complex due to the sharing of sensitive information among multiple organisations. This paper presents a blockchain-based framework for reliable access to sensitive data by utilizing a hybrid smart contract running on the Hyperledger blockchain platform. The framework incorporates evidential reasoning for pre-processing ambiguous legal evidence and an explainable deep neural network (DNN) model for transparent decision-making in insurance claims, which continuously learns from lawyers' input. It addresses the laws on the "Right to be Forgotten" by considering the immutability of the blockchain and the "Right to Explanation" by providing transparency despite the non-linear nature DNN. The study evaluates the proposed framework's effectiveness in pre-litigation decisions for the medical negligence of doctors, demonstrating the importance of periodic retraining using low-confidence samples annotated by lawyers to enhance the model's decision-making capability.
Medical indemnity claims safeguard healthcare professionals against costs and compensation
due to medical negligence [
The proposed framework utilises blockchain technology for secure, tamper-proof data
transactions among various organisations to guarantee privacy and security for sensitive medical
records. It employs a hybrid smart contract to execute both on-chain and off-chain computations,
which automates data access policies and records access by authorised parties. The data accessed
by blockchain is integrated with Explainable Deep Neural Network (x-DNN) models to increase
trust and transparency in medical indemnity claim decisions. Moreover, it meets essential
General Data Protection Regulation (GDPR) requirements on the "Right to be Forgotten" [
The blockchain architecture is evaluated and selected using a multi-criteria decision-making
approach [
Medical data is often ambiguous and incomplete, requiring specialised processing techniques. ER
algorithm is utilised to preprocess ambiguous data into explainable numerical features [
The reasoning behind each decision by CNN can be understood by Layer Wise Relevance Propagation (LRP) [22], Shapley Additive Explanations (SHAP) [23] and Local Interpretable Model-agnostic Explanations (LIME) [24]. Figure 2 demonstrates the methodology.
This study demonstrates the results of the proposed framework focusing on medical indemnity insurance for misdiagnosis and delayed cancer diagnosis. It evaluates the medical negligence of clinical oncologists before litigation, as most claims are re-solved outside of court through law firms. The dataset had 1888 instances and 23 attributes of medical indemnity claims; 8.52% were negligent, 19.40% were contributory-negligent, and the remaining 72.08% were non-negligent. Bayesian Optimisation trained the hyperparameters of CNN. The input sequence size (height × width) for the 1D-CNN is 84×1, followed by three convolution and pooling layers with kernel sizes of 50, 100, and 150, and two fully connected layers with 0.20 dropout rate.
Figure 3 shows the performance of Hyperledger Fabric Version 1.4 blockchain by the latency of creating a block and the number of valid transactions in a given time by throughput. A decision by CNN for a defendant (local explainability of an insurance case) can be understood by analysing the importance of features by heatmap of LRP values for four different techniques and the importance of the most relevant features by the SHAP and LIME, shown in Figure 4. Lawyers use these visual explanations as decision support. The ultimate verdict of a pre-litigation case depends on the lawyer's judgment, not the model's. The "negligence" decisions against the defendant were evaluated using the AUC metric through 3-fold cross-validation. The initial AUC score of a validation set by the original dataset is 0.86. The score improved to 0.91 after a second retraining iteration with newly annotated datasets by human lawyers. The accuracy can be improved after each iteration containing 100 new legal cases. The annotation activity is performed only for the least confident and unknown legal cases.
A blockchain-based framework is proposed in this study to ensure secure access to sensitive medical records across multiple organisations. It automates and enforces data access policies while incorporating an x-DNN model for transparent decision-making in medical indemnity claims, complying with GDPR requirements. The methodology is evaluated in pre-litigation decisions regarding clinical negligence by oncologists, emphasizing the significance of periodic retraining with low-confidence samples annotated by lawyers to enhance decision-making capabilities. The results show-case improved transparency, data privacy and security in medical indemnity claims.
This work is supported by the Department of Finance and Accounting at the University of Liverpool Management School. We would like to express our appreciation to the medico-legal lawyers and doctors who provided valuable feedback on the model output and decision-making variables, without whom this research would not have been possible. We extend our gratitude to Dr William Reed, Dr Mike Pierce, and Mr. Sameer Joshi for their invaluable contributions in augmenting and annotating the model, which improved the end-to-end understanding and accuracy of the models. Human Knowledge by Fuzzy Cognitive Maps: Small Business Loans." In 2023 IEEE International Conference on Blockchain and Cryptocurrency (ICBC), pp. 1-4. IEEE, 2023. [20] Turchi, Marco, Matteo Negri, M. Farajian, and Marcello Federico. "Continuous learning from human post-edits for neural machine translation." The Prague Bulletin of Mathematical Linguistics 108, no. 1 (2017): 233-244. [21] Tornetta, Gabriele N. "Entropy methods for the confidence assessment of probabilistic classification models." arXiv preprint arXiv:2103.15157 (2021). [22] Bach, Sebastian, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. "On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation." PloS one 10, no. 7 (2015): e0130140. [23] Lundberg, Scott M., and Su-In Lee. "A unified approach to interpreting model predictions."
Advances in neural information processing systems 30 (2017). [24] Ribeiro, Marco Tulio, Sameer Singh, and Carlos Guestrin. "Why should I trust you?" Explaining the predictions of any classifier." In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pp. 1135-1144. 2016.