Federated learning with IoT-based remote patient monitoring for real-time chronic disease prediction

Abstract

Chronic kidney disease (CKD) is a major health issue in the world, and there is a need to apply quick and accurate diagnosis methods to reduce chronic disease. The processing of any medical information is limited in the centralized location due to the significant latency, privacy concerns, and signal variability produced by wearable Internet-of-Things (IoT) devices. In a desire to overcome these shortcomings, the current study suggests a federated learning (FL)-based IoT to predict CKD in real-time preserving privacy. Nutcracker Optimization (ENO) heuristic was further used to select salient features out of CKD datasets, which were pre-trained convolutional architectures, such as ResNet, EfficientNet, AlexNet, and Inception, to extract high-dimensional, discriminative features. The final predictive step was a multi-head deep neural network (M-DNN), and the model was trained locally on distributed devices by use of FL, thus avoiding the exodus of sensitive patient data offsite. The evaluation of the framework was done on both independently and identically distributed (IID) and non-IID datasets, and the framework performed better in each instance. It is worth mentioning that ResNet + ENO + FL + M-DNN pipeline has achieved accuracy of 99.997%, which is higher than the KNORA-E model. AlexNet, EfficientNet, and Inception models also showed better results indicated by high-precision, recall, and area under the receiver operating characteristic curve (AUC). These results support the idea that the combination of ENO and FL significantly improves the predictive accuracy, privacy protection, and greatest use of computational resources, making the method highly applicable in the case of safe, real-time distantly monitoring CKD patients.