Research On Diagnosis Algorithm Of Type 2 Diabetes Based On Pulse Wave

Type 2 diabetes is a global epidemic that can lead to severe complications.Early and accurate detection helps reduce the risk of complications.Artificial intelligence-based diabetic pulse diagnosis,as a non-invasive,rapid,and economical detection method,is suitable for large-scale screening and has become a research hotspot.However,current models have shortcomings in terms of lightweight design,generalization ability,and noise robustness.This study addresses these issues and the main contributions are as follows:Firstly,this thesis proposes a single-channel pulse wave diabetes diagnosis algorithm based on Temporal Multi-level Convolutional Network(TMC-Net),aiming to enhance generalization performance and adaptability for resource-constrained devices.TMC-Net extracts sequential features and multi-level local features through Temporal-block and multiple convolutional layers,respectively,and employs global max-pooling to extract global features of local features at each level to reduce parameter quantity.Experiments on the PPG-BP dataset show an accuracy of 99.96%and 80.65% for intra-individual and inter-individual scenarios,respectively.Secondly,to address the limitation of single-channel pulse wave information,this thesis proposes a multi-channel pulse wave diabetes diagnosis algorithm based on Transformer and Temporal Multi-level Convolutional Network(Trans TMC-Net).Trans TMC-Net combines the advantages of Transformer and TMC-Net,integrating their excellent capabilities in channel feature and temporal feature extraction.The algorithm automatically extracts diabetes-related features from pulse waves without denoising.Experimental results demonstrate that the proposed algorithm achieves an inter-individual accuracy of 91.66% on the constructed multi-channel pulse wave dataset.In conclusion,this thesis proposes two diabetes diagnosis algorithms based on pulse waves for single-channel and multi-channel pulse wave,respectively.Experimental results show that the proposed algorithms have high performance in terms of inter-individual generalization,providing new insights for early diagnosis of diabetes.