基于CNN-Transformer-BPNN组合模型的糖尿病预测研究

Research on Diabetes Prediction Based on CNN-Transformer-BPNN Combination Model

摘要: 本文使用SMOTE方法对Kaggle公开的糖尿病数据进行过采样，并按照8:2的比例划分训练集和测试集后进行学习、建模和预测。在Transformer模型的基础上，采用双分支并行处理结构，在两侧分支分别加入CNN模型和BPNN模型，建立CNN-Transformer-BPNN组合模型，该组合模型结合了CNN的局部特征捕获能力、Transformer的全局理解能力以及BPNN的非线性映射优势，其AUC值、F2-score、精确值分别为0.9615、0.9944、0.9614，预测效果显著。本文建立的CNN-Transformer-BPNN组合预测模型可以为糖尿病早期诊察提供可靠的临床辅助，便于对病患进行及时预警和干预，对糖尿病的诊治和我国医疗系统的发展都有积极促进作用。

Abstract: In this paper, the SMOTE method is used to oversample the diabetes data published by Kaggle, and the training set and test set are divided according to the ratio of 8:2 for learning, modeling and prediction. On the basis of the Transformer model, a CNN-Transformer-BPNN combined model is established by using a two-branch parallel processing structure and adding CNN model and BPNN model to the two branches on both sides. The combined model combines the local feature capture ability of CNN, the global understanding ability of Transformer and the nonlinear mapping advantage of BPNN. The AUC value, F2-score and accurate value are 0.9615, 0.9944 and 0.9614, respectively, and the prediction effect is remarkable. The CNN-Transformer-BPNN combined prediction model established in this paper can provide reliable clinical assistance for early diagnosis of diabetes, facilitate timely early warning and intervention for patients, and have a positive effect on the diagnosis and treatment of diabetes and the development of China’s medical system.