融合注意力机制和BP神经网络的2型糖尿病风险预测研究

Study on risk prediction of type 2 diabetes with integration of attention mechanism and BP neural network

目的基于融合注意力机制和BP神经网络模型建立2型糖尿病风险预测模型,为糖尿病高危人群提供个性化的预测。方法于2019年7月—2021年2月对安徽省蚌埠市社区的慢性病队列进行纵向随访研究,共纳入2334名研究对象。采用Python 3.9软件进行分析,利用基于Python的深度学习框架PyTorch来实现注意力机制和BP神经网络结构。以是否患2型糖尿病为因变量,将数据集按照8︰2的比例划分训练集和测试集,使用融合注意力机制和BP神经网络模型进行建模与分析,并与经典的机器学习模型对比预测性能。基于准确率、精确率、召回率、F1分数以及训练过程中绘制的损失值曲线和准确率曲线进行模型评价。结果融合注意力机制和BP神经网络模型在训练集中准确率为0.9630,精确率为0.9725,召回率为0.9572,F1分数为0.9648;在测试集中准确率为0.9722,精确率为0.9756,召回率为0.9683,F1分数为0.9719。在训练过程中,融合注意力机制和BP神经网络模型得到的损失值曲线和准确率曲线均优于对比模型。结论相较于BP神经网络模型和传统机器学习模型,融合注意力机制和BP神经网络模型的预测性能更佳,可及时准确识别2型糖尿病患者,实现2型糖尿病的早发现和早治疗,以预防并减缓对其身体带来的危害。

Objective To establish a risk predictive model of type 2 diabetes based on the integration of attention mechanism and BP neural network,and to provide personalized predictions for individuals at high risk of diabetes.Methods The study subjects were derived from a longitudinal follow-up study of the chronic disease cohort in the community of Bengbu City,Anhui Province,conducted from July 2019 to February 2021,with a total of 2,334 participants enrolled.Python 3.9 software was utilized for analysis,the attention mechanism and BP neural network structure was realized by using the PyTorch deep learning framework that based on Python.The occurrence of type 2 diabetes was set as the dependent variable,and the data set was split into a training set and a test set in an 8:2 ratio.The integrated attention mechanism and BP neural network model were employed for modeling and analysis,and comparing predictive performance with classical machine learning models.Model evaluation was based on accuracy,precision,recall rate,F1 score,as well as the loss and accuracy curves plotted during the training process.Results The integrated attention mechanism and BP neural network model achieved an accuracy of 0.9630,precision of 0.9725,recall rate of 0.9572,and F1 score of 0.9648 on the training set.On the test set,the accuracy was 0.9722,precision was 0.9756,recall rate was 0.9683,and F1 score was 0.9719.Throughout the training process,the loss and accuracy curves of the integrated attention mechanism and BP neural network model outperformed the comparative models.Conclusions Compared to the BP neural network models and traditional machine learning models,the model that integrates an attention mechanism with the BP neural network exhibits better predictive performance.It can promptly and accurately identify patients with type 2 diabetes,enabling early discovery and treatment,thereby preventing and reducing the harm to their health.