深度神经网络在不规则弥漫大B细胞淋巴瘤时间序列数据分类预测中的应用

Application of Deep Neural Networks into Classification in Irregular Time Series Data of Patients with Diffuse Large B-cell Lymphoma

目的探讨深度神经网络在不规则时间序列数据中的分类效果,并对山西某医院2014-2020年362例弥漫大B细胞淋巴瘤(diffuse large B-cell lymphoma,DLBCL)患者进行复发预测。方法回顾性地收集了确诊且治疗后达到完全缓解的362例DLBCL患者的病例资料,并预测其两年内的复发。先利用LASSO回归进行变量的筛选,再构建基于GRU-ODE-Bayes(gated recurrent unirt-ordinary differential equation-Bayes)的不规则时间序列深度神经网络模型,并与传统模型及其他深度神经网络模型进行比较。结果在本文的所有模型中,传统模型的分类性能不及深度神经网络模型。其中GRU-ODE-Bayes模型最优,其AUC为0.85,灵敏度为0.84,特异度为0.71,G-means为0.77。结论关于不规则DLBCL时间序列数据,与本文其他模型相比,GRU-ODE-Bayes模型可以更精准地预测DLBCL患者的复发情况,可为患者个性化治疗和医生决策提供参考。

Objective To investigate the classification effect of deep neural networks in irregular time series data,and to predict the recurrence risk of 362 patients with diffuse large B-cell lymphoma(DLBCL)in a hospital in Shanxi from 2014 to 2020.Methods A total of 362 diagnosed DLBCL patients who achieved complete remission after initial chemotherapy were collected retrospectively,and the recurrence risk was predicted within the next two years.First,LASSO regression was used to screen the variables.Then a deep neural network model of irregular time series data based on GRU-ODE-Bayes was constructed and compared with some traditional models and other deep neural network models.Results Among all the models under study,the traditional models do not perform as well as the deep neural network models in classification.The GRU-ODE-Bayes model was the best,with AUC of 0.85,sensitivity of 0.84,specificity of 0.71,and G-means of 0.77.Conclusion Compared with other models,the GRU-ODE-Bayes model can predict the recurrence of DLBCL patients more accurately.It could benefit the individualized treatment for patients and decision-making for physicians.