基于卷积神经网络的房颤患者导管消融术后复发预测

Prediction of recurrence after catheter ablation in patients with atrial fibrillation based on convolutional neural network

通过采集患者术前的基础病史信息、影像检查信息、生化检查信息等资料,利用统计学和卷积神经网络相结合的方法对导管消融术预后情况进行预测。本研究中纳入了121例经射频消融手术治疗后的房颤患者,利用深度学习,先将生化检查的60个指标通过调整结构与参数建立3个房颤复发预测模型,复发预测精度最高为0.7(95%CI:0.536~0.864)。然后,将基础病史资料特征信息、影像检查信息进行统计学筛选和数据标准化处理,根据P值将差异性最大的10个特征与生化检查的60个特征融合,进行多因素跨模态的深度学习,建立3个深度模型,得到的房颤复发预测模型最高准确率为0.8(95%CI:0.657~0.943)。通过多组实验发现,深度模型并非越复杂越好,在样本量有限的情况下,选取合理的模型复杂度,并纳入多种模态特征可以获得更高的预测精度。

The basic medical history,and the results of imaging examination and biochemical examination of patients before operation are collected for predicting the prognosis of catheter ablation by the combination of statistics and convolution neural network.A total of 121 patients with atrial fibrillation(AF)after radiofrequency ablation were enrolled in this study.The 60 indexes of biochemical examination are used for deep learning to establish 3 different prediction models of AF recurrence by adjusting the structure and parameters,and a recurrence prediction accuracy up to 0.7 is achieved(95%CI:0.536-0.864).Then,statistical screening and data standardization are performed on the characteristic information of basic medical history and image examination information.According to the P value,the 10 features with the largest difference are combined with the 60 features of biochemical examination to carry out multi-factor cross-modal in-depth learning.The highest accuracy of the AF recurrence prediction model obtained from the 3 models reaches 0.8(95%CI:0.657-0.943).Through multiple groups of experiments,it is found that the deep learning model is not the more complex the better.In the case of limited sample size,selecting a reasonable model complexity and incorporating multiple modal features can obtain higher prediction accuracy.