Zhang Xiang, Shi Yu-ying
Beijing Tongren Eye Center, Affiliated Tongren Hospital, Capital University of Medical Sciences, Beijing 100730, China.
Zhonghua Yan Ke Za Zhi. 2007 Nov;43(11):987-95.
To train and evaluate a backpropagation (BP) neural network to predict the pseudophakic refraction of a child at any age.
The clinical data of paediatric pseudophakia were consecutively collected from the patients for subsequent visits to Cataract Center of Beijing Tongren Hospital during June to October in 2006 and 70 eyes of 41 patients that met the inclusion criteria were identified. We reviewed the case history, preoperative examinations, surgical process and follow-up results of these patients and recorded the main following data: axial length and corneal curvature of both eyes before intraocular lens (IOL) implantation surgery, targeted postoperative refraction, IOL power, laterality, age at cataract extraction and IOL implantation surgery, age and refraction at last follow up. 70 eyes were divided into a training set and a test set by simple random sampling. The training set of 55 eyes was used for training a BP neural network and updating the network weights and biases. The test set of 15 eyes was used to work out the test set prediction of the pseudophakic refraction at last follow up, which was compared with that produced by a logarithmic regression advanced by McClatchey and his colleagues.
For the test data, the correlation between network outputs and target outputs was statistically significant (r = 0.603, P = 0.017); The difference between network outputs and target outputs was not statistically significant (paired-samples t test, P = 0.270). Mean error and mean absolute error from predicted refraction were +0.69 diopters (D) and 1.34 D by BP neural network respectively and were +1.03 D and 1.98 D by logarithmic regression respectively. The differences in predictive errors and absolute errors between two predictive methods were not significant but in absolute errors the P value was close to 0.05 (P = 0.075) by paired-samples t test. The predictions by two predictive methods both underestimated the myopic shift of paediatric pseudophakia and the prediction by logarithmic regression tended towards more hyperopia.
BP neural network improved prediction of pseudophakic refraction of a child at any age compared with the logarithmic regression advanced by McClatchey and his colleagues in this study. It can be a useful tool in predicting myopic shift in paediatric pseudophakia.
训练并评估一个反向传播（BP）神经网络，以预测任何年龄儿童的人工晶状体眼屈光状态。
连续收集北京同仁医院白内障中心2006年6月至10月期间复诊患者的小儿人工晶状体眼临床资料，确定41例符合纳入标准患者的70只眼。我们回顾了这些患者的病历、术前检查、手术过程及随访结果，并记录了以下主要数据：人工晶状体（IOL）植入手术前双眼的眼轴长度和角膜曲率、目标术后屈光状态、IOL度数、眼别、白内障摘除及IOL植入手术时的年龄、末次随访时的年龄及屈光状态。通过简单随机抽样将70只眼分为训练集和测试集。训练集的55只眼用于训练BP神经网络并更新网络权重和偏置。测试集的15只眼用于计算末次随访时人工晶状体眼屈光状态的测试集预测值，并与McClatchey及其同事提出的对数回归预测值进行比较。
对于测试数据，网络输出与目标输出之间的相关性具有统计学意义（r = 0.603，P = 0.017）；网络输出与目标输出之间的差异无统计学意义（配对样本t检验，P = 0.270）。BP神经网络预测屈光状态的平均误差和平均绝对误差分别为+0.69屈光度（D）和1.34 D，对数回归分别为+1.03 D和1.98 D。两种预测方法在预测误差和绝对误差方面的差异无统计学意义，但配对样本t检验显示绝对误差的P值接近0.05（P = 0.075）。两种预测方法的预测值均低估了小儿人工晶状体眼的近视漂移，且对数回归的预测值更倾向于远视。
在本研究中，与McClatchey及其同事提出的对数回归相比，BP神经网络改善了对任何年龄儿童人工晶状体眼屈光状态的预测。它可成为预测小儿人工晶状体眼近视漂移的有用工具。
