基于人工神经网络的卵巢早衰预测模型研究

Prediction Model of Premature Ovarian Failure Based on Artificial Neural Network

目的建立基于人工神经网络(ANN)的卵巢早衰(POF)预测模型——多层向前神经网络模型,以期提高POF临床诊断总符合率。方法 2011年1—3月选取武汉市白玉山街所管辖的6个社区内符合纳入标准的妇女341例为研究对象。2011年5月—2016年6月,每隔4个月对研究对象进行1次来院随访,随访至其40岁。随访过程中2例研究对象行子宫切除术,2例服用性激素治疗,失访21例,均予以剔除,最终共纳入316例研究对象。采用无偏随机化分配法将316例研究对象分为训练样本(177例)、检验样本(44例)和坚持样本(95例)。设置输入参数为A型行为、腮腺炎病史、妇科手术史、使用促排卵药物史、婚育史、卵泡刺激素(FSH)、FSH/黄体生成素(LH)、抗苗勒管激素(AMH)、抑制素B(INHB)、窦状卵泡数(AFC)、收缩期峰流速(PSV)、阻力指数(RI);输出参数为"是否发生POF"。通过训练样本进行模型构建,检验样本对模型进行校正,坚持样本对模型进行稳定性检测。结果ANN经过剔除"冗余"后,自动构建出输入单元(12个)、单隐层(6个节点)和激活函数(hyperbolic tangent)、输出单元(2个)和激活函数(softmax)的模型。训练样本的交叉熵误差值为53.236,在预测误差未减少时终止测试,训练时间为0.42 s。影响权重在前5位的输入参数分别为AMH(26.3%)、INHB(24.1%)、AFC(21.7%)、A型行为(7.2%)、妇科手术史(6.5%)。多层向前神经网络模型预测训练样本、检验样本、坚持样本发生POF的灵敏度分别为97.8%、91.7%和92.0%,特异度分别为92.4%、84.4%和80.0%,总符合率分别为93.8%、86.4%和83.2%。在训练样本和检验样本的基础上,得到多层向前神经网络模型预测POF的受试者工作特征曲线下面积(AUC)为0.972。结论基于ANN构建的POF预测模型——多层向前神经网络模型具有较高临床诊断总符合率,不仅为临床高效诊断及优化检查提供理论基础和方法支持,而且为实现早防早治提供机会,值得临床推广。

Objective To establish a multilayer forward neural network model based on artificial neural network （ANN） for predicting premature ovarian failure （POF）, in order to increase the total coincidental rate of clinical diagnosis. Methods Three hundred and forty - one women met the inclusion criteria were selected as the study subjects, who lived in six communities under the jurisdiction of Baiyushan Street in Wuhan City during January to March 2011. From May 2011 to June 2016, each case was conducted to hospital follow -up once every four months until the age of 40. During the follow -up, 2 cases underwent hysterectomy, 2 cases received sex hormone treatment, 21 cases were lost, all these cases were excluded, and finally 316 cases were included in the study. Using unbiased randomized allocation method, 316 subjects were divided into training sample （177 cases） , test sample （44 cases） and persistent sample （95 cases） . The input parameters were set as the type A behavior, mumps history, history of gynecological surgery, history of ovulation induction drugs use, marriage and birth history, follicle stimulating hormone （FSH） , FSH/luteinizing hormone （LH） and anti mullerian hormone （AMH） , inhibin B （INHB） . the number of antral follicles （AFC）, systolic peak velocity （PSV） and resistance index （RI）; the output parameter was ＂ whether POF occured＂ . The model was constructed by training sample, and corrected by the test sample. The stability of the model was tested by persistent sample. Results After eliminating the ＂ redundancy＂, ANN automatically constructed model of input unit （12）, single hidden layer （six nodes） and activation function （hyperbolic tangent）, output unit （2） and activation function （softmax） . The cross entropy error value of the training sample was 53. 236. Abort testing when the prediction error did not decrease, and the test time was 0.42 s. The input parameters affecting the weights in the top 5 were AMH （26.3%）, INHB （24.1%）, AFC （21.7%）, type A behavior （7.2%）, and history of gynecological surgery （6.5%） . The sensitivity of multilayer forward neural network model predicting POF in the training sample, test sample and persistent sample was 97.8%, 91.7% and 92.0%, respectively; the specificity was 92.4%, 84.4% and 80.0%, respectively, and the total coincidental rate was 93.8% , 86. 4% and 83.2% , respectively. On the basis of training sample and test sample, the AUC of multilayer forward neural network model predicting POF was 0. 972. Conclusion The multilayer forward neural network model based on ANN for predicting POF has a high total coincidental rate of clinical diagnosis. It not only provides a theoretical basis and method support for the clinical efficient diagnosis and optimization examination, but also provides an opportunity to realize early prevention and early treatment, and is worthy of clinical promotion.