Construction of a pregnancy prediction model in acupuncture treatment for diminished ovarian reserve based on machine learning

通过机器学习算法构建针刺治疗卵巢储备功能减退 (Diminished Oviarian Reserve, DOR) 的临床预测模型, 为针刺改善DOR妊娠结局临床预测. 本文数据选自国际针灸病例注册登记研究平台International Patient Registry Platform Of Acupuncture-moxibustion(IPRPAM) 平台导出的377例接受针刺治疗且记录妊娠结局 (包括获得妊娠者139例以及未获得妊娠者238例) 的DOR患者.采用Spearman相关性分析及特征工程方法确定预测变量;采用Logistic回归,朴素贝叶斯,随机森林,支持向量机,极端梯度增强,k-近邻算法和线性判别分析以及神经网络方法进行模型构建;通过曲线下面积(Area Under Curve, AUC),准确率 (Accuracy, ACC) 以及模型重要性排序进行模型验证, 并对性能最好的模型进行个性化妊娠预测. 影响DOR患者针刺后是否妊娠的关键因素依次为:年龄,治疗后促黄体生成素 (LH) ,治疗后促卵泡生成素 (FSH) ,治疗后FSH/LH比值,针灸治疗史.随机森林模型的ACC值为0.95, Fβ为0.93, Logloss值为0.30, 整体模型变量的准确性和精度, 且逻辑回归损失函数最低. 基于IPRPAM平台构建的针刺对DOR患者妊娠结局影响的随机森林模型, 具有较好的临床应用价值. ChiCTR2200062293. To construct a clinical prediction model of acupuncture treatment for diminished ovarian reserve (DOR) based on a machine learning algorithm to provide a clinical prediction of acupuncture for ameliorating pregnancy outcomes in DOR. We enrolled 377 DOR patients treated with acupuncture and with records of pregnancy outcomes (139 cases of pregnancy and 238 cases failed) exported from the International Patient Registry Platform of Acupuncture-moxibustion (IPRPAM). The predictive variables were determined using Spearman's correlation analysis and feature engineering methods. The model was constructed by adopting logistic regression, naïve Bayes, random forest, support vector machine, extreme gradient boosting, the k-nearest neighbor algorithm, linear discriminant analysis, and neural network methods. The models were validated by the area under the curve (AUC), accuracy (ACC), and importance sequencing, and individual pregnancy prediction was conducted for the best-performing model. The key factors determining pregnancy after acupuncture in patients with DOR were age, luteinizing hormone (LH) level after treatment, follicle-stimulating hormone (FSH) level after treatment, the ratio of FSH to LH (FSH/LH) after treatment, and history of acupuncture treatment. Random forest model ACC was 0.95, Fβ was 0.93, Logloss was 0.30, Logloss value was the lowest, the model variables exhibited the highest accuracy and precision. The random forest model for the effects of acupuncture on pregnancy outcomes in patients with DOR, constructed based on the IPRPAM, presents a favorable value for clinical application. ChiCTR2200062293.