Sufriyana Herdiantri, Amani Fariska Zata, Al Hajiri Aufar Zimamuz Zaman, Wu Yu-Wei, Su Emily Chia-Yu
Institute of Biomedical Informatics, College of Medicine, National Yang Ming Chiao Tung University, 155 Section 2 Linong Street, Taipei, 112304, Taiwan.
Graduate Institute of Biomedical Informatics, College of Medical Science and Technology, Taipei Medical University, 250 Wu-Xing Street, Taipei, 11031, Taiwan.
Sci Rep. 2025 Mar 11;15(1):8340. doi: 10.1038/s41598-025-92986-7.
Prevention of fetal growth restriction/small for gestational age (FGR/SGA) is adequate if screening is accurate. Ultrasound and biomarkers can achieve this goal; however, both are often inaccessible. This study aimed to develop, validate, and deploy a prognostic prediction model for screening FGR/SGA using only medical history. From a nationwide health insurance database (n = 1,697,452), we retrospectively selected visits to 22,024 healthcare providers of primary, secondary, and tertiary care. This study used machine learning (including deep learning) to develop prediction models using 54 medical-history predictors. After evaluating model calibration, clinical utility, and explainability, we selected the best by discrimination ability. We also externally validated the models using geographical and temporal splits of ~ 20% of the selected visits. The models were also compared with those from previous studies, which were rigorously selected by a systematic review of Pubmed, Scopus, and Web of Science. We selected 169,746 subjects with 507,319 visits for predictive modeling from the database, which were 12-to-55-year-old female insurance holders who used the healthcare services. The best prediction model was a deep-insight visible neural network. It had an area under the receiver operating characteristics curve of 0.742 (95% confidence interval 0.734 to 0.750) and a sensitivity of 49.09% (95% confidence interval 47.60-50.58% using a threshold with 95% specificity). The model was competitive against the previous models of 30 eligible studies of 381 records, including those using either ultrasound or biomarker measurements. We deployed a web application to apply the model. Our model used only medical history to improve accessibility for FGR/SGA screening. However, future studies are warranted to evaluate if this model's usage impacts patient outcomes.
如果筛查准确，预防胎儿生长受限/小于胎龄儿（FGR/SGA）就是充分的。超声检查和生物标志物可以实现这一目标；然而，这两者往往难以获得。本研究旨在开发、验证并应用一种仅使用病史来筛查FGR/SGA的预后预测模型。我们从一个全国性的健康保险数据库（n = 1,697,452）中，回顾性地选取了对22,024名初级、二级和三级医疗保健提供者的就诊记录。本研究使用机器学习（包括深度学习），利用54个病史预测因子来开发预测模型。在评估模型校准、临床实用性和可解释性后，我们根据鉴别能力选出了最佳模型。我们还使用所选就诊记录中约20%的地理和时间分割对模型进行了外部验证。这些模型还与之前研究中的模型进行了比较，那些模型是通过对PubMed、Scopus和Web of Science进行系统综述严格挑选出来的。我们从数据库中选取了169,746名受试者的507,319次就诊记录用于预测建模，这些受试者是使用医疗服务的12至55岁女性保险持有人。最佳预测模型是一个深度洞察可见神经网络。其受试者操作特征曲线下面积为0.742（95%置信区间为0.734至0.750），敏感性为49.09%（使用具有95%特异性的阈值时，95%置信区间为47.60 - 50.58%）。该模型与之前30项符合条件的研究中381条记录的模型具有竞争力，包括那些使用超声或生物标志物测量的模型。我们部署了一个网络应用程序来应用该模型。我们的模型仅使用病史来提高FGR/SGA筛查的可及性。然而，未来有必要进行研究以评估该模型的使用是否会影响患者的预后。
