Objective: To establish a disease risk prediction model for the newborn screening system of inherited metabolic diseases by artificial intelligence technology. Methods: This was a retrospectively study. Newborn screening data (n=5 907 547) from February 2010 to May 2019 from 31 hospitals in China and verified data (n=3 028) from 34 hospitals of the same period were collected to establish the artificial intelligence model for the prediction of inherited metabolic diseases in neonates. The validity of the artificial intelligence disease risk prediction model was verified by 360 814 newborns' screening data from January 2018 to September 2018 through a single-blind experiment. The effectiveness of the artificial intelligence disease risk prediction model was verified by comparing the detection rate of clinically confirmed cases, the positive rate of initial screening and the positive predictive value between the clinicians and the artificial intelligence prediction model of inherited metabolic diseases. Results: A total of 3 665 697 newborns' screening data were collected including 3 019 cases' positive data to establish the 16 artificial intelligence models for 32 inherited metabolic diseases. The single-blind experiment (n=360 814) showed that 45 clinically diagnosed infants were detected by both artificial intelligence model and clinicians. A total of 2 684 cases were positive in tandem mass spectrometry screening and 1 694 cases were with high risk in artificial intelligence prediction model of inherited metabolic diseases, with the positive rates of tandem 0.74% (2 684/360 814)and 0.46% (1 694/360 814), respectively. Compared to clinicians, the positive rate of newborns was reduced by 36.89% (990/2 684) after the application of the artificial intelligence model, and the positive predictive values of clinicians and artificial intelligence prediction model of inherited metabolic diseases were 1.68% (45/2 684) and 2.66% (45/1 694) respectively. Conclusion: An accurate, fast, and the lower false positive rate auxiliary diagnosis system for neonatal inherited metabolic diseases by artificial intelligence technology has been established, which may have an important clinical value.
目的： 应用人工智能技术建立新生儿遗传代谢病疾病风险评估模型并验证其用于辅助新生儿串联筛查结果的判断。 方法： 回顾性研究。收集2010年2月至2019年5月来自全国31家医院新生儿遗传代谢病筛查（串联质谱法）5 907 547例数据和34家医院临床确诊的3 028例数据进行回顾性分析，建立新生儿遗传代谢病人工智能疾病预测模型；以2018年1至9月浙江大学医学院附属儿童医院360 814例新生儿筛查数据进行遗传代谢病人工智能疾病预测模型的单盲试验验证，通过比较临床确诊病例的检出率、串联初筛阳性率和阳性预测值在人工判读和遗传代谢病人工智能预测模型中的结果，验证人工智能疾病风险预测模型的有效性。 结果： 经数据筛选，共有3 665 697例新生儿串联初筛数据符合数据库建模的标准，选取所有临床确诊患儿数据3 019例共构建了16种人工智能预测模型可涵盖32种遗传代谢病；在单盲试验验证入组的360 814例新生儿中，临床确诊病例共45例，人工判读和遗传代谢病人工智能预测模型结果一致，所有临床确诊病例均为阳性或高风险。串联初筛阳性人工判读为2 684例，遗传代谢病人工智能疾病风险预测模型判读为串联初筛高风险1 694例，串联初筛阳性率分别为0.74%（2 684/360 814）、0.46%（1 694/360 814）；与人工判读相比，遗传代谢病人工智能疾病风险预测模型判读阳性人数总体减少了36.89%（990/2 684）；人工判读和遗传代谢病人工智能疾病风险预测模型的阳性预测值分别为1.68%（45/2 684）、2.66%（45/1 694）。 结论： 所建立的新生儿遗传代谢病人工智能疾病风险预测模型具有准确、快速、假阳性率低的优点，具有重要临床应用价值。.