Machine learning prediction model for early pregnancy diagnosis
FERTILITY AND STERILITY(2023)
摘要
Approximately 20-40% of women with early pregnancy will not have a definitive diagnosis at presentation and are at a higher risk of having a non-viable pregnancy and associated morbidity. After external validation of individual biomarkers, we leveraged multiple machine learning-based methodologies to evaluate combinations of biomarkers to develop a multiplexed prediction model for early pregnancy location and viability. Based on previous discovery, assay performance, and validation studies, we first assessed the predictive ability and discrimination capacity of 24 candidate biomarkers in a case control study of patients with definitive intrauterine pregnancy (IUP) n=75, pregnancy loss (SAB) n=75, or ectopic pregnancy (EP) n=68 by Area Under the Curve (AUC) with 95% Confidence Intervals and two-sample t-tests. We then utilized machine learning methods including classification and regression tree analysis (CART), random forest (RF), and logistic regression of neural networks to evaluate combinations of the 11 best markers. Analyses were performed to maximize sensitivity, sensitivity, and accuracy of predicting both pregnancy location (EP vs. IUP and SAB) and viability (IUP vs. EP and SAB). 11 biomarkers with an AUC of >0.7 and a p value of <0.001 were candidates for the development of the multiplexed prediction model. Using 10 markers, RF predicted viability in 65% of patients with 97% accuracy. When only 6 markers were used, RF predicted viability in 69-70% with 93-94% accuracy. For the prediction of location, using all 11 markers predicted the outcome in 65% of cases with 97% accuracy. RF using 6 markers predicted the outcome in 68-69% of cases with 93-94% accuracy. When models maximizing accuracy for location and viability were used serially, CART predicted the outcome in 73% of cases with 96% accuracy. We have demonstrated that a small pool of biomarkers used in combination can aid in the prediction of early pregnancy outcome. Models balancing parsimony, simplicity, and accuracy may best negate consequences from both false positive and negative predictions.
更多查看译文
关键词
machine learning prediction model,prediction model,machine learning,pregnancy
AI 理解论文
溯源树
样例
![](https://originalfileserver.aminer.cn/sys/aminer/pubs/mrt_preview.jpeg)
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要