Predictive Analytical Model for Ectopic pregnancy diagnosis: Statistics vs. Machine learning methods.

Background: Ectopic pregnancy is well-known for its serious outcome. Early detection could make the difference between life and death in pregnancy. Although, prompt diagnosis was challenged and numbers of predictive models had been developed, there’s still lack of gold standard tool. Our research introduces predictive analytical models using both conventional statistics and machine learning methods based on all three domains of features (clinical factors, serum human chorionic gonadotropin and ultrasound findings). Methods: Retrospective cohort study on 377 pregnancy of unknown location women, training and validating for model for prediction of ectopic pregnancy outcome, using set of 22 features. Analysis performed by three difference machine learning models neural networks (NNs), decision tree (DT) and support vector machines (SVMs) using cross validation technique. In addition, with traditional statistical model, logistic regression (LR). In which we compare the model performance with ROC-AUC and accuracy, PPV and NPV. Finally, new 30 PULs were tested for external validation. Result Comparing of model performance (validation) to predict EP, LR Ranked first, followed by NNs, DT and SVMs with mean ROC-AUC ± SD of 0.904, 0.895, 0.871 and 0.862 respectively. In terms of clinical parameters, we found sensitivity ± SD performed in the same sequence as AUC (LR; 87.97% ± 7.86%, NNs; 85.97% ± 9.34%, SVM; 87.47% ± 8.23% and DT; 85.47% ± 10.37%, orderly. While specificity (± SD) was ranked by DT, then followed by SVMs, LG and NNs, which were 84.08 ± 8.56%, 83.53 ± 9.18%, 82.94 ± 9.84% and 81.76 ± 9.84%, respectively. In testing data, NNs and SVM performed equally best, then DTs and LR at ROC-AUC of 0.784, 0.778 and 0.739, respectively. Conclusion The result demonstrates that both statistics and ML model could be utilized to achieve satisfied predictions for EP. Surprisingly, the highest ranked model was LR in validating test, but in new testing data machine learning, NNs and SVMs could overcome statistics. This could shade a new light for further research of unsolved medical problem with more complexity and bigger database.