(164) Individualized Preterm Labor Risk Prediction Via Neural Network Ensembles

Prediction and prevention of preterm labor (PTL) is an ongoing challenge in the field of obstetrics, with the resultant complications of fetal prematurity remaining a leading cause of neonatal mortality nationally and worldwide despite modern interventions. We introduce a computational method for early risk prediction of development of PTL via ensembles of artificial neural networks with the ultimate goal of deploying a clinical tool to provide individualized and accurate running risk assessment to allow for early detection and prevention of PTL.

Study Design:

This method includes clinical data from 228438 pregnancies obtained from the Consortium on Safe Labor (CSL) study. We randomly isolated 10000 pregnancies for testing, including 482 preterm labor events (≤ 34 weeks-gestation) , which were tracked individually throughout the study along with 482 randomly-chosen non-events. We then trained network ensembles using the remaining ~95% of pregnancies. Four different ensembles were trained, each comprised of 300 artificial neural networks. These ensembles created a running risk index, which correlates to overall risk of PTL.

Results:

The fourth ensemble correctly predicted 75% of preterm labor events (363 out of 482) and 86% of non-events (415 out of 482). Risk indices were noted to increase from the baseline risk of 4-5% through the third trimester for pregnancies ultimately positive for PTL and negative pregnancies (blue), indicating that the method typically reveals risk well in advance of the event (Figure 1). The vast majority of pregnancies ultimately positive for PTL had predicted risk indices above 95% and the vast majority of ultimately negative pregnancies have risk indices below 5% (Figure 2).

Conclusion:

This novel computational method for prediction and prevention of PTL represents a feasible and effective tool of obstetric management along with improved diagnostic specifications in terms of accessibility, equity, cost-effectiveness and sustainability.