(825) Using supervised machine learning to predict spontaneous preterm birth in patients presenting with preterm contractions

Threatened preterm labor with the occurrence of preterm contractions is the most common reason for antepartum hospital admission, encompassing between 44 – 59% of all antenatal hospitalizations. Only half of the women who present with TPTL ultimately experience spontaneous preterm birth (sPTB). The nature of preterm contractions and sPTB relationship is not fully understood. The objective of this study was to understand what contributes to sPTB when patients present with preterm contraction symptoms using machine learning (ML).

Study Design:

A retrospective case-control study of individuals who experienced preterm contractions prior to sPTBs between January 2022 to March 2024 in a single tertiary medical center. A total of 3,245 patients were manually screened. Two hundred twenty-four patients who presented with preterm contractions < 36 weeks were included, with 42 delivering > 37 weeks and 182 delivering < 37 weeks. Three supervised ML methods (Random Forest, Support Vector Machine, and k-Nearest Neighbors) were compared to predict sPTB based on key clinical features (twins versus singleton, gestational age at presentation, body mass index - BMI, smoking history, concurrent PPROM status, history of PTB, and cervical dilation). Missing information in features were imputed from median.

Results: All three ML methods achieved promising AUROC in sPTB prediction (Table 1). The top four features related to the sPTB predictions are cervical dilation (measured in centimeters), BMI, gestational age at presentation, and concurrent PPROM status (Table 2).

Conclusion: Spontaneous PTB can be predicted via clinical information readily available with ML methods. The key clinical features used in the prediction are similar across all three ML methods used. Single center data and rarity of cases were main limitations. A multicenter or state database should be used to further validate findings.