gms | German Medical Science

Introduction: Induction of labour (IOL) is worldwide a common medical intervention to end pregnancy and to prevent adverse events in mothers and infants, especially after the estimated delivery date (> 40+0 gestational weeks (GW)). Although the results from both observational studies and randomized clinical trials (RCTs) regarding the preferred timing of IOL are conflicting [1], [2], the general recommendation is to induce women around 41+0 GW and not to await a later onset of labour (“expectant management” (EM)) [3]. To provide more individualized advice for parents-to-be and medical staff, the present study used data-driven models for the prediction of caesarean delivery (CD) in a population of late- and post-term pregnancies [4] (≥ 41+0 GW).

Methods: Low-risk primiparous and parous women at or beyond 41 GW were identified by applying inclusion and exclusion criteria to the nationwide Swedish Medical Birth Register (MBR) [5] from 1998 to 2019. To assess possible changes with increasing gestational length, CD was predicted in four different subgroups (SG1: IOL 41+0-41+1 vs. EM > 41+1; SG2: IOL 41+2-41+3 vs. EM > 41+3; SG3: IOL 41+4-41+5 vs. EM > 41+5; SG4: IOL 41+6-42+0 vs. EM > 42+0). IOL itself was used as a binary feature in each of the groups. Sixty-four diagnoses and clinical characteristics were used as independent variables (“features”). To ensure predictive design, these predictors had to be known or existing at the time for decision-making regarding IOL in each subgroup. Diagnoses which occurred later in pregnancy or during labour and birth characteristics were not considered (e.g. duration of labour).

Results: The study population included 400,140 pregnancies. Restricting the initial analyses to primiparous women, the four subgroups comprised 178,932 (SG1), 129,449 (SG2), 90,448 (SG3) and 61,301 (SG4) eligible pregnancies with a caesarean section rate between 19.8% (SG1) and 26.5% (SG4) (“imbalanced data set”). In a complete case analysis four different machine learning algorithms (Random Forest, Mixed NaÏve Bayes model, Support Vector Machine and Neural Network) and a logistic regression model were compared. The Random Forest and NaÏve Bayes models had the best performance but predicted CD with a low sensitivity (“recall”) of 15%-21% in all four subgroups. In the Random Forest feature importance, IOL was not among the highly contributing features.

Conclusion: The data-driven approach is novel in its application in prolonged pregnancies and provides new methodological insights into the exact timing of IOL in late-term pregnancies. However, caesarean delivery could not be predicted with high certainty using the available features from the MBR. To enhance predictive power, further variables derived from the Swedish Pregnancy Register and the Stockholm-Gotland Perinatal Cohort will be added to the compiled feature set. This will mainly include data collected during pregnancy e.g. weight gain, antenatal ultrasound measurements or Bishop Score prior to delivery. Based on models with higher performance, the prediction of additional maternal (e.g. vaginal operative delivery) and fetal outcomes (e.g. stillbirth, neonatal death, low Apgar Score) with the entire study population (primiparous and parous women) aims to build a clinical decision-making tool supporting an informed choice.

Funding: Stefanie Schmauder is on a Walter Benjamin Fellowship funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – Project number 504494065.

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.