(184) Improving clinical fetal weight estimation using machine learning

This retrospective cohort study included all deliveries between 02/2014 and 09/2020. Multiple births, stillbirths, deliveries < 24 weeks and records without valid cEFWs were excluded (N=10,259). The analyzed cohort consisted of 38,615 deliveries (Table 1), with 20% of the data (from 05/2019 to 09/2020, N=7,723) used as a holdout test set, and the remaining 80% used to train and validate a linear regression model for estimating birthweight (BW). The model consisted of 25 variables, including gestational age (GA), gravidity, parity, previous abortions and cesarean sections, maternal height, weight, pregnancy weight gain, hypertensive disorders, gestational diabetes, placental location, fetal presentation, fetal sex and cEFW. The accuracies of the prediction model and of the cEFW were compared, using the actual BW as ground truth. The accuracy of sonographic weight estimation (sEFW) was also compared for a subset of patients. The statistical difference between estimators was assessed by McNemar’s test and paired t-test.

Results:

The prediction model improved both the absolute and relative accuracy of the cEFW, increasing the fraction of accurate estimations (within 10% of the actual BW) from 70.9% to 74.8% on the test set (P < 0.01). On the subset of deliveries with an available sEFW (N=1,707), the fraction of correct estimations by the model was 77.9%, compared to 75.8% by cEFW (P < 0.05) and 75.4% by sEFW (P < 0.05). Similar results were obtained by cross-validation on the training and validation set (Table 2). The major contributing variables of the model were gravidity, parity, cEFW, GA and fetal sex.

Conclusion:

It is feasible to combine a machine learning algorithm with clinical assessment to estimate BW with improved accuracy compared to the current practice.