Ultrasound/Imaging
Poster Session 1
Yaniv Zipori, MD
Rambam Healthcare Campus
BINYAMINA, Hefa, Israel
Einat Borohovich, PhD
Rambam Health Care Campus
Haifa, Israel, Israel
Zeev Weiner, MD
Rambam Medical Health Campus
Haifa, HaZafon, Israel
Ronit Almog, PhD
Rambam Health Care Campus
Hospital, HaZafon, Israel
Dana Vitner, MD
Rambam Medical Health Campus
Toronto, ON, Canada
Ron Beloosesky, MD
Harbor UCLA Medical Center
Los Angeles, CA, United States
Adequate estimation of fetal weight (EFW) in utero is essential to a healthy pregnancy. Ultrasound-based formulas for EFW have a random error exceeding 15% in five percent of fetuses. As a result of this inaccuracy, under and over-intervention can occur. Herein, we propose a personalized AI model that leverages a linear mixed model to forecast neonatal birthweight.
Study Design:
Our 2005-2021 dataset included information on births ≥ 36 weeks of gestation to mothers with at least one consecutive birth. We excluded cases with anomalies or birthweights below 1,000 g, consecutive birthweights delta ≥ 20%, and variables with ≥ 30% missing values. The model incorporated key maternal-neonatal demographic and clinical fixed/random variables for each birth order (2 to > 5) to capture individual variability, followed by a train-test split in a 63:37 ratio. The model was then compared to the standard ultrasound EFW performance using the Mean Absolute Error (MAE).
Results:
The dataset included 11,453 mothers, accounting for 28,629 births. The key variables used in the model were ethnic and religious background, maternal age, parity, birthweights of previous deliveries, mode of conception, comorbidities, tobacco use, uterine anomalies, gestational age at delivery, and gender. Figure 1a describes the distribution of mothers according to their parity. Figure 1b gives the EFW of the testing model vs. ultrasound. Considering all birthweight ranges, the MAE of our model was comparable to ultrasound performance (232.29 g vs. 234.02 g, p=0.279). In a sub-analysis, the MAE across our model's 3000-4000 g weight range, which represented > 75% of the testing cohort, was significantly lower than the ultrasound performance (209.29 vs. 228.86, p< 0.001). At the extremes of these birthweights, ultrasound performed better than our model.
Conclusion:
The suggested model's accuracy surpasses the sonographic EFW within specific weight groups, but both methods present comparable predictive abilities when considering the total dataset. Currently, we validate our findings prospectively.