Creation of an artificial intelligence model for calculating gestational age and predicting the risk of birth of a small fetus based on dynamic analysis of ultrasound fetometry

Low birth weight remains one of the key causes of perinatal morbidity and mortality. Timely detection of intrauterine growth disorders requires tools capable of integrating multidimensional ultrasound and clinical data of a pregnant woman.

The purpose of the study was to develop and verify a machine learning model that predicts the risk of having an underweight fetus based on fetometric measurements and maternal factors, as well as to evaluate its diagnostic value compared to classical methods.

Material and methods. The study included 5,477 pregnant women (8,396 ultrasounds; 11–40 weeks gestation). The control group consisted of 5,161 women who gave birth to full–term newborns with normal weight, the case group consisted of 316 pregnant women with full-term babies weighing less than 10th percentile for the corresponding period of pregnancy. For each ultrasound study, standard fetometric parameters and 20 clinical/social variables of the mother were collected. After data purification, a stratified division by pregnant women (80 % train / 20% test), standardization of quantitative characteristics, and binary coding of categories were performed. The following methods were compared: gradient boosting (XGBoost, CatBoost, LightGBM), transformer network and multitasking (regression + classification) neural network. Hyperparameters were selected by Optuna; the quality was evaluated using mean squared error (MAE), root mean squared error (RMSE), area under curve (AUC), sensitivity (Se), specificity (Sp).

Results and discussion. Regression (assessment of gestational age): stacking three boosts gave MAE 0.29 weeks (≈ 2 days), RMSE 0.40 weeks, R2 = 0.989. The multitasking network reached MAE 0.32 weeks. Classification (LBW / norm): The multitasking model showed an AUC of 0.96, Se of 90 %, and Sp of 96 % at the optimal threshold. The greatest contribution to the prognosis was made by the circumference of the abdomen and the length of the femur of the fetus, and maternal factors included placental insufficiency, hypertensive complications, smoking and parity. The elimination of parity reduced AUC by ≈ 0.02, confirming its additional informative value. Calibration of probabilities after isotonic regression demonstrated proximity to the ideal line, which ensures interpretability of risk for the clinician.

Conclusions. A highly accurate system for predicting the risk of having a small fetus has been created and validated, combining ultrasound fetometry and clinical and social data. The accuracy (AUC 0.96) and high sensitivity make the model a promising screening tool for obstetric practice. The implementation of the algorithm in the form of an automated report can improve the early diagnosis of fetal growth disorders and optimize the routing of pregnant women at risk. The next steps are external validation based on multicenter data and analysis of clinical and economic efficiency.

1. Lawn J.E., Ohuma E.O., Bradley E., Idueta L.S., Hazel E., Okwaraji Y.B., Erchick D.J., Yargawa J., Katz J., Lee A.C.C., … Subnational Vulnerable Newborn Measurement Group. Small babies, big risks: global estimates of prevalence and mortality for vulnerable newborns to accelerate change and improve counting. Lancet. 2023;401(10389):1707–1719. doi: 10.1016/S0140-6736(23)00522-6

2. Blencowe H., Krasevec J., de Onis M., Black R.E., An X., Stevens G.A., Borghi E., Hayashi C., Estevez D., Cegolon L., … Cousens S. National, regional, and worldwide estimates of low birthweight in 2015, with trends from 2000: a systematic analysis. Lancet Glob. Health. 2019;7(7):e849–e860. doi: 10.1016/s2214-109X(18)30565-5

3. Bekele W.T. Machine learning algorithms for predicting low birth weight in Ethiopia. BMC Med. Inform. Decis. Mak. 2022;22(1):232. doi: 10.1186/s12911-022-01981-9

4. Islam Pollob S.M.A., Abedin M.M., Islam M.T., Islam M.M., Maniruzzaman M. Predicting risks of low birth weight in Bangladesh with machine learning. PLoS One. 2022;17(5):e0267190. doi: 10.1371/journal.pone.0267190

5. Khan W., Zaki N., Masud M.M., Ahmad A., Ali L., Ali N., Ahmed L.A. Infant birth weight estimation and low birth weight classification in United Arab Emirates using machine learning algorithms. Sci. Rep. 2022;12(1):12110. doi: 10.1038/s41598-022-14393-6

6. Punyapet P., Suwanrath C., Chainarong N., Sawaddisan R., Vichitkunakorn P. Predictors of adverse perinatal outcomes in fetal growth restriction using a combination of maternal clinical factors and simple ultrasound parameters. Int. J. Gynaecol. Obstet. 2023;162(2):669–675. doi: 10.1002/ijgo.14721

7. Patterson J.K., Thorsten V.R., Eggleston B., Nolen T., Lokangaka A., Tshefu A., Goudar S.S., Derman R.J., Chomba E., Carlo W.A., … Bauserman M. Building a predictive model of low birth weight in lowand middle-income countries: a prospective cohort study. BMC Pregnancy Childbirth. 2023;23(1):600. doi: 10.1186/s12884-023-05866-1

8. Sanchez-Martinez S., Marti-Castellote P.M., Hoodbhoy Z., Bernardino G., Prats-Valero J., Aguado A.M., Testa L., Piella G., Crovetto F., Snyder C., … Bijnens B. Prediction of low birth weight from fetal ultrasound and clinical characteristics: a comparative study between a lowand middle-income and a high-income country. BMJ Glob. Health. 2024;9(12):e016088. doi: 10.1136/bmjgh-2024-016088

9. Dülger Ö., Dursun A., Osmanoğlu U. Fetal birth weight estimation with machine-learning techniques in 15-40 weeks of pregnancy. Van. Med. J. 2024;31(3):186–191.

10. Ranjbar A., Montazeri F., Farashah M.V., Mehrnoush V., Darsareh F., Roozbeh N. Machine learning-based approach for predicting low birth weight. BMC Pregnancy Childbirth. 2023;23(1):803. doi: 10.1186/s12884-023-06128-w