Poster M6 - To Investigate Machine Learning Models Performance of Predicting Small-for-Gestational-Age in Malaysia and Singapore Datasets

Introduction: Small-for-gestational-age (SGA) infants refer to birth-weight below the 10th centile for gestational age, and face increased risks of morbidities such as cardiovascular disease, diabetes, and long-term growth challenges. This study aims to evaluate the performance of various machine learning models in predicting SGA using Malaysia and Singapore second-trimester (13-28 completed gestational weeks) ultrasound scan measurement datasets. The second aim is to analyse the models’ performance in predicting SGA within high-risk and low-risk cohorts of Malaysia samples, respectively.

Materials and

Methods: This study involved women with singleton pregnancies who received routine antenatal care and subsequently delivered at the University Malaya Medical Centre, Malaysia (MREC ID NO: 2021329-9997), or the National University Hospital, Singapore (approved by the National Healthcare Group (Singapore) Domain Specific Review Board - DSRB2014/01267). Only singleton pregnancies with complete datasets, including perinatal information, neonatal birth anthropometric measurements, and complete second-trimester biometry conducted between 13 and 28 completed gestational weeks, were included. This study used 4901 Malaysia retrospective samples (2018-2021: 4272 Appropriate-for-Gestational-Age (AGA), 629 SGA) to train seven models. These models leveraged seven features for prediction: abdominal circumference, head circumference, femur length, biparietal diameter, nuchal thickness, maternal age, and ultrasound-confirmed gestational age. Trained machine learning models were evaluated using the 542 Malaysia retrospective samples (2022: 482 AGA, 61 SGA), and 178 Singapore samples (2021-2022: 162 AGA, 16 SGA). Moreover, to mitigate potential multicollinearity issues in logistic regression, features with variance inflation factor (VIF) exceeding 10 were iteratively removed for assessment. Furthermore, the study explored the impact of employing stricter collinearity thresholds by retaining features with VIF < 5. In the Malaysia prospective dataset, the classification of high-risk and low-risk pregnancies followed the criteria outlined in the Royal College of Obstetricians and Gynaecologists (RCOG) Green-top Guideline No. 31. Subsequently, the dataset (2023: 464 AGA, 53 SGA) was stratified into high-risk (110 AGA, 22 SGA) and low-risk (354 AGA, 31 SGA) groups for the analysis of SGA pregnancy outcomes. The aforementioned seven models were then tested separately in high-risk and low-risk groups.

Results, Conclusions, and Discussions:

Results: In routine clinical antenatal check, the Hadlock formula serves as the established baseline standard for assessing fetal growth and determining whether the fetus is SGA. However, as indicated in Table 1, in the second trimester, using the Hadlock formula shows limited predictive capability for SGA in both Malaysia and Singapore datasets, with balanced accuracy ranging from 0.50 to 0.55. Logistic regression (VIF < 10) emerges as the top-performing model (balanced accuracy: 0.69; area under the curve (AUC): 0.75; 95% confidence interval: 0.74-0.76) in Malaysia retrospective dataset. Conversely, in the Singapore dataset, Support Vector Classifier (SVC) demonstrates superior performance (balanced accuracy: 0.78; AUC: 0.82). Furthermore, XGBoost archives the highest AUC in the high-risk group (0.66) and low-risk group (0.81).

Conclusions: Overall, the Hadlock formula demonstrates poor performance across all datasets, suggesting its limited suitability for predicting SGA in both Malaysia and Singapore populations. In clinical practice, sensitivity plays a crucial role and is often emphasized to patients during explanations. XGBoost consistently exhibits high sensitivity (0.73-0.97) across all the datasets. This signifies its effectiveness in correctly identifying a significant proportion of SGA cases, thus reducing the risk of missing such diagnoses. However, XGBoost also shows low specificity (0.46-0.49), indicating that roughly half of AGA fetuses may be incorrectly categorized as SGA.
Discussions: All models demonstrate a greater ability to detect SGA cases in the low-risk group, while have suboptimal performance in identifying SGA samples in the high-risk group. Moving forward, future research endeavors could focus on enhancing the accuracy of identifying SGA among high-risk pregnancies.

Acknowledgements (Optional): This work was supported in part by eDANA TED 1 (MOSTI002-2023TED1).