Poster L12 - One-Size-Fits-All Prediction Model for Small-for-Gestational Age (SGA) Babies: Unraveling Demographic Disparities and Model Generalization Challenges.

Introduction: Small for gestational age (SGA) is a phenomenon where the birth weight of an infant is less than the 10th percentile for gestational age. This condition can arise due to various factors, including maternal health, placental function, genetic factors, and environmental influences. The prevalence of SGA in high-income countries such as the United States and Australia is approximately 11%, however in resource-poor countries with the highest burden in South Asia, where up to 34% of infants were born SGA. The significance of predicting SGA babies in advance lies in the potential health risks they face such as hypoglycemia, hypothermia and long-term developmental issues. Early identification of SGA babies allows timely interventions to mitigate these risks and improve outcomes.
However, predicting SGA is complex due to the diverse factors involved, and each demographic may have unique contributing factors. Furthermore, variations in measurement calibration across hospitals add complexity. The current method, utilising a mix of fetal biometric measurements to estimate fetal weight and determine its percentile based on gestational age, often lacks accuracy. Although AI research has shown promising results in this domain, there is no one-size-fits-all SGA prediction method capable of accurately predicting SGA across all demographics.

Thus, our goal is to explore the demographic variations and model generalization issues that impede the creation of a reliable, one-size-fits-all predictive model for identifying fetuses likely to be born SGA during the third trimester (weeks 28-40) of pregnancy.

Materials and

Methods: The study protocol was approved by the University of Malaya Medical Center, Medical Research Ethical Committee (MREC) with MECID.No: 2021329-9997 and AMMA Center for Diagnosis and Preventive Medicine Pvt Ltd, Kochi, Kerala, India with EC.No: 05/2023. All the data involved originated from Pusat Perubatan Universiti Malaya (PPUM), a government-funded medical institution in Kuala Lumpur, Malaysia and AMMA Center for Diagnosis and Preventive Medicine Pvt Ltd, Kochi, Kerala, India.

The dataset comprises ultrasound measurements collected from hospitals across Malaysia and India. Initially, there were 141,500 antenatal and 19,812 postnatal ultrasound measurements from Malaysia and 1904 antenatal measurements from India. These datasets contained different numbers of features. Following the data preprocessing in Figure 1, the Malaysian dataset was refined to include 27,230 antenatal ultrasound measurements from 2018 to 2021, with 11 features. The Indian dataset was reduced to 1,642 antenatal measurements from 2022 to 2023, retaining 20 features. After identifying shared knowledge-driven features from the Malaysia and India datasets, eight features were selected for analysis which include gestational age, biparietal diameter, head circumference, abdominal circumference, femur length, estimated fetal weight, maternal age and fetus’ gender.

The neural network and machine learning algorithms are trained on the Malaysian dataset and tested on datasets from India. The inner workings and weights of the neural network are examined and interpreted based on Shapley values (Figure 2). Model performance is assessed using five metrics: Balanced Accuracy, Receiver Operating Characteristics Area Under Curve (ROC AUC), F1 Score, Precision, and Recall.

Results, Conclusions, and Discussions: Results & Discussions
Using the eight common features from both the Malaysia and India datasets, the neural network and machine learning algorithms exhibit exceptional performance, achieving the best ROC AUC score of 85.30% when tested on the Malaysian demographic dataset (Figure 3a). However, when tested on the India dataset, its performance is notably reduced, with the highest ROC AUC score reaching only 68.29% (Figure 3b). Furthermore, other evaluation metrics also show suboptimal performance in the India dataset. When the model was trained and tested on the India dataset, it achieved a better ROC AUC score of 70.47% (Figure 3c). While transfer learning is explored, the neural network performance improved as shown in Figure 3d. Its effectiveness diminishes due to the model's limited capacity to learn from a small dataset. Nonetheless, using a large dataset for the base model in transfer learning yields slightly better performance compared to training and testing directly on the small India dataset (Figure 3c and 3d).

One of the main reasons is the disparity in data distribution. We illustrated the variance in data distribution between Malaysia and India datasets (Figure 4). These distinctions can influence feature learning differently, as certain features may hold more predictive value in one country due to unique population or environmental characteristics. Additionally, the lack of standardized calibration and variations in data collection methods pose challenges for the model to generalize effectively across diverse demographics. Given the restricted dataset size and feature set, the model struggles to discern underlying patterns for each feature accurately, risking overfitting specific dataset characteristics.

Conclusion
Our extensive experiments reveal that no single neural network model or existing machine learning algorithm can accurately predict SGA across all demographics. Yet, we remain optimistic that overcoming these limitations is possible. With enhanced data accuracy and expanded feature sets, we can refine model architectures to capture the true underlying patterns. This opens the door to more precise SGA predictions early in pregnancy.

Acknowledgements (Optional):