Abstract
Detecting low birth weight is crucial for early identification of at-risk pregnancies which are associated with significant neonatal and maternal morbidity and mortality risks. This study presents an efficient and interpretable framework for unsupervised detection of low, very low, and extreme birth weights. While traditional approaches to managing class imbalance require labeled data, our study explores the use of unsupervised learning to detect anomalies indicative of low birth weight scenarios. This method is particularly valuable in contexts where labeled data are scarce or labels for the anomaly class are not available, allowing for preliminary insights and detection that can inform further data labeling and more focused supervised learning efforts. We employed fourteen different anomaly detection algorithms and evaluated their performance using Area Under the Receiver Operating Characteristics (AUCROC) and Area Under the Precision-Recall Curve (AUCPR) metrics. Our experiments demonstrated that One Class Support Vector Machine (OCSVM) and Empirical-Cumulative-distribution-based Outlier Detection (ECOD) effectively identified anomalies across different birth weight categories. The OCSVM attained an AUCROC of 0.72 and an AUCPR of 0.0253 for extreme LBW detection, while the ECOD model showed competitive performance with an AUCPR of 0.045 for very low LBW cases. Additionally, a novel feature perturbation technique was introduced to enhance the interpretability of the anomaly detection models by providing insights into the relative importance of various prenatal features. The proposed interpretation methodology is validated by the clinician experts and reveals promise for early intervention strategies and improved neonatal care.