An efficient interpretable framework for unsupervised low, very low and extreme birth weight detection.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Neonatal Health

Background:

• Low birth weight (LBW) is a critical indicator of neonatal and maternal health risks.
• Accurate LBW detection is vital for timely intervention in high-risk pregnancies.
• Traditional methods often rely on labeled data, which can be scarce for anomaly detection.

Purpose of the Study:

• To develop an efficient and interpretable unsupervised learning framework for detecting low, very low, and extreme birth weights.
• To address the challenge of class imbalance in LBW detection without relying on labeled data.
• To enhance the interpretability of anomaly detection models for clinical application.

Main Methods:

• Evaluation of fourteen distinct anomaly detection algorithms.
• Performance assessment using Area Under the Receiver Operating Characteristics (AUCROC) and Area Under the Precision-Recall Curve (AUCPR) metrics.
• Introduction of a novel feature perturbation technique for model interpretability.

Main Results:

• One Class Support Vector Machine (OCSVM) and Empirical-Cumulative-distribution-based Outlier Detection (ECOD) demonstrated effectiveness in identifying LBW anomalies.
• OCSVM achieved an AUCROC of 0.72 and AUCPR of 0.0253 for extreme LBW.
• ECOD showed a competitive AUCPR of 0.045 for very low LBW detection.

Conclusions:

• Unsupervised anomaly detection offers a viable approach for LBW identification, especially with limited labeled data.
• OCSVM and ECOD are promising algorithms for detecting various categories of LBW.
• The developed interpretation methodology, validated by clinicians, supports early intervention and improved neonatal outcomes.