Automated Fetal Health Classification Using Auto-Sklearn Approach for Enhanced Clinical Sensitivity

Abstract

Accurate and timely assessment of fetal health through Cardiotocography (CTG) is critical for making less neonatal morbidity and mortality. Achieving high diagnostic accuracy with classic machine learning models often requires extensive manual hyperparameter tuning and feature engineering. This study proposes an automated approach using the Auto-Sklearn framework to classify fetal health into three categories: Normal, Suspect, and Pathological. Utilizing a dataset of 2,126 clinical instances, we implemented a "cold-start" Bayesian Optimization strategy intentionally bypassing meta learning to evaluate framework's raw optimization capacity. The results demonstrate that automated pipeline achieved an overall Accuracy of 96.94% and a weighted F1-score of 88.37%. Crucially, from a clinical safety perspective model obtained a Recall of 97.00% for Pathological class and a remarkably low Miss Rate of 0.0656, ensuring high sensitivity in detecting fetal distress. With a Specificity of 97.43%, and the model effectively minimizes false alarms, proving that AutoML can match or exceed performance of manually tuned systems. This research provides a robust and highly efficient methodology for developing clinical decision support tools in obstetrics.