Fetal health classification: a deep learning model with enhanced interpretability and lightweight deployment

Fetal health classification is crucial in the detection of potential pregnancy complications at an early level to enable timely medical intervention. Traditional diagnostic techniques rely on expert interpretation of cardiotocography (CTG) recordings, which is time-consuming and subjective in nature. To address this drawback, we suggest an optimized deep learning approach for fetal health classification based on a feedforward neural network (FNN) with hyperparameter optimization. Our methodology involves exhaustive data preprocessing, feature engineering, and hyperparameter optimization through Bayesian search, tuned with the best model consisting of 96 and 256 neurons in the first and second hidden layers, respectively, L2 regularization coefficients 0.000578 and 1.81e-05, dropout rates 0.213 and 0.458, and learning rate 0.001704. The suggested model is compared to traditional machine learning classifiers, i.e., Random Forest, XGBoost, LightGBM, and Gradient Boosting, using performance measures such as accuracy, precision, recall, F1-score, and AUC-ROC. Experimental findings demonstrate that the optimized FNN outperforms traditional models with better classification performance. Our findings reveal the potential of deep learning in fetal health assessment and its clinical usefulness in real-world environments.