A deep learning method for preterm birth prediction using wavelet representations and auxiliary features from electrohysterogram.

Electrohysterogram (EHG) has been reported as promising way of predicting preterm birth, often relying on conventional machine learning approaches. However, the EHG modeling by deep learning has not been sufficiently investigated yet, probably due to both scarcity and complexity of the EHG data. To address these limitations, we propose CWT-AuxNet, an end-to-end deep learning method for preterm birth prediction with improved capability of learning discriminative patterns directly from raw EHG signals. In our method, continuous wavelet transform (CWT) is used to provide multiscale time-frequency representations of EHG data; an auxiliary feature termed peak amplitude (PA) is employed to further enhance discriminative capacity; and a cost-sensitive function with focal loss is designed during model training to handle severe data imbalance between term and preterm classes. Our CWT-AuxNet model is designed in a multibranch convolutional architecture enabling effective feature extraction and producing fine-grained window-level predictions. These window-level predictions are further aggregated using a user-level decision strategy to make individualized decisions regarding abnormalities and preterm risk during inference. Experimental results demonstrate that CWT-AuxNet consistently outperforms both traditional and deep learning baselines, achieving AUCs of 0.741 at the window level and 0.932 at the user level. Our study indicates the effectiveness of time-frequency deep representation learning and supports the potential of CWT-AuxNet as a promising framework for noninvasive preterm birth prediction.