Deep learning models and contrastive learning frameworks for preterm birth prediction and electronic health records analysis

Abstract

The rapid growth of biomedical data has provided a wealth of information for both healthcare providers and academic researchers. To achieve better outcomes for patients and improve their quality of life, an accurate esti-mation of patients’ severity and/or risk stratification is essential. As ac-curate personalized health status prediction can support clinical deci-sion-making, many researchers have spent great effort in building prediction models in medical applications. This thesis explores deep learning models for preterm birth prediction and contrastive learning frameworks for electronic health records pre-training. The first study of this thesis aims to develop a deep learning model for preterm birth prediction based on national birth data (n = 17 378 139) in the United States between 2014 and 2018. Multi-layer perceptrons (MLPs) are developed using national birth data of women who began their first prenatal visit no later than week 24 of pregnancy. 26 and 34 routinely col-lected variables that are known before the 24 weeks of pregnancy are used for nulliparous women and multiparous women, respectively. Experiments show that MLPs surpass logistic regression, random forests, and XGBoost clas-sifiers on preterm birth prediction for nulliparous women and multiparous women. Additional analysis is performed to identify important variables. The second study of this thesis explores health status prediction based on electronic health records in a semi-supervised setting. Two novel contrastive learning frameworks, contrastive predictive autoencoders abbreviated as CPAEs, are proposed to first pre-train the deep learning models without labels, and then fine-tune on the two downstream prediction tasks, i.e. in-hospital mortality prediction and length-of-stay prediction. Our frame-work is comprised of two parts: (i) a contrastive learning process, inherited from contrastive predictive coding (CPC), which aims to extract global slow-varying features, and (ii) a reconstruction process, which forces the encoder to capture local features. This work also introduces the attention mechanism in one variant of our framework to balance the above two pro-cesses. Experiments on real-world EHR data set verify the effectiveness of our proposed framework on two downstream tasks (i.e. in-hospital mortality prediction and length-of-stay prediction), compared to their supervised counterparts, the CPC model, and other baseline models. These two studies demonstrate the effectiveness to employ state-of-the-art deep learning techniques in medical prediction problems. The MLPs built for preterm birth prediction and the contrastive learning frameworks for elec-tronic health records pre-training can serve as baselines for further studies. With further improvements, such models could be incorporated into routine health information systems to allow better risk assessments and manage-ment of patients.