Deep learning predicts extreme preterm birth from electronic health records

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Perinatal Medicine

Background:

• Extreme preterm birth (EPB), occurring before 28 weeks, causes most newborn deaths.
• Existing prediction models often focus on later preterm births.
• Predicting EPB is crucial for improving neonatal outcomes.

Purpose of the Study:

• To investigate the efficacy of deep learning models in predicting EPB.
• To leverage temporal patterns in electronic health records (EHRs) for EPB prediction.
• To compare deep learning performance against traditional machine learning models.

Main Methods:

• Recurrent neural networks (RNNs) were employed to analyze EHR data, incorporating word embedding for temporal relations.
• Ensemble methods combined multiple RNN models for enhanced EPB prediction.
• Models were trained and evaluated using a large dataset of deliveries, with undersampling techniques applied due to EPB's low prevalence.

Main Results:

• RNN ensemble models trained on balanced data achieved superior performance (AUC 0.827, sensitivity 0.965) compared to models trained on imbalanced data and traditional machine learning methods.
• The RNN ensemble models outperformed the best baseline models (AUC 0.777, sensitivity 0.819).
• Identified risk factors for EPB included twin pregnancy, short cervical length, hypertensive disorders, systemic lupus erythematosus, and hydroxychloroquine sulfate.

Conclusions:

• Temporal deep learning models demonstrate significant potential for predicting EPB.
• Accurate EPB prediction can be achieved up to 8 weeks prior to birth.
• Early prediction enables effective healthcare resource allocation and timely patient care for extreme preterm births.