SCEHR: Supervised Contrastive Learning for Clinical Risk Prediction using Electronic Health Records.

Area of Science:

• Machine Learning in Healthcare
• Clinical Informatics
• Biomedical Data Science

Background:

• Contrastive learning excels in image and text domains, offering self-supervised and supervised approaches.
• Clinical risk prediction using longitudinal electronic health records (EHR) is crucial for patient care.
• Existing methods often struggle with the inherent data imbalance in clinical datasets.

Purpose of the Study:

• To extend supervised contrastive learning to clinical risk prediction tasks utilizing EHR data.
• To propose a unified framework for both binary and multi-label classification in clinical prediction.
• To develop novel supervised contrastive loss functions adaptable to diverse clinical prediction scenarios.

Main Methods:

• Introduced a general supervised contrastive loss function ($$) for clinical risk prediction.
• The loss function comprises two components: contrasting samples with learned anchors and with supervised labels.
• Implemented and evaluated two versions of the proposed loss on real-world EHR datasets.

Main Results:

• The proposed loss functions significantly improved the performance of baseline and state-of-the-art models.
• Demonstrated effectiveness on benchmarking tasks for clinical risk predictions, including mortality prediction and phenotyping.
• Showed robust performance even with highly imbalanced clinical data, a common challenge in EHR studies.

Conclusions:

• The novel supervised contrastive loss functions offer a powerful alternative to traditional cross-entropy losses.
• This framework provides a unified approach for various clinical prediction tasks (binary and multi-label).
• The method is effective for improving predictive accuracy and handling data imbalance in longitudinal EHR data.