SCEHR

SCEHR applies supervised contrastive learning to electronic health records to improve clinical risk prediction for binary and multi-label outcomes.

Key Features:

• Supervised Contrastive Loss Framework: A combined loss \mathcal{L}_{\text{Contrastive Cross Entropy}} + \lambda \mathcal{L}_{\text{Supervised Contrastive Regularizer}} that pulls similar samples closer and pushes dissimilar samples apart, comprising Contrastive Cross Entropy (contrasts samples with learned anchors representing positive and negative clusters) and a Supervised Contrastive Regularizer (contrasts samples based on supervised labels).
• Unified Framework for Diverse Predictions: Supports binary classification tasks such as in-hospital mortality prediction and multi-label classification tasks such as phenotyping using EHR data.
• Performance Enhancement: Empirically validated on real-world EHR datasets with reported improvements over strong baseline and state-of-the-art clinical risk prediction approaches.
• Handling Imbalanced Data: Demonstrates robustness to extremely imbalanced class distributions common in clinical risk prediction.
• Replaceable Loss Functions: Proposed loss functions can substitute standard binary or multi-label cross-entropy losses in existing predictive models.

Scientific Applications:

• Mortality and readmission prediction: Predicting patient outcomes such as in-hospital mortality and readmission rates from longitudinal EHR data.
• Phenotyping: Multi-label phenotyping to identify phenotypic patterns that indicate underlying health conditions.
• Personalized medicine: Improving risk stratification and individualized predictions by leveraging detailed longitudinal EHR information.

Methodology:

Applies supervised contrastive learning adapted for longitudinal EHR data using a combined loss (\mathcal{L}_{\text{Contrastive Cross Entropy}} + \lambda \mathcal{L}_{\text{Supervised Contrastive Regularizer}), leveraging learned anchors and label-based contrastive regularization to capture intra-class similarity and inter-class dissimilarity for binary and multi-label classification.