Sparse Embedding for Interpretable Hospital Admission Prediction.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Data Science

Background:

• Electronic Health Records (EHR) contain vast amounts of patient data crucial for clinical decision-making.
• Predicting hospital admissions is vital for resource allocation and patient management.
• Current embedding techniques may not fully capture the complexity or interpretability of EHR data.

Purpose of the Study:

• To develop a novel sparse embedding technique for EHR data.
• To evaluate the predictive performance of sparse embeddings for hospital admission.
• To assess the interpretability and phenotype discovery capabilities of the proposed method.

Main Methods:

• Utilized a k-sparse autoencoder for dimensionality reduction of EHR data.
• Employed t-distributed Stochastic Neighbor Embedding (t-SNE) for 2D visualization of patient embeddings.
• Integrated sparse embeddings with various machine learning algorithms for predictive modeling.

Main Results:

• The sparse embedding achieved a competitive Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.878.
• Demonstrated superior phenotype discovery and interpretability compared to original data and traditional embeddings.
• Identified distinct patient phenotypes representing different risk groups.

Conclusions:

• Sparse embedding is an effective method for EHR data representation and hospital admission prediction.
• The proposed method offers enhanced interpretability, facilitating the identification of patient risk factors.
• This approach holds promise for improving clinical decision support systems and personalized medicine.