An Interpretable Deep Embedding Model for Few and Imbalanced Biomedical Data.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Bioinformatics

Background:

• Healthcare data often suffers from limited training examples and high labeling costs.
• The
• big p, small N
• problem, common in genomic studies, poses challenges for analysis.
• Skewed class distribution and the need for feature interpretability are critical in medical data analysis.

Purpose of the Study:

• To develop an interpretable deep embedding model (IDEM) for classifying medical data.
• To address challenges of limited, imbalanced training data and the
• big p, small N
• problem.
• To enhance feature importance identification for better clinical decision-making.

Main Methods:

• Implemented a feature attention layer to identify informative features.
• Utilized a feature embedding layer for handling diverse data types (numerical, categorical).
• Employed a Siamese network with contrastive loss for comparing sample embeddings.

Main Results:

• The IDEM model demonstrated superior generalization performance on synthetic and real-world medical data.
• Achieved better classification accuracy compared to conventional methods with scarce and imbalanced data.
• Successfully identified key features contributing to the classification of cases and controls.

Conclusions:

• The proposed IDEM is effective for medical data classification, especially with limited and imbalanced datasets.
• IDEM offers improved generalization and interpretability in healthcare machine learning.
• This approach aids in understanding feature contributions for clinical relevance.