Sparse Multidimensional Patient Modeling using Auxiliary Confidence Labels.

Area of Science:

• Machine Learning
• Medical Informatics
• Clinical Decision Support

Background:

• Learning classification models for patient Electronic Health Records (EHRs) typically requires extensive expert supervision.
• Reducing the cost of expert labeling is crucial for developing practical EHR inference models.
• Interpretability is essential for clinical models, allowing clinicians to understand decision-making processes.

Purpose of the Study:

• To develop a novel metric learning method that incorporates confidence labels to reduce the need for extensive expert supervision.
• To enhance the interpretability of EHR classification models for clinical applications.
• To create a method that produces accurate, sparse, and interpretable models.

Main Methods:

• Developed Confidence bAsed MEtric Learning (CAMEL), a novel metric learning approach.
• Incorporated confidence labels from experts to guide model training.
• Designed CAMEL to induce sparsity, generate confidence scores, and create multidimensional interpretable feature spaces.

Main Results:

• CAMEL achieved comparable or superior accuracy to existing methods using the same amount of supervision.
• When utilizing confidence scores, CAMEL learned accurate models using only 10% of the training data.
• Qualitative assessments confirmed that CAMEL's learned metrics identify and articulate key clinical factors.

Conclusions:

• CAMEL effectively reduces the cost of expert supervision in EHR model learning by leveraging confidence labels.
• The method enhances model interpretability, providing clinicians with actionable insights.
• CAMEL demonstrates a promising approach for developing accurate and efficient clinical decision support tools.