Characterizing Patient Representations for Computational Phenotyping.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Computational Biology

Background:

• Patient representation learning methods offer rich data representations for advancing computational phenotypes (CP).
• Current methods are limited by using predefined concept sets or all patient data, hindering novel discovery and explainability.
• Developing data-driven approaches is crucial for optimizing CP development.

Purpose of the Study:

• To extensively characterize the utility of patient representation learning for computational phenotype development.
• To investigate the impact of different data types and sampling windows on rare disease classification using patient representations.
• To provide a data-driven foundation for patient representation-based CP development pipelines.

Main Methods:

• Conducted ablation studies on patient representation learning methods.
• Varied combinations of data types and sampling windows were used to build patient representations.
• Evaluated the impact of these representations on rare disease classification and clustering performance.

Main Results:

• Patient representation learning performance is directly impacted by the choice of data type and sampling window.
• Classification and clustering performance varied significantly across different rare disease groups.
• Preliminary results underscore the importance of data-driven characterization in CP development.

Conclusions:

• The selection of data types and sampling windows is critical for effective patient representation learning in CP development.
• A data-driven approach is essential for optimizing the performance and explainability of computational phenotypes.
• Further research is needed to refine these methods for diverse rare diseases.