Interpreting clinical latent representations using autoencoders and probabilistic models.

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Clinical Data Analysis

Background:

• Electronic Health Records (EHRs) are rich data sources for healthcare research.
• Deep Learning (DL) models, particularly autoencoders (AEs), show promise in healthcare but often lack interpretability.
• Interpretability is crucial for clinical adoption and trust in AI-driven insights.

Purpose of the Study:

• To develop a methodology for interpreting latent representations from AE models in healthcare.
• To enhance the clinical utility of DL models by providing interpretable insights.
• To validate the proposed method using real-world patient data.

Main Methods:

• Combined probabilistic models (Gaussian mixture models) and hierarchical clustering with Kullback-Leibler divergence.
• Applied the methodology to EHR data from patients with hypertension and diabetes.
• Focused on extracting patterns from AE latent spaces.

Main Results:

• Successfully grouped patients with similar health conditions based on EHR data.
• Identified distinct patterns associated with specific diagnosis and drug codes.
• Demonstrated the potential of the approach to reveal clinical insights from AE models.

Conclusions:

• The proposed methodology offers a promising way to interpret AE models in healthcare.
• This interpretability can support clinical decision-making by highlighting patient similarities and patterns.
• Facilitates a better understanding of complex patient data for healthcare professionals.