Unsupervised probabilistic models for sequential Electronic Health Records.

Area of Science:

• Computational biology
• Health informatics
• Machine learning

Background:

• Electronic Health Records (EHR) contain complex, heterogeneous data.
• Existing models struggle to capture the dynamics and structure within long-term patient data.
• Unsupervised methods are needed to uncover hidden patterns in large-scale health datasets.

Purpose of the Study:

• To develop an unsupervised probabilistic model for heterogeneous EHR data.
• To model sequences of varying lengths, such as medications and lab results.
• To enable subgrouping and incorporate data dynamics for deeper insights.

Main Methods:

• Utilized a mixture model formulation for direct sequence modeling.
• Employed a layered latent variable structure to encode data relationships.
• Trained the model on episodic EHR data from a large healthcare system.

Main Results:

• The model successfully captures underlying structure in heterogeneous EHR data.
• Identified distinct subject subgroups and sequence dynamics.
• Demonstrated the model's utility in assessing mortality likelihood from patient data.

Conclusions:

• The developed unsupervised model provides novel insights into complex EHR data.
• It effectively handles heterogeneous data types and variable-length sequences.
• The approach has potential applications in clinical subgroup discovery and risk prediction.