Inference of dependency knowledge graph for Electronic Health Records.

Area of Science:

• Computational biology
• Health informatics
• Statistical modeling

Background:

• Electronic Health Records (EHRs) offer rich data for healthcare research but present analysis challenges.
• Knowledge Graphs (KGs) can improve predictive modeling and feature selection in EHR analysis.
• Existing KG construction methods lack statistical certainty, especially with privacy-limited EHR data.

Purpose of the Study:

• To propose the first inferential framework for constructing sparse KGs with statistical guarantees from EHR data.
• To address limitations in current KG construction techniques regarding statistical certainty and data privacy.
• To enable reliable inference on non-linear statistics within low-rank temporal dependent models.

Main Methods:

• Developed a dynamic log-linear topic model for KG construction.
• Estimated KG embeddings via singular value decomposition of the empirical pointwise mutual information matrix.
• Established entrywise asymptotic normality for the low-rank KG estimator to ensure sparse edge recovery with controlled Type I error.

Main Results:

• The proposed framework provides statistical guarantees for KG link inference.
• The method demonstrates scalability and accuracy in recovering sparse graph structures.
• Validated through simulations and application to real-world EHR data for clinical KG construction and feature embedding.

Conclusions:

• The novel inferential framework enables statistically sound KG construction from EHR data.
• This approach enhances the reliability of clinical KGs and feature embeddings for healthcare research.
• Addresses a critical gap in statistical inference for temporal dependent models with limited data.