Multi-source Learning via Completion of Block-wise Overlapping Noisy Matrices

Area of Science:

• Computational biology
• Data science
• Medical informatics

Background:

• Electronic healthcare records (EHR) offer valuable data for research.
• Representing EHR features from diverse sources, including unstructured narratives and structured data, is challenging.
• Existing matrix factorization methods struggle with multi-source data containing overlapping yet non-identical features.

Purpose of the Study:

• To propose a novel word embedding generative model for multi-source EHR data.
• To develop an efficient algorithm for optimal aggregation of multi-source pointwise mutual information matrices.
• To address limitations in current matrix completion techniques by considering non-independent missing data mechanisms.

Main Methods:

• Proposed a novel word embedding generative model for multi-source EHR data.
• Designed an efficient Block-wise Overlapping Noisy Matrix Integration (BONMI) algorithm.
• Analyzed the theoretical guarantees and statistical recovery rates of the proposed estimator for matrix completion.

Main Results:

• BONMI optimally aggregates multi-source pointwise mutual information matrices with theoretical guarantees.
• The proposed method demonstrates effectiveness in matrix completion without assuming independent missingness.
• Simulation studies confirm BONMI's robust performance across various configurations.

Conclusions:

• BONMI successfully integrates multi-lingual, multi-source medical text and EHR data.
• The method enables effective co-training of semantic embeddings and translation of medical concepts between English and Chinese.
• BONMI offers advantages over existing methods for multi-source EHR data representation and integration.