MULTIPAR: Supervised Irregular Tensor Factorization with Multi-task Learning for Computational Phenotyping.

Area of Science:

• Computational biology
• Data science
• Medical informatics

Background:

• Tensor factorization methods like PARAFAC2 are used for Electronic Health Records (EHR) mining, but struggle with irregular data and limited predictability.
• Existing models for EHR analysis lack satisfactory interpretability and predictive power, hindering their clinical application.

Purpose of the Study:

• To introduce MULTIPAR, a supervised irregular tensor factorization technique incorporating multi-task learning for improved computational phenotyping.
• To enhance the extraction of meaningful medical concepts (phenotypes) and boost predictive performance in EHR data analysis.

Main Methods:

• Developed MULTIPAR, a supervised irregular tensor factorization model utilizing multi-task learning.
• Applied MULTIPAR to temporal EHR datasets, incorporating both static and dynamic prediction tasks.
• Evaluated model scalability, tensor fit, phenotype interpretability, and predictive performance.

Main Results:

• MULTIPAR demonstrates scalability on real-world temporal EHR datasets.
• The proposed method achieves superior tensor fit and identifies more meaningful patient subgroups.
• MULTIPAR significantly improves predictive performance for downstream clinical tasks compared to state-of-the-art methods.

Conclusions:

• MULTIPAR offers a powerful approach for computational phenotyping by integrating supervised tensor factorization with multi-task learning.
• The method enhances both the interpretability of extracted phenotypes and the accuracy of predictive modeling using EHR data.
• MULTIPAR represents a significant advancement for EHR data mining and clinical decision support.