Unraveling Complex Temporal Patterns in EHRs via Robust Irregular Tensor Factorization.

Area of Science:

• Computational biology
• Medical informatics
• Machine learning

Background:

• Electronic health records (EHRs) contain rich patient data but present challenges due to irregular visit frequencies and missing entries.
• Existing tensor factorization methods like PARAFAC2 struggle with non-linear temporal dynamics and data imputation in EHRs.

Purpose of the Study:

• To introduce REPAR, a Recurrent Neural Network (RNN) Regularized Robust PARAFAC2 method, designed to model complex temporal dependencies and enhance robustness in EHR data analysis.
• To improve patient subgroup identification and clinical decision-making by addressing limitations in current EHR data processing techniques.

Main Methods:

• REPAR employs Recurrent Neural Networks (RNNs) for temporal regularization to capture complex temporal patterns.
• A low-rank constraint is integrated for enhanced robustness, particularly in handling missing data entries.
• A hybrid optimization framework is utilized to manage multiple regularizations and diverse data types within EHRs.

Main Results:

• REPAR demonstrated improved data reconstruction and robustness on three real-world EHR datasets, outperforming existing methods under missing data conditions.
• Case studies confirmed REPAR's capability in extracting meaningful dynamic phenotypes from noisy EHR data.
• The method enhanced the predictability of phenotypes derived from temporal EHRs.

Conclusions:

• REPAR offers a robust and effective approach for analyzing complex temporal dynamics in EHR data, even with significant missingness.
• The developed method facilitates more precise patient subgroup identification and supports improved clinical decision-making.
• REPAR advances the extraction and predictability of dynamic phenotypes from noisy, real-world EHR datasets.