SparGE: Sparse Coding-based Patient Similarity Learning via Low-rank Constraints and Graph Embedding

Xian Wei, See-Kiong Ng, Tongtong Zhang, Yingjie Liu

TL;DR

This paper introduces SparGE, a novel framework for learning patient similarity from Electronic Health Records (EHRs). By combining low-rank constrained sparse coding and graph embedding, SparGE effectively handles the inherent challenges of EHR data, such as missing values and noise, to significantly improve patient similarity assessment.

What Problem Does This Paper Address?

Personalized medicine relies heavily on understanding how similar patients are to one another, a task known as Patient Similarity Assessment (PSA). This is crucial for making informed clinical decisions. However, Electronic Health Records (EHRs), the primary data source for PSA, are notoriously messy. They suffer from:

  • Missing Values: Patients have different sets of tests and conditions, leading to many gaps in their records.
  • Noise: Data, especially patient-reported information, can be inaccurate.
  • High Dimensionality & Sparsity: EHRs contain a vast number of features, often requiring large datasets to train machine learning models effectively.

Existing machine learning approaches struggle with these data deficiencies, leading to less accurate patient similarity measures. This paper aims to bridge this gap by proposing a robust framework that can effectively learn patient similarity despite these challenges.

Methods

SparGE is an end-to-end discriminative learning framework that tackles EHR data challenges through a two-pronged approach:

  1. Low-Rank Constrained Sparse Coding: This initial step aims to denoise the EHR data and identify similar patients.
  2. Low-Rank Approximation: It uses low-rank constraints to reconstruct a cleaner version of the data, effectively handling missing values and reducing noise.
  3. Sparse Coding: It then represents each patient's data as a sparse combination of other patients' data. The non-zero coefficients in this representation highlight the most similar patients and their relative importance.
  4. Graph Embedding: After obtaining sparse representations, this step measures the similarity between patient pairs.
  5. Preserving Local Relationships: It uses graph embedding on these sparse representations to capture local relationships. This is achieved by learning a low-dimensional embedding of the sparse representations, where the Euclidean distance in this embedded space reflects patient similarity.

A global cost function is constructed to jointly optimize these components, allowing for a more holistic learning process.

Datasets

The proposed SparGE framework was evaluated on two real-world healthcare datasets:

  • SingHEART Dataset: A private dataset containing data from 683 male and female subjects with 467 biomarkers.
  • MIMIC-III Dataset: A large public database of critical care unit patients, comprising information from 53,423 hospital admissions.

Key Insights and Ideas

  • Joint Learning is Key: The paper emphasizes that simultaneously learning the data representation (sparse coding) and the similarity metric (graph embedding) leads to superior performance compared to separate approaches.
  • Low-Rank for Robustness: Using low-rank constraints is a powerful technique for handling missing data and denoising in EHRs, a common problem in healthcare.
  • Sparse Representation as a Bridge: Sparse coding not only identifies similar patients but also creates a more robust and interpretable representation that is amenable to further analysis like graph embedding.
  • Graph Embedding for Local Structure: Capturing local relationships through graph embedding in the sparse domain helps refine the similarity measure, ensuring that patients with similar local patterns are correctly identified.

Key Results

Experimental results on both the SingHEART and MIMIC-III datasets demonstrated that SparGE significantly outperforms existing machine learning patient similarity methods.

  • Improved Accuracy and Recall: SparGE consistently improved accuracy and recall rates across various baseline classifiers when applied to the learned representations compared to using raw data.
  • Feature Selection Benefits: Even when using a reduced set of features (14 for SingHEART, 17 for MIMIC-III), SparGE still achieved better performance.
  • Computational Efficiency: The paper also notes that using representations achieved by SparGE resulted in less computation time for classification tasks compared to using raw data.

Limitations and Future Work

  • Computational Cost: While SparGE shows improved performance, the paper acknowledges that the joint learning framework can be computationally intensive, especially for very large datasets.
  • Parameter Tuning: The performance of SparGE relies on tuning several parameters (e.g., regularization parameters $\lambda_1, \lambda_2, \gamma, \zeta$). The paper mentions cross-validation for tuning but doesn't detail an automated approach.
  • Future Directions: The authors suggest that SparGE is a flexible framework that can be extended to deeper architectures and applied to more general dimensionality reduction problems. They also hint at exploring supervised versions of graph embedding for even more tailored patient similarity learning.

Summary updated

Abstract

Patient similarity assessment (PSA) is pivotal to evidence-based and personalized medicine, enabled by analyzing the increasingly available electronic health records (EHRs). However, machine learning approaches for PSA has to deal with inherent data deficiencies of EHRs, namely missing values, noise, and small sample sizes. In this work, an end-to-end discriminative learning framework, called SparGE, is proposed to address these data challenges of EHR for PSA. SparGE measures similarity by jointly sparse coding and graph embedding. First, we use low-rank constrained sparse coding to identify and calculate weight for similar patients, while denoising against missing values. Then, graph embedding on sparse representations is adopted to measure the similarity between patient pairs via preserving local relationships defined by distances. Finally, a global cost function is constructed to optimize related parameters. Experimental results on two private and public real-world healthcare datasets, namely SingHEART and MIMIC-III, show that the proposed SparGE significantly outperforms other machine learning patient similarity methods.
BibTeX
@inproceedings{wei2022sparge,
  author = {Xian Wei and See-Kiong Ng and Tongtong Zhang and Yingjie Liu},
  title = {{SparGE: Sparse Coding-based Patient Similarity Learning via Low-rank Constraints and Graph Embedding}},
  booktitle = {IJCNN},
  year = {2022},
  url = {https://openreview.net/forum?id=CrINwfpVZ95},
}
Figure