Patient similarity: methods and applications

Leyu Dai, He Zhu, Dianbo LiuarXiv 2020 · OpenReview

TL;DR

This paper provides a comprehensive review of patient similarity analysis, a crucial technique for personalized medicine. It breaks down the process into key steps – data integration, similarity measurement, and neighborhood identification – and surveys the diverse methods and applications, highlighting its growing importance in healthcare.

What Problem Does This Paper Address?

The paper addresses the increasing need for personalized healthcare, particularly in the context of precision medicine. As healthcare systems generate vast amounts of patient-specific data (electronic health records, genetic information, etc.), effectively utilizing this data to tailor treatments and diagnoses for individual patients becomes paramount. However, the heterogeneity of this data and the complexity of patient relationships pose significant challenges. This review aims to consolidate and explain the methods and applications of patient similarity analysis, which seeks to quantify how alike patients are, thereby enabling more informed clinical decision-making and personalized interventions.

Methods

The paper outlines a typical patient similarity study workflow, which generally involves:

  • Data Integration: This is a critical first step due to the heterogeneous nature of patient data (numeric, categorical, temporal, etc.). The paper discusses two primary strategies:
  • Early Integration: Data from different sources are combined and standardized into a single feature vector before similarity is measured. This can involve transformations to a common format but may lead to information loss.
  • Late Integration: Similarity is calculated independently for each data type, and then these individual similarity scores are combined into a unified measure. This avoids some data transformation issues but might miss inter-data type relationships.
  • An intermediate integration strategy is also mentioned, which infers a joint model without explicit data transformation, aiming to minimize information loss.
  • Similarity/Distance Metrics: Various methods are presented to quantify the similarity or distance between patients:
  • Unsupervised Methods: These use predefined metrics to find patterns without explicit labels. Examples include Euclidean distance, Mahalanobis distance (which accounts for correlations between features), and cosine similarity (useful for normalized data). Jaccard similarity is mentioned for binary attributes.
  • Supervised Methods: These leverage "ground truth" labels (e.g., clinician judgments) to learn similarity metrics. This often involves adjusting feature importance or learning a generalized distance metric.
  • Semi-supervised Methods: These combine labeled and unlabeled data, offering a practical approach when complete labels are unavailable.
  • Neural Network Learning: The paper notes the growing use of neural networks, including CNNs, for analyzing patient data and predicting similarities or health status.
  • Active Learning: This strategy allows the learning algorithm to select the most informative data points for labeling, which is particularly useful when obtaining expert labels is expensive.
  • Neighborhood/Cluster Detection: Once similarity is established, methods are used to group or identify similar patients:
  • Neighborhoods: Defined by a similarity threshold or using K-nearest neighbors (KNN). Weighted sampling can also be used to select informative neighbors.
  • Clusters: Algorithms like K-means and spectral clustering are used to partition the patient population into groups based on similarity. Self-organizing maps (SOMs) are also discussed for visualizing and understanding patient clusters.

Datasets

Not explicitly discussed in the paper. The paper focuses on the methodologies and applications rather than specific datasets used for evaluation. It refers to various types of patient data, including electronic medical records, genetic data, biochemical tests, social media data, genomics, transcriptomics, proteomics, and other "omics."

Key Insights and Ideas

  • Data Heterogeneity is a Core Challenge: The paper emphasizes that patient data comes in many forms, making data integration a foundational and complex step in any similarity analysis.
  • Precision Medicine is a Major Driver: Patient similarity is presented as a key enabler for precision medicine, allowing for personalized treatments and disease management.
  • No One-Size-Fits-All Metric: The choice of similarity metric is highly dependent on the data type and the specific problem being addressed.
  • Network-Based Approaches Offer Power: Building a comprehensive patient similarity network can lead to better predictive models than focusing on isolated patient groups.
  • The Field is Maturing but Still Developing: While patient similarity research has made strides, it's still considered a relatively limited area compared to other person similarity fields, with ongoing challenges and opportunities.

Key Results

The paper is a review and does not present novel experimental results. Instead, it summarizes the applications of patient similarity analysis across various domains, including:

  • Patient and Disease Sub-typing: Identifying distinct subgroups within diseases (e.g., glioblastoma, type 2 diabetes) for better understanding and targeted therapies.
  • Personalized Medical Predictive Models: Improving the accuracy of predictions (e.g., mortality in ICUs, discharge diagnoses) by leveraging similar patients.
  • Personalized Treatment Design: Estimating treatment effectiveness and recommending personalized interventions by analyzing patient similarity networks alongside drug and gene data.
  • Selecting Informative Patients: Identifying key patients from a cohort to optimize data collection or medical decision-making, especially when data is limited.

Limitations and Future Work

Limitations:

  • Data Integration Complexity: The paper acknowledges that integrating heterogeneous data remains a significant challenge, and strategies like early integration can lead to information loss.
  • Limited Research Scope: Compared to other person similarity fields, patient similarity research is still relatively limited.
  • Data Scarcity: In some applications, obtaining complete patient data can be difficult due to resource constraints.

Future Work:

  • Intermediate Data Integration: Further exploration of intermediate data integration strategies (e.g., non-negative matrix factorization) to minimize information loss.
  • Advanced Similarity Metrics: Development and application of more sophisticated supervised and semi-supervised similarity metrics.
  • Novel Neighborhood/Cluster Detection: Utilizing advanced techniques like growing SOMs and semi-supervised clustering.
  • Domain Adaptation and Transfer Learning: Applying these modern technologies to patient similarity networks to improve accuracy and generalization, especially in personalized clinical analysis.
  • Precision Medicine Applications: Further leveraging patient similarity networks for applications in precision medicine, which is seen as a specialized form of personalized data science.

Summary updated

Abstract

Patient similarity analysis is important in health care applications. It takes patient information such as their electronic medical records and genetic data as input and computes the pairwise similarity between patients. Procedures of typical a patient similarity study can be divided into several steps including data integration, similarity measurement, and neighborhood identification. And according to an analysis of patient similarity, doctors can easily find the most suitable treatments. There are many methods to analyze the similarity such as cluster analysis. And during machine learning become more and more popular, Using neural networks such as CNN is a new hot topic. This review summarizes representative methods used in each step and discusses applications of patient similarity networks especially in the context of precision medicine.
BibTeX
@misc{dai2020patient,
  author = {Leyu Dai and He Zhu and Dianbo Liu},
  title = {{Patient similarity: methods and applications}},
  year = {2020},
  eprint = {2012.01976},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2012.01976},
}
Figure