[Paper Review] Visualization of Emergency Department Clinical Data for Interpretable Patient Phenotyping

Nathan C. Hurley, Adrian D. Haimovich+2 more|arXiv (Cornell University)|Jul 5, 2019

Machine Learning in Healthcare4 citations

TL;DR

This paper proposes a two-dimensional visualization pipeline using UMAP and Gaussian Mixture Models (GMMs) to identify interpretable, data-driven patient phenotypes in emergency department (ED) electronic health record (EHR) data. By applying non-linear dimensionality reduction and clustering to real-world ED datasets across five chief complaints, the method reveals stable, clinically relevant patient subgroups with ARI scores ranging from 0.35 to 0.74, enabling real-time triage and retrospective cohort discovery.

ABSTRACT

Visual summarization of clinical data collected on patients contained within the electronic health record (EHR) may enable precise and rapid triage at the time of patient presentation to an emergency department (ED). The triage process is critical in the appropriate allocation of resources and in anticipating eventual patient disposition, typically admission to the hospital or discharge home. EHR data are high-dimensional and complex, but offer the opportunity to discover and characterize underlying data-driven patient phenotypes. These phenotypes will enable improved, personalized therapeutic decision making and prognostication. In this work, we focus on the challenge of two-dimensional patient projections. A low dimensional embedding offers visual interpretability lost in higher dimensions. While linear dimensionality reduction techniques such as principal component analysis are often used towards this aim, they are insufficient to describe the variance of patient data. In this work, we employ the newly-described non-linear embedding technique called uniform manifold approximation and projection (UMAP). UMAP seeks to capture both local and global structures in high-dimensional data. We then use Gaussian mixture models to identify clusters in the embedded data and use the adjusted Rand index (ARI) to establish stability in the discovery of these clusters. This technique is applied to five common clinical chief complaints from a real-world ED EHR dataset, describing the emergent properties of discovered clusters. We observe clinically-relevant cluster attributes, suggesting that visual embeddings of EHR data using non-linear dimensionality reduction is a promising approach to reveal data-driven patient phenotypes. In the five chief complaints, we find between 2 and 6 clusters, with the peak mean pairwise ARI between subsequent training iterations to range from 0.35 to 0.74.

Motivation & Objective

• To address the challenge of interpreting high-dimensional, complex ED EHR data for clinical decision support.
• To develop a visual, interpretable method for identifying data-driven patient phenotypes that reduce reliance on weak diagnostic labels.
• To enable real-time triage and retrospective cohort discovery by projecting patients into a low-dimensional, stable visualization space.
• To evaluate the robustness and clinical relevance of discovered clusters using adjusted Rand index (ARI) and patient-level attributes.
• To demonstrate that non-linear dimensionality reduction (UMAP) outperforms linear methods in capturing both local and global structure in ED patient data.

Proposed method

• The authors apply Uniform Manifold Approximation and Projection (UMAP) to reduce high-dimensional EHR data to two dimensions, preserving both local and global data structure.
• Gaussian Mixture Models (GMMs) are used to identify clusters in the UMAP-embedded space, enabling probabilistic assignment of patients to phenotypic subgroups.
• The adjusted Rand index (ARI) is computed between clusterings from successive training iterations to assess stability of the identified clusters.
• The pipeline is applied separately to five chief complaints (abdominal pain, chest pain, shortness of breath, back pain, falls) from a real-world ED EHR dataset of 560,486 visits.
• The method supports inference on new patients by projecting them into the pre-trained UMAP space without retraining the entire model.
• Clinical relevance of clusters is evaluated by analyzing demographic and triage variables within each group.

Experimental results

Research questions

• RQ1 Can non-linear dimensionality reduction with UMAP effectively reveal interpretable, stable patient phenotypes in complex ED EHR data?
• RQ2 How do the identified clusters compare in stability and clinical relevance across different chief complaints?
• RQ3 Can the visualization pipeline support real-time patient triage and retrospective cohort discovery in the ED setting?
• RQ4 To what extent do cluster characteristics reflect known clinical variables such as comorbidities and triage acuity?
• RQ5 Does the use of UMAP improve cluster stability and global structure preservation compared to linear or non-parametric methods like t-SNE?

Key findings

• The method successfully identified between 2 and 6 distinct patient clusters across five chief complaints, with cluster stability measured by ARI ranging from 0.35 to 0.74.
• Falls showed the highest cluster stability (peak ARI = 0.741), while abdominal pain had the lowest (peak ARI = 0.353), suggesting higher underlying clinical heterogeneity.
• Clusters were clinically meaningful, grouping patients by demographics, triage scores, and comorbidities, such as risk of adverse outcomes in fall-related presentations.
• The UMAP-GMM pipeline enabled projection of new patients into existing visualizations, supporting real-time clinical assessment without retraining.
• The approach revealed that patients with similar pathologies (e.g., myocardial infarction presenting as chest pain or syncope) were not consistently grouped when not stratified by chief complaint, highlighting the need for multi-complaint modeling.
• In fall-related cases, higher-risk clusters were more likely to be admitted, indicating potential for risk stratification and intervention targeting.