Visualization of Electronic Health Record Sequences at Scale
\authororcid Ambre Assor0000-0002-1825-0097, \authororcid Mickael Sereno0000-0000-0000-0000 and \authororcid Jean-Daniel Fekete0000-0002-1825-0097 · 2025
Read Visualization of Electronic Health Record Sequences at Scale by \authororcid Ambre Assor0000-0002-1825-0097, \authororcid Mickael Sereno0000-0000-0000-0000 and \authororcid Jean-Daniel Fekete0000-0002-1825-0097 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-10700.txt.
Introduction
Ambre Assor, Mickael Sereno, and Jean-Daniel Fekete present ParcoursVis, an open-source tool described in a 2025 paper, for exploring patient record sequences at very large scale. Tools of this kind show the order of care events across many patients, which can reveal rare conditions and odd care paths. Older tools stall at roughly twenty thousand patients before they stop feeling interactive.
ParcoursVis instead computes an approximate result fast, shows it as an icicle tree, and then refines it step by step until the full count is done. This design keeps the tool responsive while handling millions of patients, about a thousand times more than rival tools. The authors describe the design choices that give the system its speed, its fast settling, and its stable display during updates.
Catalog record
| CREATOR | \authororcid Ambre Assor0000-0002-1825-0097, \authororcid Mickael Sereno0000-0000-0000-0000 and \authororcid Jean-Daniel Fekete0000-0002-1825-0097 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 77 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-10700
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Visualization of Electronic Health Record Sequences at Scale \authororcid Ambre Assor0000-0002-1825-0097, \authororcid Mickael Sereno0000-0000-0000-0000 and \authororcid Jean-Daniel Fekete0000-0002-1825-0097 Abstract We present ParcoursVis, a Progressive Visual Analytics tool designed to explore electronic health record sequences of patients at scale. Existing tools process and aggregate the whole dataset upfront before showing the visualization, taking a time proportional to the data size. Therefore, to remain interactive, existing tools are limited to data sizes that can be processed in under a few seconds to meet the latency constraints of human attention. To overcome this limitation and scale to larger sizes, ParcoursVis relies on a progressive algorithm that quickly shows an approximate initial result of the aggregation, visualized as an Icicle tree, and improves it iteratively, updating the visualization until the whole computation is done. With its architecture, ParcoursVis remains interactive while visualizing the sequences of tens of millions of patients, each described with thousands of events; three to five orders of magnitude more than similar systems. Managing large datasets allows for exploring rare medical conditions or unexpected patient pathways, contributing to improving treatments. We describe the algorithms we use and our evaluation concerning their scalability, convergence, and stability. We also report on a set of guidelines to support visualization designers in developing scalable progressive systems. ParcoursVis already allows practitioners to perform analyses on two large real medical datasets. Our prototype is open-source. \vgtccategory Research \vgtcpapertype algorithm/technique \authorfooter Ambre Assor, Mickael Sereno, and Jean-Daniel Fekete are with Université Paris-Saclay, CNRS, and Inria, E-mail: [email protected], [email protected], [email protected] \teaser The ParcoursVis system, designed to explore electronic health record sequences of patients at scale. Visualization of Electronic Health Record Sequences at Scale the main view, shows 10 million medical pathways for patients treated for non-cancerous prostate adenoma, visualized as an Icicle tree . Visualization of Electronic Health Record Sequences at Scale shows the number of filtered out and total patients. Visualization of Electronic Health Record Sequences at Scale shows, in the control panel, the legend and overall distribution of event types. Visualization of Electronic Health Record Sequences at Scale shows the hierarchy of event types used to summarize the tree. Visualization of Electronic Health Record Sequences at Scale zooms into a sub-tree, revealing more of the hierarchical structure. Each colored node represents a stage in a sequence of events starting from the root on the left, with time flowing to the right. For each node, the next event nodes are on the right, sorted from bottom to top by frequency. The node height is proportional to the frequency of the event, and the width to the average duration. In this dataset, most patients started with the “alphabloc” treatment (bottom left), and kept taking it for life (a few patients changed treatment since the node’s right side is almost empty). {subcaptiongroup} keywords Scalability, Progressive Visual Analytics, Temporal Event Sequences, Electronic Health Records, Tree visualization, Icicle tree. 1 Introduction As datasets continue to increase in size and complexity, traditional visualization techniques struggle to maintain interactivity and responsiveness, thereby hindering effective exploratory data analysis. Addressing the scalability challenge necessitates progress in algorithmic implementations and the development of optimized data structures within visualization systems, but more importantly, a change in programming paradigm, using progressive visual analytics (PVA) to decouple data size and algorithm complexity from latency [ 13 ] . The scalability challenge is particularly acute in healthcare, where Electronic Health Records (EHRs) contain vast amounts of patient information. For instance, the French National Health Database [ 41 ] holds decades of health data on citizens, encompassing records of drug purchases, medical diagnoses, and hospital treatments. Such data have the potential to advance healthcare by enabling a better understanding of treatment practices and patient outcomes. As an example, by reconstructing and visualizing temporal event sequences from drug purchase records, the examination of patients’ treatment journeys over time is facilitated. This approach aligns with the goal of medical professionals to assess whether patients’ treatments adhere to national guidelines and, if deviations occur, to understand the reasons. This is key to providing patients with the highest quality of care through evidence-based and safe recommendations. Yet, the volume of such datasets necessitates innovative visualization approaches capable of handling large-scale medical data while preserving analytical efficiency and usability. In this context, we developed ParcoursVis ( Visualization of Electronic Health Record Sequences at Scale ), a PVA tool created in collaboration with medical institutions to visualize patient care pathways derived from EHRs. Existing work has laid a foundation for similar visualization tools aimed at healthcare analysis (e.g., [ 5 , 28 , 31 ] ). For instance, EventFlow [ 29 ] aggregates event sequences into a prefix tree and visualizes them as an Icicle tree , as shown in Visualization of Electronic Health Record Sequences at Scale ; a visualization that has proven effective in identifying optimal drug usage patterns. We, therefore, build on the EventFlow visualization technique for ParcoursVis, along with several improvements. However, tools like EventFlow have generally been applied to relatively small datasets, sometimes encompassing fewer than 100 patients. According to our experiments, EventFlow is limited to about 20k patients. While such tools have offered valuable initial insights, they fall short in scalability, which has been reported as a main challenge for EHR visualization [ 49 ] . In practice, regional and national EHR databases present data at a far greater scale, often involving millions of patients. To our knowledge, no existing tool enables the management and interactive analysis of EHR datasets exceeding 100k event sequences [ 51 , 50 , 29 , 49 ] , which limits the scope of exploration. This constraint impedes the ability to gain a more comprehensive view of patient pathways, identify patterns and trends that may not be visible in smaller or sampled datasets, and discover rare patterns and outliers. With ParcoursVis, we allow visualizing temporal event sequences of more than 100M patients, potentially supporting the largest EHRs. Large-scale interactive data processing poses significant challenges, particularly in terms of response times. To accommodate the need for a responsive system that supports health analysts in conducting explorations on large-scale EHR data, we rely on a Progressive Visual Analytics architecture [ 42 , 13 ] . Our progressive approach provides a quick, approximate, yet useful preview of the data early on to maintain user attention within human latency limits and quick updates converging within seconds. Indeed, results from research in visualization show that user attention declines after 500 ms, and after five to ten seconds, users tend to abandon tasks or lose focus [ 52 , 25 ] . After the initial preview, our algorithm continues processing the dataset, updating the visualization every second or so until it is fully processed. Yet, the progressive updates of the aggregated tree can cause visual instability, as the nodes (i.e., rectangles on Visualization of Electronic Health Record Sequences at Scale ) are laid out at each update. The shifting of nodes can disrupt the viewer’s mental map and create distracting visual flickers. Therefore, our approach may have an impact on the usability of the tool. To mitigate this issue, we introduce a sorting with hysteresis algorithm that limits this flicker during the progressive rendering. In summary, this article presents our system and reports