ChartWalk: A Tool for Visualising Complex Medical Records

A step towards easy navigation and understanding of the extensive collection of text-based medical records.

Health Practitioners spend 32% of their time on EHR (Electronic Health Record) navigating through the patient’s health records to get a comprehensive overview of their medical history. This is needed to provide them with guidance for the following healthcare decisions. As important as it is, it is still a very time-consuming and tedious process, resulting in EHR burnout. So the question arises, Can the visual analysis strategies help in resolving this issue?

In the last 15 years, much work has been done to visualize text-based records into graphics but most of them face the same shortcomings. They only provide a summary of the text and prevent readers from delving into an in-depth contextual understanding of a single subject. The authors of ChartWalk have explored and developed systems over time including MedStory and Doccurate to bridge this gap. ChartWalk springs from the concept of Graphics+Text visualization where visual summaries give an overview of a patient’s medical history and these visual elements also trace back to the original texts to provide a holistic overview of the context to the practitioner.

Find out more about MedStory and Doccurate here.

Before indulging further into the process of how ChartWalk was developed and what key Design Principles were used, let’s understand the challenges and workflows of Chart Review in Clinical Practises.

What is Chart Review?

Chart Review refers to reviewing “previously recorded data to answer clinical queries” by a health practitioner. It consists of “collecting, distilling, and synthesizing” information from the EHR. However, these reviews do pose some challenges while analyzing due to;

• Semi-structured medical text reports
• Difficulty in the perusal of information, when data exceeds significantly
• Lack of Standardization across documents
• Fragmented and Redundant picture of patient’s medical history because of the distributed nature of data
• Occasional Conflicting, Erroneous, or Missing information in records

Getting Started with ChartWalk

The concept of Graphic+Text Visualization has been actively used in other fields like social sciences, and journalism, bridging the gap between high-level overview of texts and low-level text analysis of one document, which provides design inspiration and insights into how it can be applied to the Medical domain. This overlap is only possible because all of them share a similar need for understanding of context, causality, and nuance present in text which needs both overview and analysis of little details.

Design Principles for ChartWalk

Based on the user experience of Doccurate and Medstory:

• Overview: To provide an overall picture/summary of the record
• Time: Mapping medical events on a Timeline and drawing relation across events
• Facets: Organising information via grouping related information such as “all medications”
• Traceability & Context: Easy retrieval of original documents to understand the context better.

Design Implementations

Using Natural Language Processing tools along with the Unified Medical Language Models, the data from EHR was grouped into broader categories such as Problems, Medications, Labs, etc. The grouping was further stratified based on system-level categorization such as Cardiology and dermatology to organize and structure data for easy retrieval and analysis.

Even though, NLP progressively helped in extracting relevant details of information and populating them into timelines to detect and mark medical events. But it also raises the issue of trust generated from automated data visuals, which are prone to error and need vetting by a clinical practitioner.

The design process was carried out in 2 phases and involved constructive feedback sessions with Health Practitioners from different areas of expertise to study the overall conviviality.

(A) Mention View: Overview with a sparkline (A.1) to show variation in each category and draw a quick comparison across categories along with scatterplot pop-ups (A.2) which show a detailed report of each category. (B) Snippet View: Showing episodes of care: documentation of data from a single hospitalization by extracting admission and discharge dates. (C) Calender View: Conveying the month and number of notes associated with it. (D) Note View: Shows the reader’s selected highlights from original texts, and allows the reader to add a curated summary (E) Search Bar: Shows frequency of occurrence for the current problem. (F) Legend Bar: Filter notes by type of data

Findings:

• Mention view (A) is helpful in grouping the categories together and allowing viewers to choose the category they want to study the data from, thereby reducing the influx. However, the list seems too long and text-heavy and, therefore requires further refining. The data was simplified and structured using skyline and scatterplot diagrams.
• Calendar View (C), though difficult to read at first, conveyed wellness patterns very effectively. It also provided the flexibility to set the chronological order as per the Reader’s choice.
• Snippet View (B) provided a curated overview but was text-heavy. Organizing text using typefaces and sizes with clear separation can improve readability. This was further improved by grouping data with the same context in snippet view under ‘episodes of care’ and highlighting the chief complaint in that event, increasing the reader’s efficiency.
• Adding a Curated Summary, where practitioners can add their notes as a tool to communicate with their colleagues, for future reference, or while referring to another health expert.
• Stratifying the notes into different types, except just discharge summaries in past models, helped in further reducing cognitive load as the reader can easily filter data.

Detail categorization of data in the snippet view to filter the information by the reader, reducing cognitive load and increasing time efficiency.

• Highlighting certain notes helped reduce reading time.
• Individual’s Workflow analysis:

How different health practitioners use the same tool differently. This largely depends on their field of practice. For instance, a mental health practitioner will spend more time on Note View for an in-depth understanding, whereas, in an emergency, they will spend more time on Mentions View for a quick overview.

Chartwalk’s ability to support different workflows based on individual preferences helps them pace better on the learning curve for a new tool. The ability to quickly find relevant data as it is more structured, and organized and get an overview of the medical history at the same time while allowing the user to delve into an in-depth understanding of certain areas provided certain positive feedback but also made users a little uncomfortable. They had to learn a new tool to navigate this longitudinal data while their brain was already accustomed to the process that they might have used for decades.

Conclusions:

Reading is highly prevalent in the medical field and is likely to be so even after the creation of automated content visual-graphical summaries and text summarization tools. It provides a more detailed description of medical history even though it is time-consuming. It is also a skill that health practitioners are trained for and is considered crucial in the field.

But visual representation, using NLP (Natural Language Processing) for automated extraction, and design features like curation summary and others mentioned above, can be used to aid as support to text, in structuring the extensive text data of medical health reports which can be useful for health practitioners.

References:

Access Chartwalk at: http://chartwalk.cs.toronto.edu/

Video: [Presentation] ChartWalk: Navigating Large Collections of Notes in EHRs for Chart Review (VIS 2022)

ChartWalk: Navigating Large Collections of Text Notes in Electronic Health Records for Clinical Chart Review

N. Sultanum, F. Naeem, M. Brudno and F. Chevalier, “ChartWalk: Navigating large collections of text notes in electronic health records for clinical chart review,” in IEEE Transactions on Visualization and Computer Graphics, vol. 29, no. 1, pp. 1244–1254, Jan. 2023, doi: 10.1109/TVCG.2022.3209444.