Making Doctors More Accurate

Enhancing clinical texts with visualization.

Most of us have experienced having to reiterate our entire medical history when we visit a physician for an ailment. This process of walking the physician through our medical records takes painstakingly long and is almost as exasperating as suffering the ailment itself. At the same time, it is no picnic for the physician either. Clinical physicians face an impossible task of dealing with ample amounts of patient data. They consult hundreds of patients every day; keeping a track of this copious amount of data is both time-taking and arduous because the data is complex and extremely text-heavy.

Text is the preferred means of storing this data since it encompasses medical issues and various contextual factors. However, this exercise gets laborious when the text spans from a few lines to hundreds of records. Due to the continuous nature of medical treatments, physicians are unable to conduct a comprehensive review of the patient’s record.

Prior to seeing the patient, a physician reviews the patient’s medical chart and captures the key medical issues to form a problem list. They use this list of symptoms to compile a proto-narrative with the probable diagnosis of the patient. When there’s not enough time to put together such a list, physicians adopt other methods to save time by skimming through notes, focusing on only recent information or as a last resort, asking the patient themselves. While these methods do a good job of covering crucial information, it may leave out small but significant details hiding in the patient records.

So, What Can Be Done?

According to a Canadian research team led by Fanny Chevalier at University of Toronto, visual curation along with medical NLP can be used to overcome these challenges. The researchers argue that the use of NLP in clinical practice should be framed in a more physician-centered manner that

1. allows for a continued adaptation of automation to personal and evolving information needs,
2. fosters a more verifiable and adequate reliance on automated tools,
3. integrates into existing clinical workflows, and
4. supports efficiency.

The researchers propose Doccurate: a data visualization tool to support an overview of large patient charts.

Design Goals

1. Preserve the original text: Summarization of medical text serves no purpose. The original text (with jargons and medical terms) must be preserved.
2. Foster suitable trust over automated output: Even with automation, the control should mainly lie with the physicians to determine the extent to which they want to use automation.
3. Provide information in different levels of granularity: Physicians may or may not have the time to read the entire medical record of the patient. So, there must be a way to support understanding the text whether they want to glance through it or read it meticulously.
4. Support user-driven customization: Physicians’ personal preferences are diverse, so it is important to support customization to support their personal needs.
5. Convey time and progression: Temporal references should be thoroughly supported to understand how medical problems evolved over time.
6. Support content faceting: The patient’s record may have several medical issues pertaining to the patient. However, it is important to allow the physician to concentrate on the relevant issue without having to read the entire record.

Designing Doccurate

Building blocks — Filtering components

Filter Collections (FCs) are curation building blocks, consisting of physician-defined semantic filters to create faceted views of clinical content.

Text Pre-processing

Patient chart documents were preprocessed for entity recognition using Apache cTAKES, an open source named entity recognition (NER) tool for clinical text. Past benchmarking efforts indicated good but not excellent performance.

Interface Design

Doccurate’s interface is divided into four panels:

1. Control Panel: Lists patient demographics, document filters, and (FCs)
2. Timeline: Provides a time-oriented overview of lists of tagged content encompassed by FCs. Clicking a timeline snippet redirects the physician to the corresponding passage in the Text Panel.
3. Text Panel: Lists the documents in the patient chart. The Text Panel contains all chart documents, concatenated in a long scrolling list and chronologically sorted.
4. Curation Panel: to create and edit FCs.

Evaluation

Five residents and one physician in General Practice (GP) from seven different Canadian healthcare institutions were recruited to participate in the study.

The researchers conducted the study for two hours with predefined tasks to evaluate the working of Doccurate and tabulated the results.

Conclusion

Following the evaluation study, researchers proposed the use of Doccurate in curating chart reviews, in ongoing medical treatments, and for an overview of emergency care. Compelling evidence was found to support the fact that the curation-based approach can bring value to clinical practice. In particular, it was observed that physicians were able to easily identify automation errors and found value in the information collected by the FCs.

Doccurate represents an initial (but solid) attempt at text visualization for clinical practice. It is a great start and holds tremendous scope for future work.

This blog post is inspired by the following article:

• Sultanum N, Singh D, Brudno M, Chevalier F. Doccurate: A Curation-Based Approach for Clinical Text Visualization. IEEE Transactions on Visualization and Computer Graphics, 2018 Aug 20. doi: 10.1109/TVCG.2018.2864905.