Addressing the challenge of temporal reasoning in healthcare

In order to better understand a patient’s condition, temporal information is essential – but not always easy to find in the electronic health record.
Temporal Reasoning

“Doctor, you just requested a chest CT for one patient and an echocardiogram for another. Both patients are unsure which procedures have been ordered for them. I am sending you their most recent problem lists. Could you please review these and let me know which test to order for whom? Thank you.”

Click.

Ah, the joys of rapid patient turnover in a busy clinic…

Consider the two patients above whose problem lists include only four diagnoses without any temporal associations. Patient A and Patient B have problem lists that appear to be identical. Both have four items listed: diabetes mellitus, smoking history 30 years, lung cancer, and status post myocardial infarction.

So which patient smoked for 30 years and then developed lung cancer (for whom you wish to order a chest CT)? Which patient was diagnosed with lung cancer and has smoked for 30 years since then? And which 75-year-old patient recently had a myocardial infarction and needs an echocardiogram?

Temporal information is vital if we hope to gain some understanding about our patients. It turns out that these two different patients are in fact very different.

Temporality and the interpretation of health data

Contrast how more detailed problem descriptions alter our understanding of these individuals. A patient with diabetes mellitus for 40 years has likely developed additional sequelae of the kidney and heart, along with vascular disease that a newly diagnosed patient might not yet have. Similarly, recent lung cancer after a 30-year history of smoking is much more ominous than lung cancer 30 years prior.

With no additional information available, we might surmise that an elderly patient who suffered a myocardial infarction (MI) in their 50s has lived with a substantial burden and chronic health risk for an extended period, while a recent MI in a patient may need review of the acute cardiac impairment. Temporality moderates what we infer about the patient’s disease progress and how we tailor the monitoring and treatment plan.

Temporality fills a critical gap in the interpretation of health data, providing agency for deriving meaningful associations between observations, treatments, and outcomes. Time –  along with elements such as findings, problems, procedures, and orders – is a requisite component of an event. Temporality enables the type of phenotypic associations needed for precision medicine and longitudinal electronic medical records (LEMRs).

It is also vital for deriving meaningful links between medical history, treatment, and health outcomes; as well as for developing capabilities such as advanced decision support systems. Precision medicine, artificial intelligence (AI), data analytics, and predictive modeling are all rooted in highly granular data – and temporality can facilitate the interpretation of relationships among that data.

Two clear scenarios stand out.

  1. For precision medicine, temporality supports natural language processing (NLP) and helps assemble data into a LEMR. This provides relationships – such as age at event, length of event, and sequence of events – from records across multiple sources of data.
  2. For population health research, time associations provide a fundamental tool for AI systems and machine learning. Temporal relationships can elucidate context and enable querying across medical records in a system for very similar patients. These, in turn, can help identify similar patient records and enable the aggregation of these comparable records into cohorts.

Interpretable versus plottable temporal phrases

Due to temporal qualities an event may have an uncertain beginning or conclusion; be ongoing; have a relationship with other events; have a sequence; be momentary or have a span; or have parts, causes, results, or recurrency patterns. Text fields appearing throughout the medical record contain essential narratives describing health events, and NLP can highlight and codify these narratives and their temporal qualities so they can be used to construct a health timeline.

For a temporal phrase to be understood it will often include a specific point or range in time, a general chronology or sequence of events, or the possibility of when an event may have occurred. Unlike a computer, for a person to infer a temporal meaning for an event a phrase is not required to include the use of numbers, dates or other clearly defined time units or phases. But even the existence of an event that does not include when it occurred or has an ambiguous start may imply temporality that a person can generally understand such as previously, the patient experienced headaches, but that was some time ago.

On the other hand, at a minimum, the practical use of plotting an event on a patient’s health timeline requires some sort of measurable timeframe. With the goal of compiling a unified timeline of health-related events for a patient, organizing and incorporating the free text found in a patient’s many records provides a robust reservoir of data.

The minimum requirements to calculate temporality from free text in an entry on to a health timeline can be stated as follows: To be utilizable, temporal text must permit quantified interpretation leading to a specific point or range in time either by calling out a specific timeframe (like age or date) or by giving a quantifiable time association with a timestamp that is associated with either an element or event.

Key Takeaways

  • Text in medical records narrate significant events in patient histories
  • Those events are constructs that include both elements and time frames
  • Determining when an event occurred is of premiere importance in contextualizing the history and supporting data analytics and predictive modeling
  • NLP can identify temporal concepts which, in turn, can provide a means to determine when events occurred
  • Plotting events on a health timeline is key when comparing health histories and identifying very similar patients to assist in clinical decision support
  • Other uses for temporal knowledge include precision medicine initiatives, medical and pharmaceutical research, and population heath studies

For a deeper dive from Jonathan Gold, MD and Steven Rube, MD, watch their recorded session on “The Longitudinal Medical Record and Natural Language Processing: The Challenge of Temporal Reasoning” at the AMIA 2022 Annual Symposium.

Ideas are meant for sharing.

Sign up today and have Ideas delivered straight to your inbox.

Related Ideas