Medical records rarely tell a patient’s whole story. Even in the best of circumstances, details and nuances are often lost in records to a listing of symptoms recorded on charts. This loss can be profound when patients are dealing with pain.
But what if a tool was available to assess a patient’s progress or point to how a patient might respond to a specific treatment?
David Juckett and Eric Kasten, research faculty at the Michigan State University Clinical and Translational Sciences Institute, or CTSI, are using natural language processing to build models that represent different phenotypes, or groups of patients with similar characteristics that are experiencing pain.
Natural language processing, or NLP, is a branch of science that works to help computers understand human language. Its most common applications can be found on our smartphones and in-home smart speakers that allow us to turn on lights or change the TV channel with a simple request.
Juckett, professor emeritus in pharmacology and toxicology, and Kasten, associate director of CTSI and assistant professor in radiology, aim to link the human models to treatment outcomes to support pain research and aid decision making for clinicians in diagnosing treating chronic pain.
Their work in NLP began during a collaboration with lead physicians Fred Davis and Mark Gostine at Michigan Pain Consultants, or MPC, in Grand Rapids who began conducting patient surveys on outcomes. Between 2010 to 2014, MPC collected 70,000 patient surveys and 280,000 detailed progress notes from physicians.
These notes revealed information not typically found in patient records, including personal feelings, social circumstances and use of over-the-counter drugs. This information added a new perspective to each patient’s background, as some patients related one thing about their pain level on paper and something entirely different when talking to a doctor just 15 minutes later. This highlighted the complexity of determining a patient’s pain status and underscored the value of using both patient and physician perspectives to better understand and treat the individual.
“The notes were motivation for me to study natural language processing,” Juckett said, “because there are things in such notes that aren’t in other patient records, making access to these reports critical to real pain research.”
Juckett and Kasten’s first challenge was creating a computer program using NLP to analyze the notes efficiently.
“Typically, in medical progress notes, the English is butchered,” Juckett said. “In the case of Michigan Pain Consultants’ notes, the text is much better; nevertheless, computer algorithms need to be constructed to address the peculiarities of the profession, in this case, the medical profession in terms of pain assessment, diagnosis and treatment.”
To overcome this hurdle, Juckett and Kasten constructed an NLP approach to gathering information. Combining patient-reported outcomes and clinical progress notes could provide enough examples of the symptom-to-outcome trajectory to allow them to construct a map of how a patient’s pain would progress.
“We’re looking at a narrow field of pain medicine,” Juckett said. “So we created an ontology, which is a set of categories in a subject area that shows their properties and the relationships between these categories.
According to Juckett, the progress notes can then be processed to link words, phrases, sentences and paragraphs to the ontology. For example, one might identify two categories, "pain" and "stenosis" with the relationship "caused by," concluding that the patient’s pain is caused by stenosis.
“We wanted to extract the meaning from the document’s phrases to understand the biopsychosocial characteristics—or story—contained within the progress notes. We then tried to match the ontology to groups of people and how they might react to treatment.”
The two created a map to connect the patients’ words, the clinicians’ notes and the categories. Their work may lead to a tool physicians can use in helping make pain management decisions.
