Readers Write: Taking Clinical Natural Language Processing Mainstream for Effective Care Management

Taking Clinical Natural Language Processing Mainstream for Effective Care Management
By Kevin Agatstein

Kevin Agatstein is CEO of Kaid Health of Boston.

Across healthcare, clinical natural language processing continues to play an ever more influential role. Kreimeyer et al.’s “Natural language processing systems for capturing and standardizing unstructured clinical information” identified over 70 different CNLP systems in the literature, spanning multiple clinical domains. Unfortunately, few of these directly address the applicability of CNLP to care management. This lack of CNLP supporting care managers will and should change. Making this reality will require adapting the technology to the real-world needs of care management programs and the front-line clinicians who staff them. 

To fuel effective care management, actionable data is required across the entire workflow. Examples of actionable data include information that: 

• Identifies which patients require help.
• Stratifies patients for intervention.
• Summarizes the patient for the care manager.
• Determines the specific actions the patient needs.
• Uncovers the barriers to effective care.
• Measures intervention outcomes.

Claims data, lab data, health risk assessments, and motivational interviewing all meaningfully contribute to the above. While all of these are necessary, alas, they are not sufficient. For the care manager to meaningfully coordinate patient care, to accomplish the six steps listed above, they must have access to actual clinical data. They need the medical record. More precisely, they needs the nuggets of actionable insights buried in the massive EHR data set. Finally, they need it to be quickly digestible. Thus, CNLP can not only help, it is all but required.

This actionable data is almost all in the EHR; however, it can be hard to find. A patient’s medical record is often hundreds of pages of text, alongside hundreds of discrete data points (labs, medications, allergies, etc.) Within this morass of usually loosely organized data is the patient’s health history. While claims and labs can give some sense of the patient’s clinical experience, the chart has the diagnosed but not coded conditions, the written but not filled prescriptions, and more. It also has a plethora of exam findings, laboratory reports, radiologic data, and pathology findings that never get put into “structured” EHR fields.

Kharrazi et al., in “The Value of Unstructured Electronic Health Record Data in Geriatric Syndrome Case Identification,” found that the EHR text resulted in finding 1.5 times more patients with dementia than just reviewing the structured EHR data. That same ratio was 1.7 with decubitus ulcers, 2.9 for weight loss, and 3.2 for a history of falling.

Beyond traditional clinical data, the chart often contains insights into the patient’s family health history. It also has data on psychosocial barriers to care, limitations on activities of daily living, and other elements impacting the patient’s care journey. Just as crucial for care managers, the chart typically has data on the patient’s social determinants of health. While SDOH are almost never coded in claims, (and yes, there are ICD-10 SDOH codes), they are noted in charts. AI-powered healthcare data analysis and provider engagement platforms have found hundreds of SDOH in primary care, specialists, ED, and behavioral health charts. Kharrazi found similar results. For example, they found that it is 456 times more likely to find a patient with a “lack of social support” in the free text of the medical note than in the structured data.

For a care manager to do their job well, this data cannot be ignored.

More than just summarizing the patient’s health, the medical record can help translate the EHR text into a structured, actionable, trackable ambulatory care plan by summarizing the physician’s treatment plan noted for each encounter. Specifically, NLP can create a patient to-do list such as follow-up visits, getting testing or labs, addressing unhealthy behaviors, and more. These identified tasks can become the basis of a care management care plan or added to existing plans. As new data enters the chart, either as structured information or new medical notes, the to-do list can be updated. Tasks can be marked as completed, new tasks added, existing tasks amended, and more.

It’s important to remember that NLP algorithms do not digest a medical note the way a human does. Rather, they predict how a trained human would interpret the presented text. This is much more than finding key words. CNLP solutions also need to account for:

• Negation (“does not have cancer”).
• Family history (“the patient’s mother had an MI before age 55”).
• Uncertainty, (e.g., “initial lab findings mean early-stage chronic kidney disease possible, but additional testing is needed”).
• And more.

Making such determinations isn’t perfect, but making useful interpretations of clinical text has been proven possible. Moreover, CNLP does not fatigue as humans do. For example, Suh, et. al. found in “Identification of Preanesthetic History Elements by a Natural Language Processing Engine” that CNLP frequently identified salient clinical facts that a physician reviewer missed. 

Now, new data standards, notably FHIR, and regulatory mandates to share data combine to markedly simplify a CNLP deployment process. This, plus cloud and other emerging data exchange standards, mean CNLP go-lives can be measured in days, not months. By working with partners with rigid technological and workflow controls, extensive security training, and a culture of data security, the data can be processed safely as well.

For a real-world deployment, a care management CNLP solution should be intuitive to clinicians. It should be focused on the needs of care managers to anticipate the workflow. Care managers today deal with several different medical record and care management documentation systems. Effectively managing these variations, and the vagaries of existing workflows, comes only with experience. Most importantly, CNLP needs to add value for the user practically out of the box. They can, and they will.