- A previously developed natural language processing application was used to examine the presence or absence of documentation about palliative care discussions in electronic health records
- A major finding was that the language clinicians use to document discussions about code status and goals of care differed between 523 adults with advanced pancreatic cancer and 2,093 patients with life-threatening trauma
- Recognizing these differences and deficits will help personalize treatment to the needs of the individual patient, and identifying the root of these differences may help researchers uncover patterns in palliative care delivery
Documentation of clinician–patient communication about serious illness is more likely to be found in free-text notes in a patient's chart than in administrative codes. As an aid to developing quality measures related to care at the end of life, a previous study by Brigham and Women's Hospital researchers, published in the Annals of Surgery, used natural language processing (NLP) to identify documentation of palliative care in electronic health records (EHRs). In the study in advanced cancer patients, a computer algorithm filtered through EHRs to identify preselected words and phrases related to two palliative care processes: clarifying a patient's desired code status and discussing goals of care (GOC).
In a new study of the NLP application, Brooks V. Udelsman, MD, MHS, clinical fellow in the Department of Surgery at Massachusetts General Hospital, David C. Chang, PhD, MPH, MBA, director of healthcare research and policy development in the Codman Center, and colleagues found that the language used to document palliative care discussions is influenced by the type of illness. As they explain in the Journal of Palliative Medicine, recognition of these differences is important for clinicians, not just researchers.
Using patient data registries from Mass General and one of its partner hospitals, the researchers identified 523 adults with metastatic or unresectable pancreatic cancer who were admitted between January 1, 2011, and April 30, 2016, and who had a palliative surgical procedure. They also included 2,093 patients with life-threatening trauma who were admitted between July 1, 2016, and June 30, 2017.
The NLP "codebook"—keywords and phrases—from the previous study mentioned above was applied to each patient's EHR. For comparison, the EHRs were also manually reviewed.
Results of Initial Analysis
Among patients with advanced pancreatic cancer, the NLP codebook identified code status discussions in 54% and GOC discussions in 49%. Compared with manual review, the sensitivity and specificity of the NLP were 94% and 99%, respectively, for a code status discussion and 93% and 100% for a GOC discussion.
Among trauma patients, the NLP codebook identified code status discussions in 26% and GOC discussions in 6%. Compared with manual review, the sensitivity and specificity were 86% and 100% for a code status discussion and 50% and 100% for a GOC discussion.
Exploration of Discrepancies
The researchers reviewed all trauma admission notes in which there was a discrepancy between NLP and manual review. They found:
- Code status discussions: The reduction in sensitivity was primarily attributed to language related to the continuation of treatment. Phrases such as "wishes to remain full code" were more common in trauma patients than in patients with advanced cancer. In addition, some language in the notes on trauma patients reflected more nuanced conversations about code status limitations, such as "code reversal" or "ok to intubate"
- GOC discussions: Phrases related to family discussions tended to be missed in the original codebook. GOC discussions were more frequent for patients with traumatic injuries than those with advanced cancer, and terms such as "ongoing discussion" were used more often
Results of Re-analysis
Prompted by the discrepancies, the researchers added certain keywords and phrases to the NLP codebook. Use of the revised codebook was compared with manual record review for a randomly selected group of 40 trauma patients. The sensitivity rose to 98% for code status discussions and 100% for GOC discussions. The specificity remained 100% for both measures.
The codebook was then applied to the EHRs of all trauma patients. It documented code status discussions in 27% and GOC discussions in 18% (up from 6%).
Individualizing Palliative Care
The subtle differences between the two patient populations are probably related to differences in health trajectories and acuity. These differences include:
- Surgical and medical oncologists are likely to have had a long relationship with a patient, something that's rarely true of trauma surgeons, and this difference may contribute to differences in communication about palliative care
- Major treatment decisions for trauma patients must often be made by family members or other caregivers, and these surrogates are usually less prepared than those who have had time to plan for the patient's end of life
Clinicians should consider how communication about code status and goals of care may need to differ for patients with different types of serious illnesses. Identifying the root of these differences may help researchers uncover patterns in palliative care delivery.
Learn more about Palliative Care & Geriatric Medicine
Refer a patient to the Division of Palliative Care & Geriatric Medicine