Special Topics in Pain: Opioids
Enhancing Risk Assessment in Patients Receiving Chronic Opioid Analgesic Therapy Using Natural Language Processing.
Author(s): Haller, et al.
Journal: Pain Medicine. 2017; 18: 1952-1960. 31 references.
Reprint: Irina V. Haller, PhD, MS, Essentia Institute of Rural Health, 502 East Second Street (6AV-2), Duluth, MN 55805, USA. E-mail: Ihaller@eirh.org
Faculty Disclosure: Please see original article for disclosures. Abstracted by N Walea, who has nothing to disclose.
Objective: Review and evaluate the latest advances and newest information in the area of Opioids; Review and evaluate the latest advances and newest information in the area of Psychology of specific Groups (e.g. veterans, elderly, Children, war zones, Third world nations)
Editor’s Note: Electronic health records will eventually contain the vast majority of all documentation, and as such, allows for easy access of prior information utilizing automated techniques. Some of this can be used for good purposes, such as in this case, where the information could potentially be used to help a clinician determine a patient’s risk to abuse or misuse opioids. However, it can also be used to review charts in legal proceedings quickly, placing blame on a clinician for not reviewing every last bit of documentation in a large chart spanning years to decades. Natural language processing can help quickly review information, even in free text documentation, where a prior provider did not “check a box” indicating a medical issue, but did document it as typed text somewhere in the progress note. As EMRs become more advanced and natural language processing is utilized more and more, it can hopefully help clinicians better care for their patients and not miss important data hidden in years of progress notes.
Class: Opioid use: Legal considerations; Electronic medical records
Clinical guidelines for use of opioids in chronic pain recommend risk assessment of aberrant drug-related behaviors (ADRBs) prior to initiating opioid therapy. Investigators have successfully applied natural language processing (NLP) techniques to clinical notes and computer-assisted manual review of electronic medical records (EHRs) to identify clinician-documented problem opioid use. This suggests that semi-automated NLP techniques could be used for identifying problem opioid use among patients on long-term opioids.
This research was conducted in a regional health care delivery system with a large rural service areas in Minnesota, Wisconsin, North Dakota and Idaho, that uses Epic EHR to document health care delivery in ambulatory and inpatient settings. The study cohort for this retrospective review of EHR data included all adult ambulatory care patients who signed at least one opioid agreement during calendar years 2007 through 2012. Patients who had cancer or severe mental health conditions prior to their first qualifying opioid agreement were excluded from the study. The opioid agreement was implemented for patients on chronic opioid analgesic therapy (COAT) in primary care settings in 2006.
Structured data include information on patient demographics, diagnosis codes, medications procedures, and laboratory results. The unstructured (free-text) data include clinical notes regarding clinician-patient communication, history, physical examination, consultations, assessments, treatment plans, progress notes, and notes associated with hospital and Emergency Department admissions and discharges.
Two data extraction approaches were used to identify potential opioid-related problems documented in the EHR. Queries of the EHR relational database were performed to populate structured data tables, including coded entries in the current health care issue listing, encounter-based ICD-9 diagnosis codes, health care encounter information, opioid prescription/dose information, and patient demographics. Regular-expression NLP techniques were used for extracting unstructured data, such as text in the clinical notes. Sets of keywords and phrases for each outcome and domain of the risk stratification tool were developed. In addition, terms for negation, uncertainty, and references to people other than the patient have been adapted from the open source NLP applications.
Opioid agreement violations could be found either as a code for "violated opioid agreement" in current health issue listing or as a free-text statement ("violated opioid agreement", "breach of a pain contract", etc.). Additionally, four ADRBs were extracted from either structured or unstructured EHR data. The first two were "patient requests early refill of medication" and "patient reports lost/stolen medications". The other two were "patient abusing alcohol" and "patient is using illicit drugs".
The extracted data was used to populate an existing assessment instrument--the Opioid Risk Tool (ORT)--being used by the healthcare system.
The NLP algorithm identified nearly four times more COAT patients with opioid agreement violations compared with what was documented in the structured data only. Rates of alcohol abuse documentation were higher in the unstructured sources than the structured sources and similar between the two data sources for illicit drug use. Other ADRBs, early refill requests and reporting of lost/stolen prescription were determined on the unstructured data sources only and had the lowest rates in this population (4-5%).
The clinical notes of COAT patients are rich sources of unstructured data that can be used to support risk assessment recommended by guidelines. The large volume of clinical notes is a major barrier to clinicians retrieving information relevant for ADRB risk assessment in a timely manner, especially in busy primary care settings. These findings suggest that NLP techniques have potential utility to support clinicians in evaluating patients for risk of ADRBs during risk assessment prior to considering long-term opioid therapy or as part of ongoing monitoring of COAT patients.
This study used an innovative approach for both identifying selected ADRBs and scoring an opioid risk assessment tool using information from both structured and unstructured EHR sources. NLP is essential for consistently using free-text clinical information as a basis for clinical decision support. Moreover, the use of NLP directly responds to the expert panel recommendations for incorporating clinical decision support tools and the use of clinical data to better identify patients who may benefit or be harmed by opioid use.