Designing an automated clinical decision support system to match clinical practice guidelines for opioid therapy for chronic pain

Promoting use of good care practices is necessary for safe and effective use of opioid therapy for chronic non-cancer pain, but achieving provider adherence to clinical practice guideline (CPG) recommended care practices has proven difficult in most primary health care settings [1–3]. Increased attention to the importance of pain management has led to increased prescribing of analgesic medications [4]. Opioid analgesics are among the most prescribed medications in the US today [5, 6] and, as of 2008, hydrocodone was the top prescribed medication in the country [4]. However, increased use of these powerful and potentially addictive medications has had negative consequences. Rates of opioid overdose, prescription opioid misuse and addiction, diversion of prescribed medications toward illicit use, and opioid-related legal suits against physicians have all increased to a disturbing extent [4, 5]. Use of recommended care practices is considered essential for minimizing these negative consequences without reversing gains made in improving pain management in clinical settings.

In 2003, the Veterans Administration (VA)/Department of Defense (DOD) published a CPG for use of opioid therapy for the treatment of chronic non-cancer pain [7]. The goals included using evidence-based recommendations to improve analgesia, promote uniformity of care, and decrease related morbidity of patients with non-cancer chronic pain in the primary care setting. This guideline provides detailed information about appropriate dosing, including protocols for initiation, titration, and cessation of the most commonly used opioid medications. It provides information on potential contraindications for opioid therapy for chronic pain and suggestions for opioid management in patients at higher risk of misuse, diversion, adverse effects, overdose, and/or lack of efficacy. A substantial portion of the guideline focuses on processes of care. For example, the guideline encourages clinicians to: regularly conduct assessments of pain and functioning; use urine drug screening protocols to discourage and detect medication misuse and diversion; obtain written agreement on the parameters and responsibilities of the patient regarding the opioid prescription; provide clear education on both the risks and realistic level of benefit from opioid analgesics; and carefully document and follow treatment plans. This framework can increase clinician's confidence in appropriately prescribing opioid therapy.

Despite expert consensus on the importance of adherence to these care guidelines, there is little evidence that they are consistently followed in actual clinical practice [8]. Numerous barriers to providing guideline-adherent care exist [9]. Clinicians report lack of training in both pain management and addiction medicine and are uncomfortable assessing and treating these conditions. Moreover, patient-provider communication about opioids is complicated by: the subjective nature of pain experience, which prevents physicians from objectively verifying the severity of the pain condition; the reinforcing effects of opioid drugs, which may lead to either deliberate or unknowing attempts by the patient to obtain opioid medications; provider and patient fears about the consequences of either prescribing a potentially addictive medication or under-managing pain; and stigma associated with substance use disorders [10–12]. Because of these communication difficulties, providers may be hesitant to prescribe opioid medications initially, to discontinue medication when there is no clear sign of benefit, and to address the addictive nature of opioid analgesics and the possibility of misuse. In all cases, these behaviors lead to suboptimal care. Additionally, poor care coordination within the health care system contributes to poor opioid management [13]. Lack of clear documentation of pain management plans and opioid use agreements and lack of communication between providers can lead to inconsistent treatment and poor prescribing decisions that contribute to misuse and poor pain management. Lastly, good care practices take time, and time limitations and competing demands during outpatient visits in primary care may limit clinician adherence to guidelines.

Developing health services interventions that address these barriers is essential for improving opioid management in chronic pain. A computerized decision support system (DSS) may provide such an intervention [14, 15], and some DSSs have been shown to increase adherence to guideline recommended care [16]. Hunt and colleagues systematically reviewed randomized controlled trials of DSSs, defined as 'any electronic or non-electronic system designed to aid directly in clinical decision making, in which characteristics of individual patients are used to generate patient-specific assessments or recommendations that are then presented to clinicians for consideration' [17]. Kawamoto and colleagues identified features that were independently associated with improved clinical practice in a multiple regression analysis. These included: automatic delivery, presentation of the DSS when and where clinical decision making occurs, provision of concrete recommendations of how to proceed, and computer-based generation of decision support [18]. A model computerized DSS (ATHENA-DSS) that links with the electronic medical record (EMR) system used by the VA Health Care System (VistA) was designed to provide these key features [19–22]. ATHENA-DSS, developed using the EON guideline decision-support technology [23, 24], accesses patient information in the EMR, evaluates this information in terms of a knowledge base consisting of encoded CPG recommendations, generates patient-specific recommendations, and presents a graphical user interface with these recommendations along with tools and information support to clinicians when they open the EMR of a relevant patient at the time of the clinic visit.

We used an iterative development process involving authors of the CPG, local content experts, end-users (i.e., opioid prescribers), knowledge modelers, graphic designers, and systems software engineers to modify the initial ATHENA-DSS, ATHENA-Hypertension (HTN), to guide evidence-based opioid prescribing (Figure 1). We named this newly developed system ATHENA-Opioid Therapy (ATHENA-OT) [25]. Here, we describe the process and outcome of this iterative development via which we operationalized CPG information into a computer-interpretable knowledge base to provide patient-specific recommendations for care, and clinical tools to encourage good care practices in opioid prescribing. Iterative usability testing was also a crucial component of ATHENA-OT development, but these processes will be described elsewhere [26]. A valuable part of our process is collaboration of the DSS developers directly with the CPG authors. The process described provides a model for translating guidelines into DSSs, including methods to ensure that the DSS retains the intent of the CPG authors and encourages use of good care practices through inclusion of patient-specific recommendations and clinical tools.

The patient safety features and a thorough description of the ATHENA-OT graphical user interface have been described previously [25]. This study was approved and overseen by the Stanford University Human Research Protection Program and the VA Palo Alto Health Care System Research and Development Committee.

In the process described, the team started with the CPG and translated it into three primary products: an operationalized algorithm in Protégé/EON, a matching written Rules Document, plus a set of clinical tools (Figure 1). We based our guideline translation process on experience gained in development of ATHENA-Hypertension as well as general principles from medical informatics literature encouraging iterative design based on interim evaluation and testing (for example, the ADDIE (Analysis, Design, Development, Implementation and Evaluation) process [27]. The accuracy testing procedures for both the Protégé/EON algorithm and the clinical tools were adapted from those initially designed and successfully used by the ATHENA-Hypertension development team [28]. The Rules Document validation process was designed for ATHENA-OT and has not been previously described, and thus we report this process and findings in greater detail.

A knowledge management team (KM) consisting of the study managers, knowledge modelers (SBM and MM, medical informaticists with expertise in translation of clinical knowledge into encoded computer-interpretable formats using a knowledge acquisition program called Protégé [29]), and system software experts drafted, revised, and managed the review of these 3 products. Each of these three products were reviewed and revised through separate procedures and distinct, but overlapping teams. Revisions to the Protégé/EON algorithm and the Rules Document were made in tandem to maintain consistency, based on feedback from the accuracy and rules validation testing. These processes occurred iteratively during ATHENA-OT development. Each of the processes, as well as major revisions, are described below.

Collaboration on an iterative design process between the ATHENA-OT developers, local content experts, and the authors of the CPG for management of opioid therapy for chronic pain identified issues that arise in encoding a DSS and provided a mechanism for their resolution, resulting in an automated DSS that fulfils the intentions of the CPG authors.

Clinicians who provide care for patients based on guidelines must operationalize them to carry out care. Operationalizing CPGs for automated DSS highlights the context, assumptions, ambiguities, and gaps that are inherent in the usual formats of CPGs [30]. Encoding the DSS requires interpreting the context, specifying concepts, clarifying embedded assumptions, spanning gaps, and resolving ambiguities in source documents. Evidence sources for recommendations can sometimes serve as sources for specification of concepts (e.g., which specific diagnoses were included in the original study that forms the evidence base for a particular recommendation) or target population assumptions, but do not fill all the gaps. Consensus-based recommendations may present a particular challenge for translation into computerized DSS. These recommendations typically do not reference scientific studies on which definitions of concepts or cut-points for decisions could be based [30]. Without participation of the CPG authors, who have extensive expert knowledge beyond what is written into the CPG document, the intent or specificity of the guideline may be altered as it is translated into a computerized DSS. In developing ATHENA-OT, we used a two-phase process including accuracy testing plus a review of the guideline rules by three authors of the CPG, followed by a final round of telephone and email communications to arrive at a final version meeting approval of the CPG authors.

This process clarified details of the recommendations and better specified patient populations for whom recommendations were relevant. However, the process was made more challenging by the consensus nature of the guidelines. There was not perfect agreement among the CPG authors for all recommendations. Moreover, in some cases, it seemed apparent that some vague areas of the guideline were purposefully written to be vague, an option that could not be directly incorporated in the computerized algorithm beyond careful wording of the recommendations. Encoding the guideline into a DSS revealed areas of the guideline that lacked specificity and therefore might be difficult to implement both in the DSS and clinically. We note that the opioid therapy guidelines may have been more challenging to operationalize than other guidelines in this regard, due to the extreme heterogeneity of patients and underlying health conditions, limited evidence-base for many criteria, and controversies surrounding treatment.

Operationalizing the opioid therapy for chronic pain guideline was further complicated by the fact that good opioid prescribing decisions require consideration of behavioral, mental health, and psychosocial conditions and functioning. Moreover, patient preferences and goals necessarily influence decisions about use of opioid therapy. Prescribing decisions involve striking a balance between pain control, adverse-effect management, physical and emotional function, and addiction risk. This information can be efficiently obtained only in discussion between patient and provider; there is little to no computable information recorded in the medical record to indicate biopsychosocial functioning and patient preferences to guide such decisions. What information is available in the medical record is typically contained in free-text notes rather than structured data fields and would require advanced text mining algorithms to access, an option that may be available in the future. To design a functional DSS, the expert team was required to consider how the algorithm could usefully support prescribing decisions while not having access to crucial information required to guide opioid prescribing. To overcome these limitations, the expert team designed a DSS that would highlight important clinical information available in the medical record, provide many tools to support shared decision making, facilitate appropriate documentation, and present a checklist of important items to review when considering opioid therapy.

An issue that led to difficulties in getting expert consensus was the problem of differentiating between accurate information and information that was of high clinical priority. There was disagreement among the expert team and CPG authors about whether all accurate information should be provided by the DSS, or whether messages should be limited to issues that were clearly important enough to warrant taking primary care clinicians' time. The process of obtaining consensus on the system could be streamlined by clarifying whether the goal of rules development was to identify all clinically accurate recommendations that could be made from EMR data, or to identify important clinical recommendations for typical primary care practice. The process of analysing the encoded guideline rules presented the guideline information in a different way to guideline authors leading to new insight on their part. This insight made explicit some implicit assumptions by showing varying ways the guideline could be interpreted and applied, which may not have been the intention of the guideline authors.

The wording of recommendations and design of clinical tools evolved alongside the operationalization for the guideline. Striking a balance between detail, accuracy, and clinical utility was difficult, with disagreements among the diverse members of the development and evaluation team leading to dynamic changes in the content of recommendations in terms of prioritization, tone, and wordiness. The CPG suggested tools that would facilitate guideline recommended care processes, but expert and end-user input was needed to optimize content and design. For example, based on feedback from clinician members of the team and usability testing [26], we implemented a write back capacity of the pain assessment template.

An iterative process of drafting, testing, reviewing, and revising the DSS content enabled us to develop a DSS that usefully operationalized the written CPG for opioid therapy for chronic pain into a system that could provide real-time decision support in clinical settings. Including guideline authors alongside the local expert team in the iterative development of the content resulted in a product that was considered consistent with the intent of the guideline and amenable to implementation with the local EMR and patient population. The process and experiences described here provide a model for development of the content other DSSs attempting to translate written CPGs into actionable, real-time clinical recommendations and tools.