What are three types of clinical decision support systems?

Using openEHR GDL for computerised guidelines

Clinical decision support systems (CDSSs) are computer-based programs that provide evidence-based prompts and reminders to help caregivers improve the quality of care [17]. In a CDSS, computerised guidelines (CGs) are developed to formalise the guideline knowledge in a machine-interpretable way [18]. Computerised guidelines are mostly represented as production rules in the form of ‘IF conditions TEHN actions’. Many efforts, such as Ardern Syntax, EON, GLIF, Asbru, Gaston, and SAGE, have been proposed and validated in the past decades [19]. While these efforts have been proposed for decades, the interoperability issue still hinders the wide adoption of the CDSS. The interoperability issue in the CDSS domain is also known as the ‘curly braces problem’ [20]. The ‘curly braces problem’ refers to using nonstandard site-specific information models in CG. Nonstandard site-specific information models cannot be easily shared among organisation and therefore hampers the shareability of CGs.

It would be helpful if there is an approach that can standardise the input and output of CG and leverage existing resources. OpenEHR has such potential. From the technology point of view, openEHR aims to facilitate the interoperability between information systems. The openEHR archetype provides a standard information model that can be shared among organisations so that the ‘curly braces problem’ can be avoided. While from the domain knowledge perspective, it aims to bring informaticists and medical specialists together. Specifically, openEHR uses the archetype to capture the detailed and domain-specific clinical concepts that are modelled by clinical specialists.

The Guideline Definition Language (GDL) has been proposed by the openEHR community to facilitate using openEHR archetypes to author CIGs [21]. Recently, GDL has been upgraded to its second major version, known as GDL2. GDL improves the shareability of encoded CIGs to be shared among organisations cross-border [22].

This section briefly introduces the basic notion of GDL and the methodology of applying GDL for CG authoring. A case study of using GDL to capture the guideline knowledge of COVID-19 has been introduced subsequently.

Clinical Decision Support Systems

CDSSs are integrated analysis and deployment systems designed to facilitate decision-making in patient health care. They combine information about the current patients with information about past diagnoses and treatments stored in a database to provide feedback or recommendations that will aid in decision-making process at the point of care. The Healthcare Information and Management Systems Society (HIMSS) expands this definition to include patients as recipients of information, to permit patients to be active participants in their care. The definition of clinical decision support according to the HIMSS is:

a process for enhancing health-related decisions and actions with pertinent, organized clinical knowledge and patient information to improve health and healthcare delivery. Information recipients can include patients, clinicians and others involved in patient care delivery; information delivered can include general clinical knowledge and guidance, intelligently processed patient data, or a mixture of both; and information delivery formats can be drawn from a rich palette of options that includes data and order entry facilitators, filtered data displays, reference information, alerts, and others.

(www.himss.org/library/clinical-decision-support).

Therefore, CDSSs and HIMMS represent alternate expressions of translational biomedical informatics, which takes the results of scientific research to the bedside, to directly impact patient care.

CDSSs are separated by Plato’s Problem; the gap between knowledge and experience. Clinical knowledge is a cognitive understanding of a set of known clinical rules and principles, based on medical literature, which guide our decision-making processes. Experience is an acquired understanding of medical outcomes gained through years of practice applying various outcomes related to various particular conditions, the majority of which cannot be learned sufficiently through reading and acquiring cognitive knowledge. This important distinction arises because there is an immense number of medical subjects in the literature that could be researched and taught. In addition, there are so many variables in the medical decision-making process that outcomes based on knowledge versus those based on experience are often discordant. It is practicably impossible to teach physicians all of the knowledge acquired by experience, because the environmental variables are constantly changing so the body and nature of our experiences evolve through time, reflecting particular outcomes under specific conditions. Cognitive knowledge, however, is always associated with a limited scope of outcomes that are out of date, by necessity – there is a time-lag between subject outcomes and the reporting of them. For example, this distinction could come sharply into focus when choosing a physician to remove your kidney in surgery: would you choose one who has mastered a surgical textbook, or one who has the experience of 300 successful operations?

CDSSs can be classified into two types of systems:

Knowledge-based support systems that are defined by a well-established set of rules that guide decisions, based on the interpretation of the medical conditions judged in the medical literature to be the best practice.

Non-knowledge based systems that do not use a set of defined a priori rules, but instead use artificial intelligence algorithms to induce the rules through machine learning methods, allowing the system to learn from hundreds or even thousands of encounters, rebuilding the “model” set of rules as environmental variables change. These systems can be based on neural networks, genetic algorithms, support vector machines, decision trees, or any other machine learning technology, which “learns” to recognize patterns in data sets case by case.

Hybrid CDSSs

Hybrid CDSSs have been developed to allow the end user to synthesize the results from both knowledge and clinical experience, and make a clinical decision based on the results of both. (examples in the literature include Santelices et al., 2010).

In such a hybrid system, multiple predicted outcomes are posed for the physician, based on data input from knowledge and experience bases, and furnished with associated probabilities to permit the selection of the appropriate decision. As we continue to learn more about cognitive science, and distil this knowledge into principles, we can apply them to improve these “intelligent” systems to help us make the best clinical decisions possible at the time. This practice of continuous incorporation of patient data, cognitive knowledge, and clinical experience is referred to often as “rapid learning.” Rapid learning approaches that continuously update the CDSS as new data become available provide an ability to create decision models that adapt to the availability of new treatments, interventions, and metrics (variables) that can be input to the modeling process. This paradigm is shown in Figure 3.2.

Figure 3.2. Illustration of the pathways in a hybrid CDSS (Clinical Decision Support System); these have been developed to allow the end user to synthesize the results from both knowledge and clinical experience, and make a clinical decision based on the results of both. Copyright © 2013 Nephi Walton.

Many CDSSs provide information on drug interactions and can generate allergy alerts. These alerts, however, are very basic, and do not include any information on many other factors, such as dose, time of administration, and the context in which the medications are given. Consequently, many physicians discount these warnings. On a given work day, it is very common for a physician to dismiss dozens of these warnings, as he or she prescribes medications in the hospital. In some instances physicians become so used to ignoring these warnings that they may accidentally disregard an important one. It is time to make these alerts more “intelligent,” by using predictive analytics to predict levels at which problems occur, and to set thresholds to control when alerts will be generated. In addition, these alerts should provide information about the effectiveness of the drug for the given clinical scenario, and suggest more effective options, if a suboptimal treatment is selected. Such a system could include analysis of a patient’s antibiotic prescription history before presenting a list of drugs for choice, or checking its database for any information about the susceptibility of the patient to a bacterial invasion if the chosen antibiotic does not provide broad enough coverage.

1.7.5 Clinical Decision Support

CDS systems can be defined as “HIT applications that relate individual patient health data to established knowledge bases and thereby assist in clinical decision making and health management” [70]. Providing personalized insights to oncologists with CDS is one of the primary objectives of CancerLinQ. There are countless areas in oncology where data are needed to make clinical decisions but are largely absent. As the number of new molecular diagnostic tests proliferates, propelled in part by the explosion of cancer panomics, and the number of therapeutic agents expands proportionally, it will become increasingly challenging to practice high-quality cancer medicine without computerized CDS.

At the time of publication of this text, a specific CDS solution has not been determined. It is anticipated that the ASCO clinical practice guidelines [71], derived from high-quality evidence by expert ASCO panels and addressing disease-specific clinical situations or modalities, such as tests or procedures, may serve as a robust source of content knowledge for the development of CDS in CancerLinQ. CancerLinQ will include a feature that will identify relevant guidelines to present CDS that is relevant to the characteristics of an individual patient to assist physicians in treating specific patients. Furthermore, the clinical insights gained by oncologists from the redacted data sets in the system may be used as feedback to inform the refinement of existing guidelines and suggest new clinical areas where guidelines may be necessary.

CancerLinQ will also include observational CDS, which will enable a physician to query the system about a specific case to learn about the treatments and outcomes for similar patients. CancerLinQ will match a particular patient’s data against the experience of similar cancer patients whose data reside within the system and provide the oncologist with aggregate information pertinent to the possible treatments available to the patient and the range of outcomes experienced by other similar patients. Boxes 1.1 and 1.2 provide representative use cases of the CDS capabilities envisioned for CancerLinQ.

Box 1.1

Personalized Medicine CDS Use Case

Phillip Alk is a 72-year-old white male who presents to his primary care physician with nonproductive cough and pleuritic chest pain. He is otherwise healthy except for high cholesterol, for which he takes atorvastatin, mild hypertension, and gastroesophageal reflux treated with omeprazole. He has no fever or weight loss and a Karnofsky performance status (PS) of 90. Mr Alk smoked cigarettes for a few years during college but has not smoked in the last 50 years. A routine chest X-ray reveals a 3 cm nodule in the right middle lobe (RML) confirmed by follow-up chest computerized tomography scan that also demonstrates scattered 1 cm nodules in the left lung and ground glass opacities in the right lower lobe. Routine blood chemistries are normal except for a serum creatinine of 1.8 mg/dL. A core needle biopsy of the RML nodule is performed and reveals malignancy consistent with adenocarcinoma. Upon receiving the diagnosis, the patient is referred to a local oncologist for discussion of treatment options.

Dr Gene Profile is a practicing general oncologist in a group of six oncologists. He has been in practice for 20 years. His practice was a participant in ASCO’s QOPI program for the last 8 years and transitioned to participation in CancerLinQ when offered the opportunity to do so. The practice’s QOPI reports revealed high adherence with measures of adjuvant chemotherapy use and appropriate antiemetic use but also demonstrated need for improvement in chemotherapy use in the last 2 weeks of life and counseling patients about smoking cessation. During his initial visit with Mr Alk, Dr Profile reviews the pathology reports and wonders if the diagnosis of adenocarcinoma reflects a primary lung cancer or metastases from another organ. The patient’s medical history and review of systems do not suggest other problems. A positron emission tomography (PET) scan is performed for staging purposes and demonstrates the RML lung nodule as well as several smaller nodules scattered throughout the left lung. No other areas of increased fluorodeoxyglucose (FDG) uptake are noted. Dr Profile recalls seeing some literature in the office from a lab that offers gene expression profiling to assess carcinomas of unknown primary. He considers ordering the test but first consults the CancerLinQ clinical decision support system. He immediately receives a link to an ASCO guideline on carcinoma of unknown primary that recommends against use of commercially available RNA expression profiling panels due to lack of evidence of clinical utility in helping to establish a more definitive diagnosis. Dr Profile ultimately concludes that primary adenocarcinoma of the lung is the most likely diagnosis and records that diagnosis in the patient’s EMR. He immediately receives a guidance notification from CancerLinQ recommending a standard molecular work-up of the tumor to include testing for sensitizing and resistance mutations to epidermal growth factor receptor (EGFR) inhibitors, EML4-ALK translocation, ROS1 and KRAS mutation. The notification includes a link to ASCO’s provisional clinical opinion (PCO) on integration of palliative care into standard oncology care as well as a link to ASCO’s PCO on EGFR mutation testing for patients with nonsmall cell lung cancer. Dr Profile orders the recommended tests to be performed on the patient’s tumor biopsy. Mr Alk is anxious to begin treatment but is advised to wait for the molecular test results to see if they impact Dr Profile’s treatment recommendation. Two weeks later, Dr Profile receives a report from a commercial laboratory of a 400 gene mutation panel that was performed on the biopsy. His hospital’s pathology department had determined that it is more cost-effective to obtain the multigene panel rather than the individual tests for the mutational analysis recommended by CancerLinQ. The report describes an L858R mutation in exon 21 of the EGFR gene, wild type KRAS, mutation in the p53 gene, mutation at the V600E locus of the BRAF gene, as well as alterations in other genes. Uncertain of how to proceed, Dr Profile uses the CancerLinQ physician portal to submit the molecular profiling test results and a brief case summary to the ASCO University Molecular Oncology Tumor Board. Two days later, he receives an explanation of the molecular profiling results that confirm his plan to prescribe an EGFR inhibitor drug for Mr Alk. Upon entering the order for erlotinib in the patient’s EHR, Dr Profile receives a notification from CancerLinQ of a possible drug interaction with omeprazole and is advised to discontinue that medication. He is also advised to carefully monitor the patient’s renal function in view of the toxicity profile of erlotinib and the patient’s preexisting renal insufficiency.

After 3 months on treatment, Mr Alk’s cough and chest pain have fully resolved and his CT scan reveals complete resolution of all pulmonary nodules. He continues on treatment but 3 months later notices a return of his cough. When these new symptoms are recorded in his EHR, Dr Profile receives a message from CancerLinQ advising him to consider erlotinib-induced interstitial lung disease as a possible complication of treatment. A repeat CT scan reveals new pulmonary nodules consistent with recurrent and progressive lung cancer. Dr Profile recalls that the molecular testing of Mr Alk’s tumor had revealed a BRAF V600E mutation. He wonders if a new BRAF inhibitor, recently approved to treat melanoma, could possibly benefit his patient. He uses CancerLinQ’s observational clinical decision support function to query the CancerLinQ database and receives a report of 30 patients in the system with adenocarcinoma of the lung and a BRAF V600E mutation. Of these, 20 had been treated with vemurafenib and 10 had experienced significant tumor shrinkage. Dr Profile prescribes off-label vemurafenib for his patient. Within 2 months, Mr Alk’s symptoms have worsened, his performance status has declined, and his cancer has progressed on CT scan. He consults Dr Profile about other treatment options, particularly whether he should give chemotherapy a try. He is advised to transition to hospice care at this point in his illness.

Box 1.2

Observational CDS Use Case

Mr Jones is a 72-year-old African American man with advanced prostate cancer whose disease has been well controlled with conventional antiandrogen treatment. However, during the past month he has developed increased skeletal pain, rising prostate-specific antigen (PSA), and new visceral metastases detected on CT scan. Mr Jones consults his oncologist about treatment options and receives a recommendation to change treatment to abiraterone acetate, a recently approved CYP17 inhibitor that inhibits the extragonadal conversion of pregnenolone to testosterone, a pathway frequently exploited by castrate-resistant prostate cancer. Abiraterone was approved by the FDA in 2011 based on a randomized clinical trial that compared abiraterone plus prednisone to prednisone alone and demonstrated an improvement in median overall survival from 10.7 to 14.8 months. Mr Jones is inclined to proceed with this new treatment until his oncologist tells him that the cost of treatment is about $6000/month. Mr Jones is concerned that he cannot afford the expected copayment of $1200/month and will have to forego further treatment. His oncologist recalls that the antifungal drug ketoconazole inhibits the same enzyme as abiraterone and has been reported in several small clinical trials to have activity in advanced prostate cancer. He explains to Mr Jones that ketoconazole was commonly used off-label to treat patients with castrate-resistant prostate cancer until the FDA approval of abiraterone. A generic drug, ketoconazole is considerably less expensive than abiraterone, costing approximately $500/month. No clinical trials have ever been performed that directly compare the effectiveness of ketoconazole and abiraterone. Mr Jones wonders if ketoconazole might be an acceptable alternative to abiraterone in his case and one that he could much more easily afford.

Mr Jones’ oncologist is a participant in ASCO’s CancerLinQ learning health care system. Using the observational clinical decision support function of CancerLinQ, he enters Mr Jones’ characteristics into the system and requests a report on treatment administered to similar patients and the outcomes achieved. Minutes later, CancerLinQ provides a report on several thousand similar patients in the system. Of these, 15% had received treatment with ketoconazole and 25% treatment with abiraterone over the past 4 years. The patients in each group experienced similar survivals of about 12 months, although patients receiving ketoconazole experienced higher rates of fatigue and nausea. Upon learning of these data, Mr Jones opts for treatment with ketoconazole and receives a prescription from his oncologist to begin immediately.

Clinical Decision Support Systems

To provide the best healthcare to their patients, health professionals need updated, evidence-based information [17]. Electronic resources that present relevant medical information at the point of care, integrating patient-specific data, are known as clinical decision support systems (CDSS). CDSS can be defined as “software that is designed to be a direct aid to clinical decision-making in which the characteristics of an individual patient are matched to a computerized clinical knowledge base, and patient-specific assessments or recommendations are then presented to the clinician and/or the patient for a decision” [18].

CDSS have become increasingly useful for health professionals and patients to access filtered information (Fig. 19.2). These systems should help close the gap between optimal practice and actual clinical care [23,24]. They are designed to assist the health professional-patient encounter at multiple points (from initial patient-clinician interaction to diagnosis, treatment, and follow-up) and are expected to significantly enhance both the quality and safety of healthcare at all levels [19,21,25]. Moreover, they can be seen as a pathway to address some of the inefficiencies of healthcare by rationalizing the use of diagnostic tests, treatments, and referrals [19].

CDSS encompasses a variety of tools [16,19,21], including computerized alerts, reminders or visual summaries of complex data, documentation templates, and contextually relevant reference information, among others (Fig. 19.2). They have been implemented in several clinical areas, such as pharmacology and pathology [25], and have been used in the management of various chronic conditions, such as diabetes, hypertension, and dyslipidemia [21].

One of the core components of CDSS is patient data, which includes specific data related to the patient for which the decision will be made. Also, assessing and integrating the patients’ perspective at the time of the clinical encounter can enable health professionals to adjust their management meaningfully. Indeed, clinical decisions such as treatment change and screening for the need of additional medical services (e.g., psychologic care) may be based on patient-reported data [26]. This is also true outside of healthcare facilities and clinical visits, e.g., when patients are at home or in the community, where patient-reported data can support patients to optimize their self-management. In fact, between each clinical encounter, it is the patient who makes health-related decisions, such as to follow or not the agreed treatment plan.

Assessment and monitoring of patients’ preferences, health, well-being, and behavior have the potential to inform the management of chronic diseases at the point of care [27]. Using patient-reported outcome measures (PROM), which are health measurements reported by patients about themselves and include (but are not limited to) symptoms, functioning, health perceptions and health-related quality of life [28], is one of the ways of assessing the patient’s perspective. However, collecting PROM at the point of care with the conventional paper-and-pencil method of administration can be time-consuming and hinder its use at the time and place (e.g., clinical encounter; home) of patient healthcare. Therefore, PROM are being successfully integrated into real-time CDSS [29,30].

The widespread implementation of CDSS in clinical practice has been delayed by both the difficulties inherent to the creation of an adequate CDSS and the acceptability of its use within the clinical workflow [22,25]. Nevertheless, published studies of CDSS are increasing rapidly, and their quality is improving. A systematic review of 70 randomized controlled trials identified that CDSS with (1) decision support provided automatically as part of clinician workflow, (2) decision support delivered at the time and location of decision making, and (3) actionable recommendations provided, are strongly associated with clinical practice improvement [31]. These favorable features should be considered in the improvement of current CDSS as well as in the design of new ones.

18.1 Introduction

A clinical decision support (CDS) system is a computer-based system that analyzes available data to guide people through a clinical decision-making process. The availability of data may be considered to be the most fundamental prerequisite of a CDS, because analysis and guidance depend on it. Its usefulness further depends on the data being well-structured and unambiguous. Coding of data items is essential in order to understand and be able to manipulate them. Chapter 17 reviewed approaches to standardizing the terminology and information models used for data items. The latter includes the need for the data type of each item to be specified, including units for quantifiable data types or categorical values for dictionary/directory-based data types. Lack of ambiguity of these aspects of a data item is essential when data are communicated between the user and the computer.

Two frequently occurring tasks in the clinical workflow where the health care provider and the computer communicate are during clinical documentation (both in structured data input and in report production) and in computerized-provider order entry (CPOE). These tasks are facilitated by grouped knowledge elements, i.e. structured documentation templates and order sets respectively. A structured documentation template is an organized collection of data items relevant to a particular clinical context (e.g. patient presenting with a headache) that can be used to collect or present codified information about a patient. An order set, similarly, is an organized collection of actions that can be ordered by a health care provider for the care of a patient in a specific clinical context (e.g. orders for a patient being admitted with stroke).

These grouped knowledge elements may be considered as vehicles for providing clinical decision support. In a passive sense, they remind the health care professionals about questions to ask the patient, specific observations to be made during an examination, tests to be ordered, or medications to be administered that are relevant to that clinical context. Documentation templates that display information can organize information in such a way as to play a mnemonic function, by facilitating identification of important data, recognition of trends, or making other associations. More actively, those who design the groupings of knowledge elements can use them to drive the behavior of the health care professional user. For example, in an order set, an intervention that is known to be more effective, per current evidence, can be made easier to order than alternatives by having that intervention be selected by default. Additionally, order items and documentation items can be dynamically presented on the basis of the context, e.g. a documentation template can suggest asking about past history of rheumatic fever only if a murmur is noted as having been heard on auscultation. By anticipating needs for data entry or access, or for orders, such grouping not only provides CDS but also facilitates workflow, by eliminating extra steps that would otherwise be needed.

The specification of an order set’s or a document template’s structure and content is a form of knowledge. Standards are being developed for such specification to encourage the collection of higher quality, more interpretable, more comprehensive data, and to encourage reuse of document specifications, or parts thereof, where appropriate. This chapter will review those efforts, in terms of their degree of maturity and harmonization, and how they relate to clinical decision support.

The management of a collection of grouped knowledge element specifications has not received much attention in the clinical informatics and standards development communities. This is a knowledge management (KM) task, the purpose of which is to reconcile collections of document specifications in an enterprise, and encourage convergence on and reuse of those that foster best practices and conformance with enterprise goals. Knowledge management for grouped knowledge elements is only recently being recognized as an important challenge, as KM systems begin to be introduced into health care enterprises to manage their knowledge content. Examples of approaches to curating documentation specifications, and supporting authors and editors in creating and updating them, are discussed in Chapter 28.