Abstract
Clinical Decision Support Systems (CDSS) provide aid in clinical decision making and therefore need to replicate human, data interactions, and cognitive functions of decision makers. This paper classifies important considerations of CDSS design into six elements that formulate a high-level reference model (CDSS-RM). These elements are introduced in form of questions, and examined with the use of illustrated scenario-based and data examples. The six considerations are: (i) Do CDSS mimic the cognitive process of clinical decision makers? (ii) Do CDSS provide recommendations with longitudinal insight? (iii) Is the model performance contextually realistic? (iv) Is the ‘Historical Decision’ bias taken into consideration in CDSS design? (v) Do CDSS integrate established clinical standards and protocols? (vi) Do CDSS utilize unstructured data? The CDSS-RM reference model can contribute to optimized design of modeling methodologies, to improve response of health systems to clinical decision making challenges.
Keywords
Clinical decision support system, decision making, reference model, systems design, theoretical framework
Introduction
Proper use of clinical information is especially important in an effort to provide quality health services [1]. Decisions made in hospitals rely on information which once acquired, is further processed by healthcare professionals’ cognitive skills, such as in the differential diagnosis [2]. Combining clinical information and the cognitive element is therefore critical to clinical decision making. Clinical Decision Support Systems (CDSS) provide clinicians with knowledge, intelligently filtered or presented at appropriate times, to enhance health and health care [3]. CDSS consist of an effective pathway to improve patient safety and reduce errors of clinical practice [4]. For instance, use of alerts is an error reduction strategy [5].
The role of medical informatics is important in improving patient-centered care by developing decision support systems. However, development of such systems is a complex task that requires the integration of knowledge from the clinical domain, and decision science to adapt the CDSS to the hospital practice and clinical work flows [6]. The information that CDSS provide needs to reflect the decision-making process and the intellectual effort of clinicians in a contextually relevant way. For cognitive tasks, like the diagnosis decision, the aim of CDSS is to assist, rather than to override clinicians. In recent approaches, such as CONFlexFlow [7], integrating flexible clinical pathways into CDSS was recognized as critical for system success, as, also, a means for better understanding the clinical context through ontologies, in deciding the right rules for a certain activity, to address the needs of a specific case. Thirty years ago, during the eighties and early nineties, there was an open debate on how “recent progress in computer-based diagnosis has been encouraging enough to consider the concept of computer diagnosis” [8]. Nowadays, while this is still an open debate, the impressive progress in Machine Learning and Artificial Intelligence, provides new opportunities for more targeted and accurate clinical predictions and recommendations [9].
Additional considerations for CDSS design include the annotation of any clinical attributes that are commonly used together to drive specific decisions. Pieces of information are often combined by healthcare professionals who arrive at clinical conclusions during patient assessment [10]. CDSS and cannot rely on static, prefabricated ‘in vitro’ methods. Instead, CDSS should make dynamic predictions, allowing interactions with clinicians, for user feedback, and take into consideration the longitudinal nature of repeated physiological properties and the disease progression [11]. To respond to the above challenges, the objective of this paper is to establish a comprehensive theoretical framework for successful decision support systems in healthcare and present a reference model built around six important principles.
While there are available theoretical frameworks such as the Google TITE (Time-Interactions-Trends-Events) [12], outlining important components for decision support systems in general, there are barely any comprehensive efforts specific to CDSS and clinical decision making.
Peripheral work includes a general purpose artificial intelligence framework to address challenges in the modern healthcare system [13]. This framework serves as a simulation environment for exploring various healthcare policies, and payment methodologies, forming the basis for clinical artificial intelligence. Other efforts include the work of Fox et al. [14] who developed the PROforma method, for specifying clinical guidelines and protocols via graphical notation and a formal knowledge representation language. In their paper, they discuss the need for flexible and well understood knowledge representations which are capable of capturing clinical guidelines and protocols for decision support systems.
This paper starts by defining the context that the CDSS-RM reference model is constructed on. Then it outlines four properties of health data that should be taken into consideration during the design of CDSS, to eventually introduce six interrelated principles for successful CDSS, built around the data properties previously discussed. These six principles altogether, form an illustrative reference model for CDSS, that we named CDSS-RM. This reference model has its roots on initial work that was presented at the Petra 2017 International Conference. During that work, we introduced principles. In this paper, we move forward to further develop the considerations and formulate the CDSS-RM reference model.
Definition of the context for the CDSS-RM
Upon patient admission, clinical decision considerations start at the point shown in Figure 1. Clinical decisions include (i) selection of appropriate diagnostic tests, (ii) patient diagnosis, (iii) selection of optimal treatments and (iv) prediction of the patient prognosis This is the decision-making context that the reference model is based upon. These decisions are interdependent and are characterized from five data use patterns summarized below:
1. Ordered steps, each leading to new data, useful for a new decision: Patient history, symptoms, and physical examination contribute to the decisions for ordering diagnostic tests. The test results form the basis for patient diagnosis. The diagnosis, in turn, is decisive for the choice of an optimal treatment.
2. Feedback loops and temporal repetitions: The result of a diagnostic test may direct physicians to order additional tests, as a requirement of successful differential diagnosis (Fig. 1, point 2). In addition, a diagnostic test may be repeated, during a periodic assessment, to confirm or alter the therapeutic schema in response to the updated diagnostic test results (Fig.1, point 6).
3. Combining data: A variety of different data need to be combined in clinical decision making. Examples include the diagnostic test results combined with the patient history, physical examination and symptoms, to make the diagnosis (Fig.1, point 4). Or the combination of lab test results, the diagnosis, patient history and physical examination and treatment, to predict the patient prognosis (Fig. 1, point 7).