What does it mean to provide decision support to a responsible and competent expert?

Abstract

Decision support consists in helping a decision-maker to improve his/her decisions. However, clients requesting decision support are often themselves experts and are often taken by third parties and/or the general public to be responsible for the decisions they make. This predicament raises complex challenges for decision analysts, who have to avoid infringing upon the expertise and responsibility of the decision-maker. The case of diagnosis decision support in healthcare contexts is particularly illustrative. To support clinicians in their work and minimize the risk of medical error, various decision support systems have been developed, as part of information systems that are now ubiquitous in healthcare contexts. To develop, in collaboration with the hospitals of Lyon, a diagnostic decision support system for day-to-day customary consultations, we propose in this paper a critical analysis of current approaches to diagnostic decision support, which mainly consist in providing them with guidelines or even full-fledged diagnosis recommendations. We highlight that the use of such decision support systems by physicians raises responsibility issues, but also that it is at odds with the needs and constraints of customary consultations. We argue that the historical choice to favor guidelines or recommendations to physicians implies a very specific vision of what it means to support physicians, and we argue that the flaws of this vision partially explain why current diagnostic decision support systems are not accepted by physicians in their application to customary situations. Based on this analysis, we propose that decision support to physicians for customary cases should be deployed in an “adjustive” approach, which consists in providing physicians with the data on patients they need, when they need them, during consultations. The rationale articulated in this article has a more general bearing than clinical decision support and bears lessons for decision support activities in other contexts where decision-makers are competent and responsible experts.

Introduction

Decision support is an activity that consists in helping a decision-maker to improve his/her decisions, through a better understanding of the stakes of the decisions, a more thoughtful examination of the relevant data, or/and a more rigorous utilization of relevant theories and practices. Decision support is usually provided upon demand, but clients requesting decision support are often themselves knowledgeable, at least to some extent, about the topic concerning which they ask decision support. Moreover, clients requesting decision support are often taken by third parties and/or the general public to be responsible for the decisions they make. In such cases, the task of the decision analyst (or decision support provider) is delicate in the sense that s/he risks infringing upon the expertise and responsibility of the decision-maker. The case of attempts at providing decision support to physicians in customary consultations is paradigmatic. Physicians are experts in medical matters and they are responsible for the medical decisions they make, but numerous decision support tools are developed in the literature and in practice in hospitals to provide them with decision support. How can one make sure that these tools do not infringe upon physicians’ expertise and responsibility? In this article, we set out to answer this question, based on a literature review and a critical methodological analysis of medical decision support approaches.

The decision support systems that we are about to analyze here are part of the larger set of information systems in healthcare environments, more commonly called Health Information Systems (HISs). HISs have been developed in the last decades mainly to support and improve healthcare processes, decisions, and outcomes of patients. Nowadays, HISs are ubiquitous in hospitals and it is difficult to find a hospital without an information system. One can distinguish, among HISs, different kinds of systems dedicated to healthcare support. According to Shortliffe and Cimino (2014)’s review of computer applications in healthcare, one of the first systems developed in healthcare environments corresponded to systems allowing the recording of healthcare information. These are Electronic Health Records (EHRs), including databases, indexing systems, and research systems using healthcare information. With a similar objective, Computer Physician Order Entry (CPOE) (Kuperman and Gibson 2003) is a system developed to digitize physician’s orders.

Another subset of HISs is composed of clinical decision support systems (CDSSs) (Musen et al. 2014; Berner 2016). CDSSs include all kinds of tools designed to transmit information to clinicians to help them to make decisions or simply to facilitate their daily processes. The main objective of CDSSs is to minimize the risk of medical errors. CDSSs themselves include a variety of systems. Alert Systems provide alert messages to clinicians when an emergency occurs, e.g. when a hospitalized patient undergoes a heart attack. Alert Systems are also integrated into some CPOEs to prevent mistakes in drug prescriptions and/or drug dosages (Van Der Sijs et al. 2006). Reminder Systems (Garg et al. 2005) are likewise developed to avoid omission errors.

Lastly, diagnostic decision support systems (or DDSSs) are a subset of CDSSs dedicated to providing support to physicians in their clinical diagnosis. These systems will be our main topic in the present article. According to a recent systematic survey of DDSSs (Yanase and Triantaphyllou 2019), there are currently two main types of DDSSs:

• Diagnostic decision support systems (DDSSs) based on “gold standard” rules or guidelines defined by experts of the domain or health authorities (thereafter: “Guideline-based DDSSs”).

  Clinical practice guidelines, including diagnostic guidelines, are lists of instructions to follow in a specific situation. They are generally based on current best practices and can be represented by a flowchart. Figure 1 shows an example of a flowchart from the MIMS website1 and based on the guidelines for diabetes treatments produced by the UK’s National Institute for Health and Care Excellence (NICE)2. Other examples of clinical guidelines can be found on the NICE website3, on the website of the French “Haute Autorité de Santé” (HAS)4 or in reports of International Classification of Diseases (World Health Organization et al 1992).

  Guideline-based DDSSs encompass not only “expert systems”, which integrate “gold-standard” flowcharts/rules into their process to produce full-fledged diagnosis recommendations to physicians  (Yanase and Triantaphyllou 2019), but also systems that prescribe to physicians the steps they should follow to abide by the “gold-standard” (this is the case, for example, of the systems found on the NICE website or the Quick Medical Reference (QMR) linked to the INTERNIST expert system (Miller et al. 1986; Miller 2010)).

• Diagnostic decision support systems (DDSSs) based on machine learning (ML) algorithms, or ML-based DDSSs, are used to support diagnoses of specific diseases, with the aim to minimize error rates by treating large amounts of data on patients (Dua et al. 2014; Yanase and Triantaphyllou 2019).

  Machine learning (ML) algorithms are methods used to learn how to approximate a classification function based on a learning dataset. Classification functions could be, for example, functions anticipating the value of an exogenous variable y depending of the value of an endogenous variable x, or functions distinguishing pictures of healthy from pictures of diseased organs by analyzing a matrix of pixels. ML problems are generally divided into three subclasses, depending of the degree of knowledge included in the learning dataset: supervised learning (full knowledge), semi-supervised learning (some pieces of information are not available) and unsupervised learning (no predefined class).

  Many ML algorithms have been proposed to handle these classification problems, from Naive Bayes algorithms to Artificial Neural Networks and Support Vector Machine algorithms. In this paper, we used the term “ML-based DDSSs” to refer to all the DDSSs using one of these ML algorithms.

As we will see in this paper, in their application to support customary diagnostic decisions, these DDSSs are currently in a paradoxical situation. On the one hand, their potential usefulness appears unquestionable, but on the other hand, they are generally poorly accepted by physicians. In addition, the use of DDSSs raises responsibility issues and involves patient safety risks. This paradoxical situation reflects, in our view, the more general difficulty to provide decision support to a competent, responsible decision-maker. By analyzing the specific case of DDSSs for customary consultations in detail, we aim to develop a new approach to address this general difficulty. To that end, we analyze here the reasons underlying the current failure of DDSS, and we draw the constructive lessons from this analysis.

By tackling this issue, this article aims to contribute to a broader research program devoted to analyzing the challenges facing decision support approaches and methodologies, as developed mainly in decision sciences and operational research, when they are applied to decisions involved in the design, implementation, and evaluation of public policies (Tsoukiàs et al. 2013; De Marchi et al. 2016). This research program has already produced applications to the evaluation of environmental policies (Jeanmougin et al. 2017), the design of policy options (Ferretti et al. 2019; Pluchinotta et al. 2018, 2019), the development of methodological tools for large scale environmental policies (Choulak et al. 2019), among others. In the wake of these contributions, we endorse the methodological and epistemological approach clarified in Tsoukiàs et al. (2013), Meinard and Tsoukiàs (2019), Meinard and Cailloux (2020).

Our reasoning unfolds in three steps. In Sect. 2, we begin by reviewing historical choices that led to the current development policy of DDSSs and past experiences in the elaboration of DDSSs. Section 3 explores the adverse impact of HISs, CDSSs, and DDSSs, responsibility issues raised by the use of DDSSs, as well as gaps between DDSSs’ design and the reality of customary consultations, to highlight potential reasons behind the failure of DDSSs in these situations. Section 4 discusses the conceptual approaches underlying the current DDSSs and sets out to determine which approach should be favored in the case of customary consultations. Section 5 briefly concludes the paper.