G Model

ARTICLE IN PRESS

ARTMED-1365; No. of Pages 10

Artificial Intelligence in Medicine xxx (2014) xxx–xxx

Contents lists available at ScienceDirect

Artificial Intelligence in Medicine journal homepage: www.elsevier.com/locate/aiim

From decision to shared-decision: Introducing patients’ preferences into clinical decision analysis Lucia Sacchi a,∗ , Stefania Rubrichi a , Carla Rognoni a,b , Silvia Panzarasa a , Enea Parimbelli a , Andrea Mazzanti c , Carlo Napolitano c , Silvia G. Priori c,d , Silvana Quaglini a a

Department of Electrical, Computer and Biomedical Engineering, University of Pavia, Via Ferrata 5, 27100 Pavia, Italy Centre for Research on Health and Social Care Management, Bocconi University, Via Roentgen 1, 20136 Milano, Italy c Molecular Cardiology Laboratories, IRCCS Fondazione Salvatore Maugeri, Via Salvatore Maugeri, 8-10, 27100 Pavia, Italy d Department of Molecular Medicine, University of Pavia, Via Forlanini, 6, 27100 Pavia, Italy b

a r t i c l e

i n f o

Keywords: Shared decision-making Decision trees Patient preferences Utility coefficients Atrial fibrillation

a b s t r a c t Objective: Taking into account patients’ preferences has become an essential requirement in health decision-making. Even in evidence-based settings where directions are summarized into clinical practice guidelines, there might exist situations where it is important for the care provider to involve the patient in the decision. In this paper we propose a unified framework to promote the shift from a traditional, physician-centered, clinical decision process to a more personalized, patient-oriented shared decision-making (SDM) environment. Methods: We present the theoretical, technological and architectural aspects of a framework that encapsulates decision models and instruments to elicit patients’ preferences into a single tool, thus enabling physicians to exploit evidence-based medicine and shared decision-making in the same encounter. Results: We show the implementation of the framework in a specific case study related to the prevention and management of the risk of thromboembolism in atrial fibrillation. We describe the underlying decision model and how this can be personalized according to patients’ preferences. The application of the framework is tested through a pilot clinical evaluation study carried out on 20 patients at the Rehabilitation Cardiology Unit at the IRCCS Fondazione Salvatore Maugeri hospital (Pavia, Italy). The results point out the importance of running personalized decision models, which can substantially differ from models quantified with population coefficients. Conclusions: This study shows that the tool is potentially able to overcome some of the main barriers perceived by physicians in the adoption of SDM. In parallel, the development of the framework increases the involvement of patients in the process of care focusing on the centrality of individual patients. © 2014 Published by Elsevier B.V.

1. Introduction Patients’ preferences are progressively emerging as an essential requirement in health decision-making [1,2]. Over the past three decades, the proportion of patients willing, and asking, to be involved in clinical decisions during encounters with their physicians has been constantly growing and accounts now for the majority of the patients [3]. This changing attitude does not only include the need for information but also the need to consider personal preferences as an essential part of medical interventions. In the perspective of providing patient-centric care

∗ Corresponding author. Tel.: +39 0382985981. E-mail address: [email protected] (L. Sacchi).

[4], attention is focused on addressing individual attitudes, considering a patient’s perception of his/her health conditions, personal context, job-related requirements and economic conditions. This new trend is also reflected by the new Horizon2020EU calls for eHealth projects, which address personalized medicine as the central topic of the “Personalizing Health and Care” focus area (http://ec.europa.eu/research/participants/portal/desktop/en/ opportunities/h2020/index.html, accessed 07 October 2014). Clinical decision analysis refers to the systematic exploitation of a decision-theoretic model to evaluate the choice between two or more alternatives [5]. As an example, alternatives may concern the choice between two similarly effective pharmacological treatments, between a surgical intervention and a drug, or deciding whether to undergo cancer preventive screening that requires an invasive examination. As a matter of fact, even in an ideal

http://dx.doi.org/10.1016/j.artmed.2014.10.004 0933-3657/© 2014 Published by Elsevier B.V.

Please cite this article in press as: Sacchi L, et al. From decision to shared-decision: Introducing patients’ preferences into clinical decision analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.10.004

G Model ARTMED-1365; No. of Pages 10

ARTICLE IN PRESS L. Sacchi et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx

2

evidence-based and clinical practice guideline (CPG) decisional environment, there might exist situations in which it is desirable to involve the patient and his/her preferences in the decision. This may happen due to the limitation/poor applicability of the scientific evidence or to an explicit indication of the CPG that suggests the use of the patient’s preferences [6]. The process during which the patient and his/her care provider reach a clinical decision together is known as shared decision making (SDM) [7–9]. In this paper we present a general web-based framework that can be used by physicians to exploit some methodologies commonly used in clinical decision-making and to couple them to instruments for eliciting patients’ preferences and performing SDM. The concept of preference we refer to involves the patient’s actual perception of his/her health condition, the perception of the consequence of a therapeutic choice and also the impact of such choices on the patient’s out-of-pocket costs. Our approach includes a utility model and a cost model. Such models are coupled to a theoretical decision model framework to solve the decision task. The final decision will thus account for patient-specific parameters, which might be different from population parameters derived from the literature. To the best of our knowledge, a system incorporating CPGs, decision-theoretic models and tools for preferences elicitation is still not available in the literature. The framework was developed in the context of MobiGuide (http://www.mobiguide-project.eu/, accessed: 07 October 2014), a project funded under the European commission 7th framework program and carried out by a consortium of 13 partners from several European countries (Israel, Italy, The Netherlands, Spain, and Austria). It is aimed at developing a knowledge-based patient guidance system based on computer-interpretable guidelines (CIGs) and designed for the management of chronic or subacute illnesses, including atrial fibrillation (AF). The main components of the MobiGuide System are a decision support system (DSS), devoted to the representation and execution of CIGs, and a body area network (BAN), including a network of sensors and a smartphone, to support telemonitoring of the patient. The data collected by the system are stored in a patient health record (PHR). Of the project objectives, here we focused on the identification and analysis of the CPG recommendations requiring a shared decision due to the lack of robust clinical evidence. After this identification step, a suitable framework is set up to support the physician to properly manage the SDM process. This paper extends a previous work [10] by providing deeper insight into both the architectural and theoretical aspects of the framework. In addition, a first evaluation of the system is presented. The paper is organized as follows: in Section 2 we introduce the theoretical basis of the proposed framework and the implementation strategy within the MobiGuide system. Section 3 presents the proposed models and the implemented interfaces on a specific application regarding thromboembolism prevention in AF patients. Section 4 shows the results of a preliminary clinical application. In Section 5 we discuss the proposed framework and, in Section 6, we draw our concluding remarks.

2. Methods The definition of the shared-decision framework presented in this paper is the result of a set of methodological steps. First, we developed a general model of all the concepts required for the framework definition and of how these are interconnected. The second step included the identification of methodologies and technologies for (i) the collection of patients’ data and preferences and (ii) the development and implementation of the decision-theoretic

models. Finally, we set up the system architecture to make the framework available to the MobiGuide project. In this section we provide a detailed description of each of these steps. 2.1. An ontology for shared decision-making The general framework for SDM presented in this paper takes into account several aspects, starting from the implementation of decision theoretic models relying on CPG recommendations and providing different facilities for eliciting patients’ preferences and automatically including them in the models. The ontology depicted in Fig. 1 represents how all the concepts and methods analyzed in this paper interconnect to create the SDM scenario. In this representation, classes of concepts are shown in rectangles and connecting arrows express the relationships between the concepts. In general, a guideline presents one or more decision points, which are the clinical problems in which a shared decision is suggested or needed (class DecisionProblem). Every shared decision problem has multiple possible solutions (class DecisionOption), which are the options from which the patient and the care professional will select the final decision (the ‘do-nothing’ option may also be present). The selected decision (class FinalSharedDecision) is an Option and it is chosen from all the available decision options. A decision problem can be represented by a decision model (class DecisionModel). A decision model considers several decision options and includes a set of variables. These are probabilistic variables (class ProbabilisticVariables), which can be either health states (class HealthStateVariable), the results of diagnostic tests (class TestResults) or some relevant patient behaviors (class PatientBehavior). Since, in a decision model, events are generally represented following their temporal sequence, we have included the arc precedes to take this into account. An important concept in decision analysis is quantification. It represents the step in the modeling that allows the variables that are part of the model to be assigned values. Health states can be associated with specific values (class HealthStateValue), which represent the patient’s perception of a health condition. These values can be either utility coefficients (class UtilityCoefficient) or some ranking values (class RankingValue), as we will detail in Section 2.2.1. The class ProbabilityValue is used to represent the quantification of the probabilistic values of the model variables. Demographic data, such as age and gender, can be useful in selecting the correct probability value and have been included in the ontology using the class ‘Demographic’. Another quantification component, although not exploited in all the shared decision problems, is the cost (class CostValue). In general, costs can be divided into those with an impact on the national healthcare service (class NHSCost), costs that impact on society (class SocietyCosts) and costs that directly impact on the patient (class OutOfPocketCost). This latter class is central in our SDM framework, as it can be quantified through the patient’s direct participation of. For this reason, it will be better detailed in Section 2.2.2. The decision model output consists of a set of results (class Result), obtained on the basis of the quantification settings and related to the decision options. Results can be of a varied nature: they can be the expected values for some quantities (class ExpectedValues), they can be indices (class Indices) or they can be the results of additional analyses, such as Monte Carlo simulations (class MonteCarloSimulationResults) or sensitivity analysis (class SensitivityAnalysisResults). The expected values that are calculated by our decision models are: life years (class LifeYears), Quality Adjusted Life Years (QALYs – class QALYs) and costs (class CostValue, which, in this case, can be either a quantification step or a result when it represents the expected value of a decision option). Among the indices, we have selected ICUR (Incremental Cost/Utility

Please cite this article in press as: Sacchi L, et al. From decision to shared-decision: Introducing patients’ preferences into clinical decision analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.10.004

G Model ARTMED-1365; No. of Pages 10

ARTICLE IN PRESS L. Sacchi et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx

3

Fig. 1. The ontology for SDM. Rectangles represent classes of concepts and arrows represent the relationships connecting them.

Ratio) and ICER (Incremental Cost/Effectiveness Ratio). These two indices are computed starting from the analysis results, as the change in costs divided by the change in benefits. In particular, ICUR is calculated using QALYs and costs and ICER is calculated using costs and life years. The individuals taking part in the shared decision process are represented by the class ‘Agent’ in the ontology. The patient (class Patient) and the care professional (class CareProfessional) work together to take the shared decision. In this schema we have divided the care professionals into physicians (class Physician) and psychologists (class Psychologist), to allow for the presence of a specialist supporting the physician in the process. These two professional figures are considered the most suitable to present the proposed framework to the patients. Although other professionals might be added, the psychologist is the most skilled for supporting the utility elicitation task. As a matter of fact, it is important to point out that this ontology includes mainly the features that have been identified during the definition of our system. It thus represents a first step toward the definition of a comprehensive ontology of SDM. 2.2. Collecting patients’ preferences – values, utilities and costs A patient’s health condition (status) may vary during his/her follow up. The probability of occurrence of each health state, and of transition between states, is highly dependent on the selected treatment option. Intuitively, different patients may have a different perception of the quality of life related to health states.

Moreover, one patient might consider the economic impact of a specific treatment as more relevant for the choice compared to another individual. For this reason, it is very important to tailor the decision process to the single patient, taking into account this variety of aspects. 2.2.1. The utility model From the observations above, it is clear that it is important to measure the quality of life the patient associates to specific conditions. Quality-adjusted life years (QALYs) [11] is one of the most known and used indicators, combining in a single value the life expectancy and the subjective perception of the health states considering physical, mental and social aspects. Given a time span T divided into n time intervals ti , i = 1, . . ., n, each one spent in a particular health state si , QALYs are defined as



i

= 1, . . ., n(ti ∗ ui ),

where ui is known as the utility coefficient (UC) for si . To define QALYs, we thus need to characterize each health state by a UC, ranging from 0 (usually associated to death) to 1 (perfect health). Literature and the web provide UCs for several health states (e.g. [12,13]). Such coefficients can be also conveniently elicited from the single patient, according to the physician/psychologist judgment about the patient’s capability to understand both the clinical conditions associated to each health state and the methodologies used for elicitation. It is to be noted that UCs are related to

Please cite this article in press as: Sacchi L, et al. From decision to shared-decision: Introducing patients’ preferences into clinical decision analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.10.004

G Model ARTMED-1365; No. of Pages 10 4

ARTICLE IN PRESS L. Sacchi et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx

health states so if different decision options involve the same states, UCs need to be elicited only once. However, periodical reassessment of UCs may be necessary, as patients’ perception of health states changes over time. The proposed SDM framework is provided with two traditional methods for eliciting utility coefficients, i.e. standard gamble (SG) and time trade-off (TTO). In addition, the rating scale method (RS) is available. In the latter, the patient is asked to rate all the states represented in the decision model on a specific scale (for example ranging from 0 to 100). While unsuitable for QALY calculation [14], values collected using RS and properly rescaled to range between 0 and 1, represent a patient-specific ranking of the states, which can be useful for further consistency check. The TTO method was specifically developed for use in health care [15]. To elicit utility for a specific health state s, the patient is asked to compare two different scenarios. In the first, he is prospected to stay in s for a time t computed as the life expectancy of an individual with the same age and the same chronic condition. In the second scenario, the patient is prospected to live a totally healthy life for a time x < t. During elicitation, time x is varied until the patient is indifferent between the two alternatives and the UC for the examined state is computed as the ratio x/t. In the SG approach [16], the patient is offered two alternatives. Alternative 1 is a treatment that, if successful, might enable him/her to live in perfect health for the rest of his/her life. Such treatment, though, carries a certain risk of death r (think, for example, of a surgical intervention). Alternative 2 is that the patient lives all his/her life in the specific health state under analysis. During the face-to-face encounter, the value for r is varied until the patient is indifferent between the two alternatives, and the UC is computed as 1 − r. 2.2.2. The cost model Besides quantifying the patient’s perception of different health states, it is also useful to involve him/her in the quantification of a cost model. While, in the utility model, the patient gives his/her opinion about the health states taking into account all the nonmonetary aspect related to them, for the cost model he is asked to provide the information needed for quantifying the monetary costs related to the clinical paths that are generated as a consequence of the different decision options. In this model we consider the socalled “out-of-pocket” costs, which are the costs directly burdening the patient and causing an economic impact on his/her activities. To build a general model, we have considered four categories of costs: (a) costs related to the appointments the patient has to undergo during his/her treatment; (b) costs related to domiciliary care the patient may be in need of; (c) home adaptation costs and (d) costs related to the drugs the patient has to purchase. As regards points (a) and (d), the value for the costs of the appointments and drugs directly imputed to a patient depends on the position of the patient with respect to the national healthcare service. In some countries, such as Italy, some patients might have these costs entirely covered, while, in some others, costs might wholly impact upon the patient’s resources. As regards appointments, besides considering the specific health care related costs, other costs have also been taken into account. Examples of these are: • the cost of the trip to the medical center where the examination is performed: this is a patient-specific value that depends on the distance the patient has to travel, transportation availability, etc.; • the cost of the meals a patient might have to pay for during the appointment day; • patient’s productivity loss, if the patient is self-employed or retired;

Fig. 2. Simple generic decision tree model. This model shows a decision node (square in the graph), from which a set of branches (the decision options) start. Circles symbolize chance nodes and triangles denote terminal nodes.

• the cost related to an assistant possibly needed to help the patient to reach the appointment location: the model takes into account the cost of travel and meals. In addition, it is possible to quantify the assistant’s time in terms of productivity loss or salary. The inclusion of costs related to domiciliary care takes into account the possibility of domiciliary assistance required after severe events (e.g. a stroke). In cases where the assistant is a professional employed by the patient, cost (b) is quantified by the salary given to the assistant. If, on the other hand, the assistant is a member of the patient’s family, this cost is quantified in terms of productivity loss. Assistance to the patient after a specific health event may also imply some home adaptation to manage the impairments the patients may experience after the event. These costs are assessed based on [17], where the authors present an analysis of the overall social costs of stroke in Italy. In this analysis, a full-costing procedure is applied to identify cost-generating components and to attribute appropriate costs in terms of direct costs and productivity loss. Since several cost components are related to the specific patient’s context, the quantification holds as long as the context remains the same. When the physician perceives the context of the patient might have changed, a reassessment of the costs should be considered. 2.3. The decision model The decision models that have been selected for building the presented SDM framework are decision trees (DT) [5]. Even though other formalisms, such as influence diagrams or decision tables, are available, DTs allow easier knowledge elicitation by the medical experts. DTs are one of the most widely used formalisms in the analysis of the logical structure and timing of clinical decisions. They connect the alternative decision options to their expected effects and the final outcomes of each possible scenario. This is done following a formalism based on the combination of nodes and branches. Fig. 2 shows a portion of a simple generic DT structure. In a DT there are decision nodes, chance nodes and terminal nodes. Decision nodes are the starting points for the alternative options the study is considering. In the example shown in Fig. 2, the decision node is the square node labeled as “Eligible Population”, i.e. the patients eligible for that DT. Chance nodes (circles in Fig. 2) symbolize uncertain events. Each event is characterized by a finite number of possible outcomes, which must be exhaustive and mutually exclusive. Each outcome is associated with its occurrence probability. A terminal node (triangles in Fig. 2) identifies the end of a path, and it is associated with a payoff value, characteristic of that path. A payoff is an outcome that the decision maker wants to maximize (e.g. QALY) or minimize (e.g. cost). If events recur over time and transitions among states must be represented, a Markov model (MM) [18] is integrated in the DT. Running or solving a decision tree means calculating the expected values of the payoffs for all the possible decision options,

Please cite this article in press as: Sacchi L, et al. From decision to shared-decision: Introducing patients’ preferences into clinical decision analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.10.004

G Model ARTMED-1365; No. of Pages 10

ARTICLE IN PRESS L. Sacchi et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx

by weighting the values at the end of the paths with their probability of occurrence. After a tree is run, the decision node shows, for each strategy, these expected values. The solution suggested by the model is the one optimizing the payoff. In case of multiple, competing payoffs, the results of the tree cannot be used as direct suggestions. Instead, they represent useful quantitative values on which the physician and the patient can discuss and reason together. For example, suppose decision option 1 gives better results in terms of QALYs than decision option 2, which instead gives less out-of-pocket costs. In this case, the final choice could depend on whether the patients’ financial status allows him to spend more money to gain a higher quality of life. In our framework, DTs are initially fed with probabilities and UCs found in the literature, and can be used as they are, when physicians judge that the patient is not able to provide additional, more personalized information. For the most frequent case, though, when the patient is able to provide his/her own preferences and context details, DTs can be personalized with patient’s preferences and his/her profile features collected and stored through the aboveillustrated generic models. As a result, decision-options are ranked on a personalized basis. It should be noted that, while the utility and the cost models presented so far are generic models, i.e. they can be used for any type of shared decision, each specific decision problem requires the implementation of a specific decision-theoretic model. In the MobiGuide project, trees are built using the commercial TreeAge (TreeAge Software Inc, www.treeage.com, accessed: 07 October 2014) tool. Moreover, we have developed a web interface using the TreeAge Pro Interactive tool, to also make the models available to the end users less familiar with the modeling technique. Physicians will thus be able to browse probabilities, utilities, and costs, and to adjust values according to their knowledge, if needed. They will then run the decision tree and see the results. 2.4. The framework architecture in the MobiGuide setting The SDM framework presented in the previous sections has been implemented into the MobiGuide system according to the architecture shown in Fig. 3. The access to the framework is provided through the caregiver interface, which is the main component through which the care professional manages his patients, edits their data, and gets CPG recommendations through the decision support system. The caregiver interface is directly connected to the PHR, the repository of patient’s clinical data. When a specific CPG recommendation triggers the possibility of performing a shared decision, the physician is presented with three different links on his interface. One link directs to a set of specific aids to explain the decision problem and illustrating to the patient the possible health states related to it. The second link directs to the interface for eliciting utility coefficients and values. The third link directs to the interface that allows the decision tree to be run. The possibility to have an automatic link from the PHR interface to the SDM opportunities was recognized in a recent study as one of the possible facilitators for practical implementation of SDM [19]. Both the interface for utility coefficients elicitation and the interface for running DTs rely on a support relational database to store the tree characteristics necessary for user interaction (i.e. the represented health states and some numerical parameters) and the tree results. Results are stored together with data about the interaction session, such as the identity of the patient and physicians participating in the encounter, and their opinion on the usefulness of such interaction. This data model is derived from the ontology described in Section 2.1. It is composed of static tables describing the health states considered in the different decision trees, and dynamic tables storing all the utility coefficients elicited from patients using the three methods.

5

3. Results – implementation of the framework for the management of thromboembolism risk in atrial fibrillation In this section we present the implementation of the proposed SDM framework. As already mentioned when introducing decision models, every specific problem requires a DT to be designed and built. To illustrate the application of the framework, we focus on one of the SDM paths developed for the MobiGuide project: the prevention and management of the risk of thromboembolism for AF patients. The guideline itself explicitly advises the physician to consider patients’ preferences in the selection of the antithrombotic therapy for the subgroup of patients presenting with a low thromboembolism risk. The AF guideline [20] reports the following indication: For primary prevention of thromboembolism in patients with nonvalvular AF who have just 1 of the following validated risk factors, antithrombotic therapy with either aspirin or a vitamin K antagonist is reasonable, based upon an assessment of the risk of bleeding complications, ability to safely sustain adjusted chronic anticoagulation, and patient preferences: age greater than or equal to 75 y (especially in female patients), hypertension, HF, impaired LV function, or diabetes mellitus (Class IIa, Level of Evidence A1 ). Another recommendation [21] nominates dabigatran as an alternative to vitamin k antagonists: Dabigatran is useful as an alternative to warfarin for the prevention of stroke and systemic thromboembolism in patients with paroxysmal to permanent AF and risk factors for stroke or systemic embolization who do not have a prosthetic heart valve or hemodynamically significant valve disease, severe renal failure (creatinine clearance 15 mL/min) or advanced liver disease (impaired baseline clotting function) (Class I, Level of Evidence B2 ). These recommendations involve the choice whether or not to treat and, in case of treatment, the selection between three drug options. These are: acetyl salicylic acid (ASA), oral anticoagulant therapy (OAT) with vitamin K antagonists (e.g. warfarin), or OAT with dabigatran, a new oral anticoagulant recently made available for clinical use. As explicitly stated in the CPG text, patients’ preference should be taken into account and weighted in reaching the final decision. This configures a typical case of SDM as defined in our study. We modeled this scenario using a modified and adapted version of a published DT [22] combined with MMs to compare the clinical pathways of an AF patient who may undergo one of the aforementioned treatments for stroke prevention, or who takes no drug therapy at all. For the model quantification we relied on published clinical studies [23–27]. During the Markov process, individuals’ moves between health states have transition probabilities, which may also vary over time. In agreement with the CPG recommendations, the set of decision options includes: no treatment, ASA, warfarin and dabigatran. As outcomes, we have considered life years, QALYs, and costs. Fig. 4 shows a simplified representation of the implemented MM. A patient enters the process in the nonvalvular AF state (NVAF-only). During the course of the disease, he can experience events such as myocardial infarction (MI), ischemic stroke (IS), intracranial hemorrhage (ICH), and extracranial bleedings. Temporary IS or ICH are events that cause only a transient disability and after which the patient recovers and goes back to the NVAF-only state. A patient experiencing more severe events, such as a mild/moderate-severe IS or ICH, is often subject to permanent impairment. The occurrence of these events depends on the different transition probabilities that are related both to the treatment and to the patient’s risk of stroke, calculated on a CHADS2 score basis [28]. The administration of warfarin or ASA decreases the probability of occurrence of IS, but increases the probability of ICH and extra-cranial bleedings. On the other hand, the choice of not prescribing any therapy increases the probability of IS while limiting the occurrence of ICH and extra-cranial bleedings. If, while undergoing therapy with warfarin, a patient in the AF-only state experiences an ICH or major extra-cranial bleeding, OAT therapy is interrupted and replaced by ASA to decrease the probability of further bleeding. In order to make the decision analysis a shared decision, the physician may access the interface allowing him to elicit UCs from his patients. This interface is designed to be used in face-to-face encounters between patients and physicians. Using the proposed framework, the physician can interact with the patient to elicit values and UCs using the entire set of approaches presented in Section 2.2.1 The interface has been designed to give the patient the best possible understanding of the questions to be answered. For example, the interface for the SG method (that requires reasoning about a possibly curative procedure with a risk of death) is enhanced with a graphical representation. For each value of the risk a chromatic scale is provided that turns a set of yellow smileys red corresponding to the risk percentage. Besides utilities, costs are the other aspect that tailors the framework to the specific patient. On the basis of the cost model presented in Section 2.2.2, a

1 Class IIa: Weight of evidence/opinion is in favor of usefulness/efficacy. Level of Evidence A: data derived from multiple randomized clinical trials or meta-analyses (20). 2 Class I: Conditions for which there is evidence and/or general agreement that a given procedure/therapy is beneficial, useful, and effective. Level of Evidence B: data derived from a single randomized trial, or nonrandomized studies (20).

Please cite this article in press as: Sacchi L, et al. From decision to shared-decision: Introducing patients’ preferences into clinical decision analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.10.004

G Model ARTMED-1365; No. of Pages 10 6

ARTICLE IN PRESS L. Sacchi et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx

Fig. 3. The architecture of the proposed framework in the MobiGuide system.

questionnaire has been designed to gather the necessary personal data for cost calculation. The part of the questionnaire related to the cost for anticoagulant treatment monitoring in the case of treatment with vitamin K antagonists is reported in Fig. 5. DT results are presented in the interface as shown in Fig. 6. The expected values for all the defined payoffs for each decision option are listed. The

most common case is that the most expensive option is also the most effective from the health outcomes point of view. Nevertheless, it may be that an option shows higher life expectancy but lower QALY. These are the situations that require the patient to reason with his/her doctor about the best choice.

Fig. 4. Simplified Markov model representation. Patients enter the Markov process in the non-valvular atrial fibrillation (NVAF-only) state, where they can remain or move to the health states: IS (ischemic stroke), ICH (intracranial hemorrhage), MI (myocardial infarction), minor extracranial bleedings, major extracranial bleedings. The states represented with dotted lines are transient. From each state, patients can move to the absorbing state (death).

Please cite this article in press as: Sacchi L, et al. From decision to shared-decision: Introducing patients’ preferences into clinical decision analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.10.004

G Model ARTMED-1365; No. of Pages 10

ARTICLE IN PRESS L. Sacchi et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx

7

Fig. 5. A portion of the questionnaire for quantifying the cost model. International Normalized Ratio (INR) is a laboratory test required to monitor anticoagulant therapy effects, and it must be done, in general, every 15 days. INR evaluation is requested for vitamin K antagonists, while it is no longer necessary for drugs like dabigatran (new oral anticoagulants).

Fig. 6 represents an example of a DT run on a sample patient using only utility coefficients taken from the literature. In this example it is noticeable that dabigatran is the treatment option leading to the best values for QALYs and Life Years. The cheapest option is, intuitively, ‘no treatment’. Patients not on anti-thrombotic therapy will, nevertheless, undergo medical appointments and they might be in need for assistance if they enter some particular path during follow up. For this reason the cost related to the no treatment option is not zero. As a matter of fact, choosing not to treat the patient leads to lower values for the other outcomes. The cost difference of the two oral anticoagulants is mainly due to the need for regular international normalized ratio (INR) value controls with vitamin K antagonists (e.g. warfarin), which is not required with newer oral anticoagulants (such as dabigatran). The example shown in Fig. 6 shows the expected values for the outcomes in the case of a specific patient. These expected values come from a probability distribution, which can be shown to the patient (Fig. 7) to better compare the variability of the possible outcomes for the different options. In this figure it is, for example, possible to note the higher prevalence of low survivals in the no treatment option.

4. Clinical evaluation study For an initial assessment of the framework, a pilot evaluation study was carried out in collaboration with the Cardiology Division at the IRCCS Fondazione Salvatore Maugeri hospital, located in Pavia (Italy) and partner in the MobiGuide project. Twenty patients, 10 males and 10 females, participated in the evaluation of the

interface for utility coefficients elicitation. The mean age of these patients was 66.2 years (range: 34–79). The distribution of the AF types in the subjects was as follows: 5 patients were affected by permanent AF, 9 patients suffered from persistent AF, while, for 6 subjects, AF appeared as a consequence of cardiothoracic surgery they had just undergone. The patients, guided by a physician, were taken through the pages of the utility coefficients elicitation interface and, after a thorough explanation of the three methodologies, were asked to answer the corresponding questions related to the AF state. One of the selected patients had cognitive impairment and was not able to interact with the physician for the mentioned tasks. To evaluate the interface, we first measured how long it took to each patient to understand and answer the questions proposed by the physician. Rating scale administration required an average time of 1.5 min (range 1–2), TTO required an average time of 3.2 min (range 2–5) and Standard Gamble required, on average, 5.9 min (range 2–15) to be completed. These values include the time taken by the physician to explain the methodology-specific scenarios (the explanation time accounted for around 2/3 of the total time taken for elicitation for TTO and SG). Given that the average times are

Fig. 6. Example of the results provided by running the decision tree on a specific patient: expected values of the payoffs (rows) for the possible therapeutic options (columns). In particular, a 66 year-old patient is considered, with utility default values taken from the literature for all the health states defined in the model.

Please cite this article in press as: Sacchi L, et al. From decision to shared-decision: Introducing patients’ preferences into clinical decision analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.10.004

G Model ARTMED-1365; No. of Pages 10 8

ARTICLE IN PRESS L. Sacchi et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx

Fig. 7. QALY distribution for dabigatran and no treatment obtained running a Monte Carlo simulation with a set of 10,000 random trials. Vertical lines show the 10th, 50th, and 90th percentiles.

quite low, these findings ensure that using the utility coefficients elicitation interface would not prolong the appointment duration too much. This is very important as it has been shown that one of the main barriers to actual implementation of SDM in clinical practice is time constraint [19]. After collecting the utility coefficients using both TTO and SG for all the patients, we found an average UC associated to the NVAFonly health state of 0.976. This value was higher than the one used for the default quantification of our DT (0.779) [29]. One reason for this difference is related to the utility elicitation method. In [29], coefficients were calculated using the EuroQOL questionnaire (http://www.euroqol.org/, accessed: 07 October 2014), which considers the overall condition of the patients that may include comorbidities. In our case, we focused on the UC related only to AF because, in our population, the presence of chronic comorbidities was limited (only 4 patients had hypertension and one also had hyperthyroidism). The observed variation in the UC values with respect to the default creates a suitable context for showing how the proposed SDM framework could work using personalized values in the case of a population with specific characteristics. To this end, we ran two decision trees for each patient: one decision tree using the default UC and one using the personalized coefficient related to the NVAF-only health state directly elicited from the patient. The utility coefficients for the other health states were left as defaults, since they were not elicited for the examined sample. Results were considered in terms of QALYs and compared across the entire population set. A paired Wilcoxon test was used to determine significant differences of QALYs for the available decision options between the two cases (personalized UC vs default UC). As expected, we obtained significant results for all the treatments

(p-value

From decision to shared-decision: Introducing patients' preferences into clinical decision analysis.

Taking into account patients' preferences has become an essential requirement in health decision-making. Even in evidence-based settings where directi...
2MB Sizes 1 Downloads 8 Views