Will 'Computable' Clinical Guidelines Be Compatible with Personalised Care?

INTRODUCTION: The potential benefits from digitalisation processes will only be fully realised if the conceptual challenges they uncover are accepted and addressed, alongside the technical ones such as interoperability. Will 'computable' clinical guidelines be compatible with personalised care if the definition of the relevant disease embeds preferences that pre-empt those of the individual patient? METHOD: As a case study we investigated the definition of diabetes in glycaemic management guidelines. RESULT: The dominant component of its definition - HbA1c ≥6.5% - embeds the consensus preference judgement of a 2009 International Expert Committee. DISCUSSION: This preference-sensitive threshold for the diagnosis of diabetes has subsequently been endorsed in many guidelines relating to glycaemic management, though there are signs of awareness and concern with its implications. CONCLUSION: Those seeking to digitalise guidelines by making them 'computable' need to acknowledge and address their inbuilt preference-sensitivity - if they wish to further care that respects patient's preferences.

Inferring Causality Is Preference-Sensitive: We Need a Book of Who as Well as Why.

In multiple publications over 3 decades, most recently in The Book of Why, Judea Pearl has led what he regards as the 'causal revolution'. His central contention is that, prior to it, no discipline had produced a rigorous 'scientific' way of making the causal inferences from observational data necessary for policy and decision making. The concentration on the statistical processing of data, outputting frequencies or probabilities, had proceeded without adequately acknowledging that this statistical processing is operating, not only on a particular set of data, but on a set of causal assumptions about that data, often unarticulated and unanalysed. He argues that the arrival of the directed acyclic graph (DAG), a 'language of causation' has enabled this fundamental weakness to be remedied. We outline the DAG approach to the extent necessary to make the key point, captured in this paper's title regarding DAG's potential contribution to improved decision or policy making.

The Causal Plausibility Decision in Healthcare.

The explosion of interest in exploiting machine learning techniques in healthcare has brought the issue of inferring causation from observational data to centre stage. In our work in supporting the health decisions of the individual person/patient-as-person at the point of care, we cannot avoid making decisions about which options are to be included or excluded in a decision support tool. Should the researcher's routine injunction to use their findings 'with caution', because of methodological limitations, lead to inclusion or exclusion? The task is one of deciding, first on causal plausibility, and then on causality. Like all decisions these are both sensitive to error preferences (trade-offs). We engage selectively with the Artificial Intelligence (AI) literature on the causality challenge and on the closely associated issue of the 'explainability' now demanded of 'black box' AI. Our commitment to embracing 'lifestyle' as well as 'medical' options for the individual person, leads us to highlight the key issue as that of who is to make the preference- sensitive decisions on causal plausibility and causality.

Individualising Life Expectancy Is Necessary for Optimal Prescribing.

One possible cause of overprescribing (or insufficient deprescribing) is the failure to explicitly address the individual's life expectancy (LE). For example, if a LE estimate shows the person has six months to live, this should influence the prescribing of a medication that offers benefits only over a much longer LE. Predicting exactly the number of years a person will live is impossible, but probabilistic forecasting is possible and arguably essential, both for the selection of the optimal intervention and for meeting the 'reasonable patient' standard of information about the harms and benefits of alternative options. One side-effect of the COVID-19 pandemic has been to bring mortality into greater prominence, hopefully facilitating its discussion in the clinic as part of the 'new normal'. This paper outlines the case for introducing LE into prescribing decisions as a way of making more individualised decisions and potentially reducing overprescribing. It concentrates on how the clinical task of arriving at individualised estimates of LE might be tackled, especially in the case of the growing number of older patients with heterogeneous sociodemographic characteristics who are experiencing multiple long term conditions of varying severity and are frequently subject to 'polypharmacy'.

Communicating About Mortality in Health Decision Support: 'What and Why and When, and How and Where and Who'.

The Covid-19 pandemic has only accelerated the need and desire to deal more openly with mortality, because the effect on survival is central to the comprehensive assessment of harms and benefits needed to meet a 'reasonable patient' legal standard. Taking the view that this requirement is best met through a multi-criterial decision support tool, we offer our preferred answers to the questions of What should be communicated about mortality in the tool, and How, given preferred answers to Who for, Who by, Why, When, and Where. Summary measures, including unrestricted Life Expectancy and Restricted Mean Survival Time are found to be reductionist and relative, and not as easy to understand and communicate as often asserted. Full lifetime absolute survival curves should be presented, even if they cannot be 'evidence-based' beyond trial follow-up limits, along with equivalent measures for other criteria in the (necessarily) multi-criterial decision. A decision support tool should relieve the reasonable person of the resulting calculation burden.

Measures of Decision Aid Quality Are Preference-Sensitive and Interest-Conflicted - 1: Normative Measures.

The belief that following rigorous inclusive methods will eliminate bias from 'quality' measures ignores the preferences necessarily embedded in any formative instrument. These preferences almost always reflect the interests of its developers when one uses the wide definition of 'interest' appropriate in healthcare research and provision. We focus on the International Patient Decision Aid Standards instrument, a popular normative measure of decision aid quality. Drawing on its application to a set of 23 breast cancer screening decision aids, we show the effects of modifications that reflect our own different interest-conflicted preferences. It is emphasised that the only objection is to the implication that any formative instrument should be promoted or treated as the 'the gold standard', without a conflict of interests disclaimer, and to the implication that other instruments cannot provide equally valid, high-quality measures.

Measures of Decision Aid Quality Are Preference-Sensitive and Interest-Conflicted - 2: Empirical Measures.

Empirical measures of 'decision aid quality', like normative ones, are of a formative construct and therefore embody interest-conflicted preferences in their criteria selection and weighting. The preferences of the International Patient Decision Aid Standards consortium distinguish the quality of the decision-making process and the quality of the choice that is made '(i.e., decision quality)'. The Decision Conflict Scale features heavily in their profile measure of the former and Decision Quality Instruments (DQIs), have been developed by members of the consortium to measure the latter. We confirm that both of these, and other components, like the higher-level measures, are preference-sensitive and interest-conflicted. Non-financial interest-conflicted preferences are endemic in healthcare research, policy-making, and practice. That they are inevitable means the main problem lies in the denial of this and attitude to and behaviour towards alternatives, equally interest-conflicted.

Multicriteria Decision Support Would Avoid Overdiagnosis and Overtreatment.

Population-level studies confirm the existence of significant rates of overdiagnosis and overtreatment in a number of conditions, particularly those for which the screening of asymptomatic individuals is routine. The implication is that the possibility of being overdiagnosed and/or overtreated must be mentioned as a possible harm in generating informed consent and participation from the individual invited to be screened. But how should the rates of such preference-insensitive population-level phenomena be introduced into preference-sensitive individual decision making? Three possible strategies are rejected, including the currently dominant one that involves presenting the rates relevant to overdiagnosis and overtreatment as discrete pieces of information about a single criterion (typically condition-specific mortality). Extensive quotation from a review of cancer decision aids confirms that processing this complex and isolated information is not a practical approach. However, the task is unnecessary, since an outcome-focused multicriteria decision support tool will incorporate the effects of overdiagnosis and overtreatment - along with the effects of any underdiagnosis and undertreatment.

COVID-19, the Swedish 'Experiment', and Me.

In this personal vision paper the Swedish approach to COVID-19 prompts an exploration of how and why assuming individual rationality coupled with minimal social restriction may be as good a solution as any and better than most. A COVID sub-model is developed and populated with probabilities for four outcomes of infecting another person (asymptomatic, sick, hospitalized, dead), conditional on three observable characteristics (sex, age, and BMI), and (dis)utilities for three categories of person (nearest/dearest, friends/colleagues and unknown others) experiencing those outcomes. The implications for a liberal democracy are drawn, based on the assumptions that individual citizens will and should maximise their informed expected utility, exhibiting 'commons sense' as well as common sense.

Evaluations of Decision Support Tools Are Preference-Sensitive and Interest-Conflicted: The Case of Deliberation Aids.

The questions 'What constitutes a good health care decision?', and, by extension, 'What constitutes good healthcare decision support?' continue to be asked. The most developed answers focus largely, often exclusively, on the quality of the 'deliberation' component as the determinant of the quality of the decision or decision aid. We argue that these answers and resulting aids reflect the preferences of healthcare professionals and aid developers and that these preferences are closely aligned with their interests. Some interests are material, but many professional, institutional, intellectual, methodological, and ethical. Successful promotion of a particular preference-sensitive, interest-conflicted decision aid does not change its ontological nature. Conflicts of interest are therefore universal and of concern only when this ontology is denied and if aids based on alternative interest-based preferences, such as technologies involving numerical analytic calculation, are subjected to discrimination.

People Living with Multiple Long-Term Conditions: Meeting the Challenges of Personalized Decision Making.

As with any diagnosis, the underlying purpose of a 'multimorbidity' one is to identify and establish the impact of a person's health conditions on their lives and to facilitate personalized decisions regarding proposed interventions. Clinicians routinely make decisions about the use of interventions for people with multiple long-term conditions. This is challenging because evidence to support this process currently relies on guidance on single health conditions for people without multimorbidity, typically taking fewer medications. Establishing the person's preferences over relevant criteria is central to a person-centered decision-making process, and it is particularly challenging, given the complexities of the person's multiple conditions. The final challenge is in combining the clinician's best estimates of the benefits and harms of possible interventions with the person's preferences. A review of these challenges, drawing on the NICE guidelines, leads to a proposal for using a Multi-Criteria Decision Analysis-based support tool for personalized shared decision making for multiple long-term conditions.

Are Clinical Decision Support Systems Compatible with Patient-Centred Care?

Few, if any, of the Clinical Decision Support Systems developed and reported within the informatics literature incorporate patient preferences in the formal and quantitatively analytic way adopted for evidence. Preferences are assumed to be 'taken into account' by the clinician in the associated clinical encounter. Many CDSS produce management recommendations on the basis of embedded algorithms or expert rules. These are often focused on a single criterion, and the preference trade-offs involved have no empirical basis outside an expert panel. After illustrating these points with the Osteoporosis Adviser CDSS from Iceland, we review an ambitious attempt to address both the monocriterial bias and lack of empirical preference-sensitivity, in the context of Early Rheumatoid Arthritis. It brings together the preference data from a Discrete Choice Experiment and the best available evidence data, to arrive at the percentage of patients who would prefer particular treatments from those in the listed options. It is suggested that these percentages could assist a GRADE panel determine whether to produce a strong or weak recommendation. However, any such group average preference-based recommendations are arguably in breach of both the reasonable patient legal standard for informed consent and simple ethical principles. The answer is not to localise, but personalise, decisions through the use of preference-sensitive multi-criteria decision support tools engaged with at the point of care.

Decision Quality Is a Preference-Sensitive Formative Concept: How Do Some Existing Measures Compare?

The primary output of a decision making process is a decision and a key outcome measure is therefore decision quality. However, being a formative construct, 'decision quality' is both preference- and context-sensitive and legitimate alternative measures accordingly exist. A decision maker wishing to measure decision quality in the evaluation of a decision or decision making process needs to be aware of the attributes of the measures on offer. This paper establishes some of the key conceptual differences by examining two measures: Decision Quality Instruments and MyDecisionQuality. Four of their main conceptual differences relate to: the timing of the measurement (at the point of decision or at follow-up when the 'downstream' outcome is known); (whether or not an objective assessment of the information state of the individual is included (as opposed to self-reported state); whether the instrument itself is preference-sensitive; and whether the measure is to be used in the context of individualised clinical practice at the point of care or only in research to produce group level feedback. Establishing agreed measures of decision quality is necessary and useful, so long as it is accepted that it is a preference- and context-sensitive construct, in the way that is widely acknowledged in relation to, for example, Health-Related Quality of Life, with its many measures.

A Generic Rapid Evaluation Support Tool (GREST) for Clinical and Commissioning Decisions.

A fast and frugal generic tool can provide decision support to those making decisions about individual cases, particularly clinicians and clinical commissioners operating within the budget and time constraints of their practices. The multi-national Generic Rapid Evaluation Support Tool (GREST) is a standard preference-sensitive Multi-Criteria Decision Analysis-based tool, but innovatory insofar as an equity criterion is introduced as one of six. Equity impact reflects the number of population QALYs lost or gained in moving from Old (current intervention) to New (contemplated intervention). In the exemplar UK implementation Claxton's NHS Willingness to Pay per QALY is the numeraire. Any weight from 0 to 100% may be assigned to the equity criterion but its presence affirms that it is persons-as-citizens who experience any opportunity harms or benefits arising from actions within the health service commons. A fully-operational but demonstration-only version is available on open access, as proof of concept and method.

The Evaluation of Decision Support Tools Needs to Be Preference Context-Sensitive.

Individuals have different preferences in how they wish to relate to healthcare professionals such as doctors. Given choice, they also have preferences in relation to the type and location of support they want for their health and healthcare decisions. We argue that preference-based clusters within this heterogeneity constitute different contexts and that evaluations of decision aids should be context-sensitive in this respect. We draw attention to two distinct preference-based clusters: individuals with a preference for 'intermediative' decision support as a patient, implemented in a largely qualitative deliberative model, on the one hand, and for 'apomediative' decision support as a person, implemented in a largely quantitative multi-criteria decision analytic model, on the other. For convenience, we refer to the latter as Person Decision Support Tools (PDSTs), leaving Patient Decision Aids (PDAs) for its former, conventional use. Seeking to establish proof of method, we present an online PDST that can help individuals establish which of these two types of decision support they would find optimal. It is based on nine key attributes on which PDAs and PDSTs can be contrasted. Within population heterogeneity, preference clusters should be identified, and acknowledged and respected as contexts relevant to the evaluation of decision support tools.

Strong Recommendations Are Inappropriate in Person-Centred Care: The Case of Anti-Platelet Therapy.

A 'Rapid Recommendation' has been produced by the GRADE group, in collaboration with MAGIC and BMJ, in response to an RCT showing Dual Anti-Platelet Therapy (DAPT) is superior to Aspirin alone for patients who had suffered acute high risk transient ischaemic attack or minor ischaemic stroke. The interactive MAGIC decision aid that accompanies each Rapid Recommendation is the main route to their clinical implementation. It can facilitate preference-sensitive person-centred care, but only if a Multi-Criteria Decision Analysis-based decision support tool is added. A demonstration version of such an add-on to the MAGIC aid, divested of recommendations, is available online. Exploring the results of different preference inputs into the tool raises questions about the strong recommendation for DAPT.

Why a Global PROMIS® Can't Be Kept.

Composite multi-dimensional constructs, such as 'global mental health' and 'global physical health', in PROMIS® instruments and ICHOM standard outcome sets, are formative, not reflective. Their preference-insensitivity means they are potentially misleading in both clinical and policy decision making practice. Their frequent validation by reflective psychometric tests is also improper methodologically. The spread of these instruments is occurring without sufficient awareness on the part of patients, clinicians, researchers and policy makers that the need for group-specific preference bases ('tariffs') for such measures rules out any possibility of 'international gold standard metrics'.

PROMs Need PRIMs: Standardised Outcome Measures Lack the Preference-Sensitivity Needed in Person-Centred Care.

A growing number of condition-specific standard outcome sets have been developed by the International Consortium for Health Outcomes Measurement in pursuit of 'value-based care'. These sets embrace many Patient-Reported Outcome Measures (PROMs), reflecting a simultaneous commitment to 'patient-centred care'. However, none of these sets embody recognition of the preference-sensitive nature of the decisions that eventually generate the outcome database. 'Patient-Reported Importance Measures' (PRIMs) are the valid source of the required preferences. The ICHOM Stroke standard set is input into a hypothetical Multi-Criteria Decision Analysis-based decision support tool to provide simple confirmation that PROMs should be preference-mix adjusted as well as case-mix adjusted. PROMs need PRIMs if value-based care is to be personalised values-based care.

Translating the Results of Discrete Choice Experiments into p-/e-/m-Health Decision Support Tools.

The rapidly growing number of health-related Discrete Choice Experiments (DCEs) has not been matched by studies of their impact on decision or policymaking. However, it is widely assumed that this impact has been very limited, despite the potential relevance of the resulting average preferences to group policy development. The main, but at the moment essentially speculative, explanation offered, focuses on the methodological quality of the DCEs and their reporting. An alternative explanation, equally speculative, lies in the research-practice gap created by the conceptualisation of the DCE as a purely research exercise, not supplemented by any attempt to translate the findings into analytic decision support form. This also applies in the clinical decision context, where there are frequent claims that DCE results can assist in an individual's decision making. In the absence of suggestions as to how group results can analytically facilitate preference-sensitive care (and legally informed consent), we propose a generic add-on for DCEs with 'real' options, attributes, and attribute levels. This takes the form of a multi-criteria analysis-based decision support tool. Exemplars, showing how preference-sensitive individualised opinions can be derived from published DCEs for Heavy Menstrual Bleeding and Prostate Cancer Screening, may be consulted online.

A Multi-Criterial Support Tool for the Multimorbidity Decision in General Practice.

The magnitude and seriousness of the challenge posed by heterogeneous multimorbidity in most health services is now unquestioned. Equally well-acknowledged is the fact that existing guidelines essentially set out the principles of best practice in an idealised setting, concentrating on information gathering and not providing any personalisable decision support for the general practitioner aiming to share decision making with a person with multiple morbidities in the reality of routine practice. Existing decision aids have been developed largely within the single condition context and can draw on a body of robust research largely absent in the multimorbidity context. The need for a more flexible and generic approach which draws more on the clinician's expertise-based judgments has been called for, and we introduce a Decision Support Tool, GREST CLIN, in response to this call. Based on Multi-Criteria Decision Analysis it reflects the quantitative calculation approach to decision support rather than the deliberative reasoning of mainstream decision aids. It is emphasised that any evaluation should use the current decision making process as empirical comparator, rather than idealised normative standards, and that the primary outcome measure should be decision quality at the point of care, not downstream outcomes. A demonstration version of the tool is available online as proof of method.