Article Text

Fibrosis, atrial fibrillation and stroke: clinical updates and emerging mechanistic models
  1. Patrick M Boyle1,
  2. Juan Carlos del Álamo2,
  3. Nazem Akoum3
  1. 1 Bioengineering, University of Washington, Seattle, Washington, USA
  2. 2 Mechanical Engineering, University of Washington College of Engineering, Seattle, Washington, USA
  3. 3 Cardiology, University of Washington School of Medicine, Seattle, Washington, USA
  1. Correspondence to Dr Nazem Akoum, Cardiology, University of Washington School of Medicine, Seattle, Washington, USA; nakoum{at}cardiology.washington.edu

Abstract

The current paradigm of stroke risk assessment and mitigation in patients with atrial fibrillation (AF) is centred around clinical risk factors which, in the presence of AF, lead to thrombus formation. The mechanisms by which these clinical risk factors lead to thromboembolism, including any role played by atrial fibrosis, are not understood. In patients who had embolic stroke of undetermined source (ESUS), the problem is compounded by the absence of AF in a majority of patients despite long-term monitoring. Atrial fibrosis has emerged as a unifying mechanism that independently provides a substrate for arrhythmia and thrombus formation. Fibrosis-based computational models of AF initiation and maintenance promise to identify therapeutic targets in catheter ablation. In ESUS, fibrosis is also increasingly recognised as a major risk factor, but the underlying mechanism of this correlation is unclear. Simulations have uncovered potential vulnerability to arrhythmia induction in patients who had ESUS. Likewise, computational models of fluid dynamics representing blood flow in the left atrium and left atrium appendage have improved our understanding of thrombus formation, in particular left atrium appendage shapes and blood flow changes influenced by atrial remodelling. Multiscale modelling of blood flow dynamics based on structural fibrotic and morphological changes with associated cellular and tissue electrical remodelling leading to electromechanical abnormalities holds tremendous promise in providing a mechanistic understanding of the clinical problem of thromboembolisation. We present a review of clinical knowledge alongside computational modelling frameworks and conclude with a vision of a future paradigm integrating simulations in formulating personalised treatment plans for each patient.

  • atrial fibrillation
  • stroke

Statistics from Altmetric.com

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

Introduction

Stroke remains a leading cause of mortality and morbidity in patients with atrial fibrillation (AF).1 It is estimated that 30% of all ischaemic strokes occur in patients with AF.1 Patients with this arrhythmia who have a stroke experience higher mortality and disability compared with ischaemic strokes not associated with AF.2 3 When AF is diagnosed, oral anticoagulation may be recommended based on assessment of the individual’s risk of stroke.4 Risk estimation tools include demographic and comorbid conditions. Their predictive accuracy is modest, and they lack mechanistic association with thrombogenesis.

Cryptogenic strokes, including embolic stroke of undetermined source (ESUS), account for another 30% of ischaemic strokes and are suspected to follow thromboembolisation from a cardiac source. Occult AF is suspected in these patients and cardiac monitoring is recommended to diagnose it.4 The current paradigm of thromboembolisation in AF is challenged by temporal dissociation between the timing of arrhythmia episodes and that of thromboembolic events in patients with cardiovascular implantable electronic device (CIED)-detected arrhythmia.5 6 Similarly, AF detection in patients who had cryptogenic stroke is relatively uncommon,7 8 leaving most patients without an adequate explanation for their stroke.

This review discusses risk stratification tools in AF, device-detected atrial high rate episodes, rhythm monitoring and clinical trials of oral anticoagulants in ESUS. We will also discuss recent evidence linking atrial fibrosis as a common substrate leading independently to AF and stroke. We then provide an overview of computational simulations that can model patient-specific fibrotic substrate to reveal AF perpetuation mechanisms or characterise the relationship between atrial fluid dynamics and thrombogenesis. We conclude with an exploration of transformational work interfacing computational approaches and clinical medicine, as anticipated in the near future.

Clinical updates

Stroke risk assessment in AF

Many scores have been used to risk-stratify patients with AF; CHA2DS2VASc and ATRIA are two of the most recent. A common association is increased risk of stroke with older age. This is captured in CHA2DS2VASc, which assigns 1 point to patients aged 65–74 and 2 points to patients over 74.9 ATRIA assigns different points to these age categories and adds categories with points assigned to those over 85 and those younger than 65 with prior stroke.

Both CHA2DS2VASc and ATRIA recognise the association of diabetes, hypertension, female sex and heart failure with risk of stroke. CHA2DS2VASc includes vascular disease, consisting of peripheral arterial disease, aortic plaque and coronary disease associated with myocardial infarction. ATRIA does not include vascular or coronary disease; however, it includes kidney disease (end stage or chronic disease with estimated glomerular filtration rate <45 mL/min/1.73 m2, proteinuria). Table 1 compares the variables and point systems for CHA2DS2VASc and ATRIA.

Table 1

Comparison of variables and point systems for CHA2DS2VASc and ATRIA scores

CHA2DS2VASc identifies individuals at low, intermediate and elevated risk (0, 1 and ≥2 points, respectively), but its predictive accuracy is modest.10 For ATRIA, low, moderate and high risk correspond to ≤5, 6 and 7 points, respectively. The predictive accuracy of risk stratification depends on the prevalence of stroke in the population studied.11 In a large Swedish AF cohort, ATRIA slightly outperformed CHA2DS2VASc, with areas under the curve of 0.708 and 0.690, respectively.12 While this difference was statistically significant, both scores remain imperfect predictors of risk of stroke. Other clinical parameters have also been associated with stroke in AF but are not commonly used clinically or included in risk assessment tools. The role of catheter ablation in reducing stroke is another variable that requires further study especially in light of the findings of the EAST-AFNET 4 study.13

Atrial high rates in patients with CIED

In CIED patients, temporal dissociation between device-detected atrial high rate episodes and thromboembolic events has been observed, with most patients lacking atrial arrhythmia episodes 30 days before thromboembolic events.5 6 While risk of stroke associated with device-detected episodes seems lower than that associated with clinical AF, this temporal dissociation questions whether arrhythmias are essential in cardioembolic stroke. The need for chronic anticoagulation in device-detected arrhythmia episodes is currently being addressed by two clinical trials: ARTESiA (Apixaban for the Reduction of Thrombo-Embolism in Patients with Device-Detected Sub-Clinical Atrial Fibrillation; NCT01938248)14 and NOAH (Non-Vitamin K Antagonist Oral Anticoagulants in Patients with Atrial High Rate Episodes; NCT02618577).15

Arrhythmia monitoring and anticoagulation in ESUS

AF is highly suspected in cryptogenic stroke and ESUS. Extended rhythm monitoring for 4 weeks and 3 years demonstrates an increase in the yield (10% and 30%, respectively) of detection of AF episode lasting ≥30 s.7 8 With these results, the majority of patients who had ESUS are found without detected AF over a 3-year period, casting further doubt on the causal relationship between AF and stroke.

Non-selective anticoagulation for secondary stroke prevention in patients who had ESUS has been studied in two clinical trials. NAVIGATE-ESUS (Rivaroxaban for Stroke Prevention after Embolic Stroke of Undetermined Source trial; NCT02313909) was terminated prematurely due to futility, since daily rivaroxaban (15 mg) was associated with increased risk of bleeding without significant decrease in recurrent stroke.16 Likewise, RESPECT-ESUS (Dabigatran for Stroke Prevention after Embolic Stroke of Undetermined Source; NCT02239120) showed that treating patients who had ESUS with dabigatran 110 mg or 150 mg two times per day did not reduce risk of recurrent stroke.17 Post-hoc analysis of NAVIGATE-ESUS demonstrated that patients with enlarged left atrium (LA) diameter, a marker of AF, benefited from anticoagulation, indicating a role of atrial remodelling in identifying which patients who had ESUS may benefit from secondary prophylaxis with oral anticoagulation.18 ARCADIA (Atrial Cardiopathy and Antithrombotic Drugs in Prevention After Cryptogenic Stroke; NCT03192215) is currently enrolling patients who meet the criteria for atrial cardiopathy, including increased LA diameter measured by echocardiography, increased P wave terminal force in V1 or increased serum N-terminal pro-B-type natriuretic peptide for randomisation to apixaban versus aspirin.19 ATTICUS (Apixaban for Treatment of Embolic Stroke of Undetermined Source; NCT02427126) is also enrolling patients who had ESUS for randomised evaluation of apixaban versus aspirin, but with more selective inclusion criteria than ARCADIA and a primary endpoint of examining brain lesions using MRI.20 Table 2 summarises the ESUS trials, including the entry criteria and primary outcomes.

Table 2

Studies of anticoagulation in embolic stroke of undetermined source

Fibrosis as a nexus between arrhythmia and stroke: understanding relationships to electrical remodelling, electromechanics and haemodynamics

Shortcomings in stroke risk stratification tools for patients with known or suspected AF reflect gaps in the current mechanistic understanding of thrombogenesis in these settings. These limitations have prompted a more comprehensive evaluation of the atrial substrate, specifically atrial fibrosis, as an underlying cause of both AF and stroke.21 22 This section summarises our understanding of the relationships between these factors and describes computational methods for studying arrhythmogenesis and thrombogenesis. One approach is computational simulation of the electrical substrate based on atrial fibrotic remodelling. The second is the computational modelling of blood flow inside the LA. The link between these approaches is that atrial haemodynamics depend on wall kinetics, which are governed by electrical activation and contraction mechanics of the atrial wall. Computational modelling of these phenomena could facilitate identification of thrombogenesis determinants and further our understanding of disease mechanisms. These modelling-based approaches constitute an update to the classic Virchow’s triad for thrombosis in vascular beds.

Association of fibrotic remodelling and AF

Fibrosis results from myocyte loss and increased collagen deposition, altering tissue architecture and electrophysiology. Fibrosis quantified by late-gadolinium enhancement (LGE)-MRI is associated with stroke and atrial thrombosis in AF.23 24 Biomarkers of increased collagen deposition and crosslinking have also been associated with AF and heart failure.25

The association between atrial fibrosis and stroke is supported by studies relating atrial enlargement, dysfunction measured by MRI and abnormal electrocardiographic indices such as P wave terminal forces with stroke independent of AF presence.26–28 Fibrosis measured using LGE-MRI was also shown to increase in patients who had ESUS compared with control patients. Also notable was that patients who had ESUS demonstrate a similar amount of fibrosis as patients with known AF (figure 1).29 Fibrosis is therefore emerging as a likely unifying mechanism for stroke and AF. Alternatively, fibrosis may also be a marker of different pathophysiological processes linked to stroke in patients with AF.

Figure 1

Atrial fibrosis (AFib) in embolic stroke of undetermined source (ESUS). Atrial fibrosis is higher in patients who had ESUS compared with age-matched and sex-matched healthy controls and similar to that seen in atrial fibrillation. Panels represent examples of atrial fibrosis (blue: non-enhancing tissue; green: hyperenhanced fibrotic tissue; A: anterior; P: posterior). PA, posterior anterior; AP, anterior posterior. Data are adapted and redrawn from Tandon et al.29

Computational modelling of electrophysiological consequences of fibrosis

Electrophysiology simulation in patient-specific models has emerged as a promising framework for understanding AF and custom-tailoring treatment to each individual’s disease manifestation. Since fibrosis is associated with both arrhythmogenesis and thrombogenesis, computational methods are poised to clarify electrophysiological consequences of atrial remodelling in patients who had a stroke, even those without AF (eg, ESUS). The basis of these simulations is realistic multiscale representation of the heart. At the cellular scale, differential equations are solved to represent ion channel gating kinetics and other intrinsic processes contributing to atrial action potential dynamics. In AF models, parameters are modified to reflect changes associated with disease and fibrotic remodelling. Figure 2A shows AF variants of a human atrial model,30 based on clinical data and in vitro models of fibrosis.31 Both model variants have abbreviated action potential duration compared with normal myocytes; the fibrotic variant has slower upstroke and prolonged refractoriness compared with the non-fibrotic variant. At the tissue scale, excitation propagation is simulated in a finite element model that represents myocardium as a resistive continuum,32 with spatially heterogeneous anisotropic conduction arising from regions of fibrotic remodelling and local cardiac fibre orientations (figure 2B). Since the latter cannot be obtained via in vivo imaging,33 model reconstruction involves mapping fibre orientations from a human atlas geometry.34 At the organ scale, patient-specific anatomies from clinical LGE-MRI scans are used to build three-dimensional (3D) computational models (figure 2C); areas of fibrotic remodelling are delineated from images and used to calibrate regional modification of cell-scale and tissue-scale parameters. Simulated arrhythmia dynamics in such models correlate well with those observed clinically via electrocardiographic35 or intracardiac mapping.36

Figure 2

Multiscale framework for simulation of atrial electrophysiology. (A) Top: schematic showing scaling factors used in recent simulation studies to represent changes due to chronic AF in the absence (blue) or presence (green) of fibrotic remodelling. Bottom: membrane voltage over time traces showing action potentials elicited by a decremental train of electrical stimuli in simulated non-fibrotic (blue) and fibrotic (green) atrial cardiomyocytes; the response to pacing in the baseline (ie, non-AF) version of the model is also shown (dashed black line). Basic cycle length of pacing ramps down from 300 to 200 ms in 20 ms steps. (B) Activation map showing excitation of a fibrotic atria model in response to electrical stimulation. Major features arising from tissue-scale modelling components are highlighted in yellow: anisotropic propagation results from mapping of myocardial fibres; slow conduction (and in extreme cases conduction block) results from decreased conductivity and ionic remodelling (see A) in fibrotic regions. (C) Example of atrial LGE-MRI scan (left) prior to creation of a personalised three-dimensional finite element model (right) including patient-specific regions of fibrotic remodelling (ie, hyperenhanced LGE regions). AF, atrial fibrillation; AFib, atrial fibrosis; LGE-MRI, late-gadolinium enhancement-MRI; LIPV, left inferior pulmonary vein; LSPV, left superior pulmonary vein; RSPV, right superior pulmonary vein; RIPV, right inferior pulmonary vein.

Computational modelling of atrial electrophysiology has revealed key mechanistic insights on arrhythmia perpetuation. LGE-MRI acquisition in patients with AF (an increasingly widespread practice) enables creation of virtual cohorts that reflect interpatient heterogeneity in extent and distribution of fibrosis. One study used models from 20 patients with persistent AF to demonstrate that re-entrant drivers dynamically localise to fibrotic tissue boundaries with a particular spatial pattern of remodelling37; this was later corroborated via clinical findings.38 Modelling has also been used to explain how epicardial fibrosis can cause transmural dissociation and the simultaneous presence of focal and re-entrant activity.39 The multiscale nature of the approach means that subcellular effects can be linked to emergent pathophysiological properties at the organ level. For example, incorporation of remodelling caused by Pitx2 knockout in simulations revealed that this AF-associated genetic variant has a dual arrhythmogenic effect, increasing both triggered activity and vulnerability to re-entry initiation.40

Electrophysiology modelling can also be used translationally to guide clinical procedures. In the OPTIMA (Optimal Target Identification via Modelling of Arrhythmogenesis) study, 10 sequentially recruited patients with persistent AF received a set of ablation lesions custom-designed based on simulations based on each individual’s preprocedure LGE-MRI.41 The study reported promising clinical outcomes, with 9 of 10 patients in sinus rhythm at the end of 309 (245–386) days of follow-up. A randomised clinical trial comparing OPTIMA with pulmonary vein isolation is now underway (NCT04101539). For simulation-guided treatment in the CUVIA-AF1 trial (Clinical Usefulness of Virtual Ablation of AF; NCT02171364), five virtual ablation templates (four linear lesions, one electrogram-based) are tested in each model and the one that terminates re-entrant arrhythmia most rapidly is carried out clinically.42 Preliminary results are compelling, with lower AF recurrence in the CUVIA-guided group (20.8%; n=53) compared with the empirical ablation arm (40%; n=55) based on clinician preference.

Recent cardiac LGE-MRI studies29 have enabled use of computational modelling to understand potential electrophysiological consequences of atrial fibrosis in patients who had ESUS.43 Preliminary findings indicate that patients who had ESUS and AF are indistinguishable in terms of spatial distribution of fibrosis and re-entry inducibility. This suggests a subset of patients who had ESUS might have preclinical fibrotic AF substrate; our initial speculation is that their lack of AF is attributable to other factors such as a low rate of triggered activity, but this has not yet been proven.

Computational fluid dynamics analyses of LA haemodynamics

Blood flow dictates the transport of enzymes, coagulation factors and platelets to and from endothelial damage sites.44 Blood stasis permits these species to initiate clot formation. Assessing left ventricular (LV) blood stasis can predict risk of LV thrombosis in patients with myocardial infarctions45 and cerebral microembolism in a large animal model of acute myocardial infarction.46 However, there is still an unmet need for personalised assessment of blood stasis and stroke risk in patients with AF.

Computational fluid dynamics (CFD) analysis solves the equations of fluid motion inside blood vessels or cardiac chambers, using patient-specific images (figure 3A) to establish boundary conditions (figure 3B). With proper numerical methods and imaging data, CFD calculates four-dimensional (4D) (3D in space, time-resolved) LA blood velocity fields (figure 3C, left) that closely match measurements by phase-contrast MRI or echocardiographic Doppler.47 48 In particular, this analysis can predict stagnant regions by computing blood residence time—the time spent by blood particles along their trajectories—and identifying regions of high residence time (figure 3C, centre). Furthermore, CFD provides high-resolution data about haemodynamic variables like shear stresses (figure 3C, right), which are difficult to access via medical imaging. Finally, CFD analysis offers the possibility to simulate how interventions like left atrium appendage (LAA) occlusion alter atrial haemodynamics,49 which could help personalise intervention planning.

Figure 3

Patient-specific CFD analysis of risk of thrombosis. (A) Time-resolved segmentation and registration of 4D medical images provide patient-specific information about LA wall position and kinematics, PV position, and LV and LA volume changes across the cardiac cycle. (B) This information is used together with a reduced-order model of pulmonary circulation to determine the boundary conditions for blood flow inside the LA, namely the flow rate profiles through the pulmonary veins (QRSPV, QRIPV, QLSPV and QLIPV) and the mitral valve (QMV), and the velocity of the LA wall (Embedded Image ). The boundary conditions are imposed into the CFD solver by a variety of methods (eg, the immersed boundary method). CFD analysis calculates blood velocity fields and blood stasis by solving incompressible Navier-Stokes equations and the residence time equation. (C) These and other (eg, shear rate) haemodynamic variables can be visualised for each patient to identify regions of altered blood haemostasis, where risk of thrombosis is increased. 4D, four-dimensional; CFD, computational fluid dynamics; LA, left atrium; LIPV, left inferior pulmonary vein; LSPV, left superior pulmonary vein; LV, left ventricular; RSPV, right superior pulmonary vein; RIPV, right inferior pulmonary vein.

Being a closed, small chamber bypassed by the main route of atrial blood transit makes the LAA prone to blood stasis. Consistently, 90% of thrombi in non-valvular AF are found in the LAA.50 Interest in simulating LA haemodynamics to describe LAA blood stasis has surged in recent years. There have been two classes of analyses, both involving small patient cohorts. The first class explored how altered atrial wall kinetics affect LA haemodynamics. Some of the most sophisticated simulations modelled AF by prescribing synthetic random motions to patient-specific anatomies;W51, W52 however, patient-specific electro-mechano-fluidic simulations are still lacking. As an alternative, recent studies prescribed realistic LA wall motion captured by high-resolution 4D CT imaging.W53 Researchers have also considered fixed-wall simulations with modified inflow–outflow boundary conditions (eg, suppressed transmitral A wave) as a simplified AF model.W54, W55 Overall, these CFD studies have not conclusively uncovered a causal relationship between impaired atrial wall kinetics and LAA blood stasis. This ambiguity could result from small cohorts analysed so far, but it also suggests that LAA stasis critically depends on additional factors. Indeed, clinical data suggest relationships between LA/LAA morphology and risk of stroke.W56–W58

The second type of analyses have focused on how LA and LAA morphology influence LAA blood stasis. Of note, some of these works singled out LAA morphology from other factors by merging different patient-specific LAAs into a common atrial body for each simulation.W52, W55 Taken together, these CFD analyses suggest that LAA blood stasis increases during AF in a patient-specific manner that often defies the reported correlations between discrete LAA shape categorisations (eg, chicken wing, cactus, windsock, cauliflower)W56 and risk of stroke. This discrepancy highlights the need for more accurate clinical correlates of risk of stroke and the need for CFD validation endpoints that are closer to clinical outcomes than LAA blood stasis.

Multiphysics models and future directions

The outlook for integrating custom-tailored computational models into the clinical diagnosis and management of patients with AF and/or stroke (as illustrated schematically in figure 4) relies on several ongoing transformational advances. First, there is a need for multiscale models to couple the electrophysiology and mechanics of the heart wall with LA haemodynamics. Emerging models can now incorporate wall thickness, fibrosis and fibre organisation to resolve relationships between electrophysiology and wall mechanicsW59 and fluid–structure interactions.W60 In the case of stroke, novel multiscale CFD analyses can calculate the biochemical coagulation cascade, the polymerisation of fibrin and platelet aggregation,W61 as well as embolus transport towards the brain.W62 However, full electro-mechano-chemo-fluidic models that integrate all these factors in patient-specific configurations are still missing. In particular, there are no multiscale models completely incorporating fibrosis into prediction of LA thrombogenesis.

Figure 4

Schematic for envisioned use of modelling and simulation to augment imaging, resulting in better, personalised treatment strategies for patients who had stroke, atrial fibrillation or both. Electrophysiological simulations facilitate detailed assessment of patient-specific consequences of fibrotic remodelling. Computational fluid dynamics simulations enable prediction of thrombus formation and can be further integrated with modelling tools to reflect the coagulation cascade and clot transport towards the brain. Both modelling methodologies integrate medical imaging with measurements from biophysical experiments to produce patient-specific predictions that can be integrated with direct analysis of clinical data to produce better treatment options (eg, custom-tailored drug dosing, recommendations for ablation procedures or appendage closure). LAA, left atrium appendage; LGE-MRI, late-gadolinium enhancement-MRI.

A rigorous computational assessment of risk of LAA thrombosis not only needs to be patient-specific, but also requires uncertainty quantification (UQ) to estimate how input variability propagates forward to model predictions. UQ is relevant for variables like segmented LA geometry; pulmonary veins (PV) inflow profiles; blood viscosity; anisotropy of electrical conduction arising from myocardial fibre orientations; and specific expression levels of relevant ion channels and subcellular components such as calcium handling mechanisms. The latter are subject to interindividual and intraindividual heterogeneities that must be considered when conducting UQ. Input parameters affect excitation propagation, LA wall kinetics, haemodynamics, the kinetics of the coagulation cascade and localisation dynamics of re-entrant activity in AF.44 48 W61, W63, W64 However, patient-specific values are difficult to obtain non-invasively and are thus usually estimated. While there have been significant UQ analyses in the context of CFD for other cardiovascular flows,W65 no formal UQ analyses exist for modelling of LA fluid dynamics and LAA thrombosis risk. Considerable progress has been made in UQ for computational cardiac electrophysiology models, with robust methods for assessing the impact of underlying cell-scale and tissue-scale parametersW66 as well as anatomical uncertainty related to intrinsic limitations of clinical imaging methodologies.W67

Multiscale modelling as discussed above requires significant computing power and UQ requires running multiple simulations per patient to interrogate the space of input parameters, further increasing the computational cost. This poses two crucial barriers for adoption in the clinical setting. First, it limits simulation campaigns to very few cardiac cycles and small patient cohorts. Second, it prevents adoption of new software tools in clinical workflows. We anticipate that advances in hardware and software technologies will lower the adoption barrier in the near future. In parallel, the search for translational breakthroughs should also consider low-cost, reduced-order models of LAA thrombogenesis. These reduced-order models are attractive to make predictions over time scales longer than it would be feasible to simulate with high-resolution models (eg, transitions from paroxysmal to persistent AF, which take place over the course of months/years). Accuracy and computational cost could be balanced by developing nested computational frameworks where a reduced-order model runs over multiple beats, feeding on the results of a high-dimensional, high-resolution model that runs only for a few key beats.

Additional references are provided in the online supplemental file 1.

Supplemental material

References

Supplementary materials

  • Supplementary Data

    This web only file has been produced by the BMJ Publishing Group from an electronic file supplied by the author(s) and has not been edited for content.

Footnotes

  • Twitter @pmjboyle

  • Contributors All three authors substantially contributed to the conception or design of the work, or the acquisition, analysis or interpretation of data; contributed to drafting the work or revising it critically for important intellectual content; provided final approval of the version published; and agree to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

  • Funding This study was funded by The John Locke Charitable Fund.

  • Competing interests None declared.

  • Patient and public involvement Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.

  • Patient consent for publication Not required.

  • Provenance and peer review Not commissioned; externally peer reviewed.