Advanced cardiovascular risk prediction in the acute care setting

The purpose of this amendment is to include additional outcome data from the civil registration dataset. This dataset will give mortality outcomes on the cohort that is the subject of this data sharing agreement.

The additional data fields requested in this dataset are - date of death, date of registration and cause of death.

Date of registration has been included to serve as an alternative data point for date of death when the latter was missing. As mortality is one of the primary outcomes that is being investigated in this study – having this surrogate marker will allow for greater data completeness and more robust analysis.

Secondly, date of registration acts as a validation procedure marker. If there is a substantial amount of time/difference between the two points – then the outcome could be classed as erroneous. Again – this will allow for more robust analysis.

These additional data fields will allow the project team to ascertain the risk of major adverse cardiovascular events, which includes death. When predicting a cardiovascular event, death is an important outcome for patients and clinicians.

---------- Original Application ------------

This application from The University of Manchester seeks to determine whether a diagnostic chest pain algorithm can be updated with machine learning techniques to prevent known loss in accuracy over time.

The purpose of the study is to improve cardiovascular risk prediction in the emergency department (ED). The project seeks to do this in two parts. Firstly University of Manchester aim to maintain and improve an existing acute myocardial infarction clinical prediction model currently in clinical use and in the process validate a method for updating all clinical prediction models. Secondly University of Manchester intend to examine the prognostic value of emergency department data in predicting long term cardiovascular outcomes.

University of Manchester are requesting data on cardiovascular outcomes that occur subsequent to the index attendance at an emergency department. The time frame is of interest is short term and long term (1 - 10 years) for the first cohort and only short term (1 year) for the second. This will enable the use of machine learning to update the diagnostic algorithm, and also assess the prognostic ability of ED data for long term cardiovascular outcomes.

The Troponin-only Manchester Acute Coronary Syndromes (T-MACS) decision aid was derived by the group to improve the early diagnosis of acute coronary syndromes (ACS) or ‘heart attack’. T-MACS uses data available when patients arrive in the Emergency Department (ED) to calculate the probability of ACS. Patients are then assigned to four risk groups including a ‘rule-out’ group that can be immediately discharged, and a ‘rule-in’ group that can receive early treatment. T-MACS has been validated extensively and has been used at Manchester University NHS Foundation Trust (trust in the geographical area of the project) since June 2016, avoiding unnecessary hospital admission for approximately two thirds of patients. T-MACS won the MFT Transformation Prize (2016) and will be implemented across Greater Manchester as a Health Innovation Manchester exemplar project.

All clinical prediction models degrade over time, in a process called calibration drift. In this work University of Manchester will seek to update and improve the TMACS algorithm, two methods will be used; a single ‘one-off’ update and a method that continuously updates itself. This will enable the algorithm to be continually refined and optimised, avoiding the problem of ‘calibration drift’ due to changes in clinical practice and patient demographics. This will promote patient safety (by reducing the potential for missed diagnoses); enable patients to benefit from more accurate diagnoses at the earliest opportunity; will facilitate increasingly personalised healthcare; will reduce the need for future, expensive and prolonged clinical research studies to update diagnostic algorithms; and will prevent over-use of healthcare resources.

Emergency medicine has been at the forefront of the implementation of clinical prediction models. The need for timely, accurate exclusion of high-risk pathologies combined with an NHS wide efficiency drive has been adeptly met by such tools. Current extensively used risk assessment tools include the Wells’ Score for deep vein thrombosis and pulmonary embolism, the CURB-65 score to aid the treatment of community acquired pneumonia(6), Ottawa ankle rules to guide the use of x-rays in suspected ankle fractures, and also the Canadian C-spine rule to determine if it is possible to clinically exclude a cervical spine fracture.

These prediction models are derived and validated in peer-reviewed papers, before being implemented clinically and audited for clinical accuracy and impact. The populations being similar and comparable is an integral part of each model’s applicability to the individual urgent care settings. However, this static comparability is an inherent flaw, as the model cannot be tailored to each site accounting for different patient populations, further it is static and therefore cannot adapt to changes in evolving diagnostic technology through time.

Classical statistical and modern machine learning methods have previously been proposed to overcome such limitations machine learning would allow prediction models to be updated (re-derived) without human input. This has several advantages over classical statistical methods as it can be done quickly, continually, and accurately/ other clinical prediction models have been updated using similar methodologies. EuroSCORE is a risk prediction model used in European cardiac surgery which was shown to demonstrate calibration drift due to changing demographics.

The updating of these models is often done from large cumulative data sets from multiple different hospitals. The use of these collective datasets poses challenges of comparability, different demographics in the larger hospitals may make the clinical prediction model at a smaller hospital more inaccurate than if it had just used its own data. Wiens et al explored this issue and the different methods to combat it. They investigated the use of only combining data for analysis when similar features were available from all sites. They suggest that such an analysis is at its greatest strength when it has more shared data features. This would be the case for Greater Manchester -TMACS as the tool would collect similar data from all sites, therefore it would have almost identical data features making it a near perfect data set for such a cumulative analysis.

Long term cardiovascular risk prediction

In a pilot trial comparing T-MACS to standard care, University of Manchester evaluated patient satisfaction. While overall satisfaction was high (mean overall score 3.78/5), patients gave lower ratings (mean 2.78/5) for “advice you got about ways to avoid illness and stay healthy”. Patients are dissatisfied with an approach that simply informs them that they ‘do not have ACS’ but that does not address future cardiovascular risk. This sentiment was echoed by two patient groups. Whereas such tasks may previously have fallen to inpatient teams, the widespread use of early rule out strategies means that emergency physicians must increasingly bear responsibility for informing patients of their future risk.

There were 23.4 million patient presentations to UK Emergency Departments (ED's) in 2016, and this has been increasing by 10% each year. Patients who do not see their General Practitioner (GP) frequently are more likely to attend the ED, meaning that ED's are interacting with a portion of society under-served by primary care. Patients now expect clinical staff in the acute care setting to have tools to inform them of their long-term cardiovascular risk.

Cardiovascular disease (CVD) remains the leading cause of premature death in the western world. Primary prevention to reduce mean blood pressure and cholesterol by 10% could reduce the incidence of major CVD by 45%. However, because important risk factors for CVD such as hypertension and hyperlipidaemia are usually asymptomatic, identification of at-risk individuals can be challenging.

While in the ED, all patients with suspected ACS will have vital signs recorded. However, these data are not currently used to identify patients at risk of CVD, which represents an important missed opportunity. Previous research has demonstrated that patients with hypertension in the ED have over 90% probability of having persistently elevated blood pressure in the community setting. Contrary to popular belief, hypertension in the ED cannot be wholly attributed to pain or anxiety. Recent studies have shown that hypertension measured in the ED is predictive of 10-year adverse cardiovascular outcomes.

Furthermore there is limited evidence from Farkouh et al, that acute chest pain algorithms are predictive of long term cardiovascular disease. Farkouh et al demonstrated that patients deemed high risk had a hazard ratio of 2.45 (95% CI 1.67-3.58) for cardio and cerebro-vascular events at a follow up of 7.3 years.

In primary care, the QRISK-2® (or QRISK-3®) tool is routinely used to predict patients’ 10-year risk of CVD. If the 10-year risk exceeds 10%, the National Institute for Health and Care Excellence (NICE) recommends that statin therapy should be considered. A range of other measures (advice on smoking cessation, weight loss, diet, exercise and review of comorbidities) should also be undertaken.

This tool could potentially be used in the ED because most of the data required are already routinely collected. This could identify patients at high risk of CVD who would otherwise have been unidentified. Thus patients would not only be better informed but University of Manchester could prevent more incident CVD.

Patients in the ED have different characteristics to those attending for screening in primary care. University of Manchester must validate prediction model in this setting before use.

Legal Basis for Processing under GDPR Justification:

Article 6(1)(e) - public interest - The accurate prediction of heart attacks in the emergency department is vital as it is such a high risk condition. Currently a diagnostic algorithm (T-MACS) is deployed across Greater Manchester, and University of Manchester are seeking to maintain and improve its accuracy. Unfortunately, diagnostic algorithms like T-MACS also have some important disadvantages. Over time, they tend to become less accurate because important factors change including the age of patients, the number of long-term medical conditions they have, the tests that are used and the ways in which healthcare workers practice. This means that the research must be repeated, which is inefficient and expensive.

Cardiovascular disease is the leading cause of premature death in the western world, and unfortunately Greater Manchester currently has double the national average of preventable cardiovascular deaths. Emergency Departments present an opportunity with; increasing footfall, pre-existing data collection systems, teachable moments around chest pain presentations and interacting with a portion of society that does not regularly interact with primary care.

Article 9(2)(j) - The accurate prediction of heart attacks in the emergency department is vital as it is such a high risk condition. Currently a diagnostic algorithm (T-MACS) is deployed across Greater Manchester, and University of Manchester are seeking to maintain and improve its accuracy. Unfortunately, diagnostic algorithms like T-MACS also have some important disadvantages. Over time, they tend to become less accurate because important factors change including the age of patients, the number of long-term medical conditions they have, the tests that are used and the ways in which healthcare workers practice. This means that the research must be repeated, which is inefficient and expensive.

The study has been approved by the research ethics committee and the confidentiality advisory group

This whole project will form part of a PhD thesis. Only summary level data with small numbers suppressed in line with the HES Analysis Guide will be reported and no record level data will be published. The person undertaking the PhD is a substantive employee of the University of Manchester.

This project will handle routinely collected data from patients who had presented to the emergency department with chest pain.

University of Manchester plan to analyse short and long term cardiovascular outcomes; short term outcomes to refine and prevent the degradation of an acute coronary syndrome rule out strategy and long term outcomes to identify predictors for long term cardiovascular disease in the acute care setting.

University of Manchester plan to include two separate patient cohorts, which will enable the University of Manchester to maximise the value of the work, as follows:

To evaluate short-term outcomes (refinement of the T-MACS mathematical algorithm) University of Manchester intend to use a cohort from 2016 to the present day. This will include patients presenting to Emergency Departments across Greater Manchester. These patients have full data for T-MACS (required to optimise short-term risk prediction), and the follow-up period (12 months) is sufficient for that analysis. (Estimated 16000 patients).

University of Manchester will use the data to optimise the machine learning approach to refine the T-MACS algorithm.

University of Manchester will send to NHS digital two cohorts of data, both will contain patient identifiers and the date and time of index event (attendance to an emergency department with the presenting complaint of chest pain).

University of Manchester intend to identify the optimal predictors of long-term cardiovascular disease among patients presenting to the Emergency Department with chest pain. To do this University of Manchester will use cohorts that cover the introduction of conventional cardiac troponin and high sensitivity cardiac troponin. This long term outcome cohort (Cohort 1) will be exclusively from Manchester University NHS Foundation Trust and match the time periods of covering the introduction of the technology (2009-2010, 2011-2012 , 2016-2017). The cohort will consist of patients who presented to the adult Emergency Department with chest pain and will include any patient over 18.

Summary:

Cohort 1 - 10 year outcome cohort

Contributing NHS trust: Manchester University NHS Foundation Trust

Estimated sample size: 21,000

Data requested from NHSD: diagnostic & intervention codes, codes including cardiovascular death, acute myocardial infarction, stroke or coronary revascularization, mortality data if applicable.

Multiple NHS trusts are providing Manchester University NHS Foundation Trust with data as part of a Greater Manchester service improvement programme.

This data forms part of the second cohort, for cross-linking. NHS trusts to include Manchester University NHS Foundation Trust, Stockport NHS Foundation Trust, Wrightington Wigan and Leigh NHS Foundation Trust, and East Lancashire Hospitals NHS Foundation Trust will send the data to University of Manchester who will then send onto NHS Digital.

Cohort 2 - 1 year outcome cohort.

Contributing NHS trusts - Manchester University NHS Foundation Trust, East Lancashire Hospitals NHS Foundation Trusts, Stockport NHS Foundation Trust, and Wrightington Wigan and Leigh NHS Foundation Trust

Estimated sample size: ~ 15,000

Data requested from NHSD: diagnostic & intervention codes, readmission [1 year from index event] - all intervention and diagnostic codes, mortality data if applicable.

For the short term outcome at one year from index event the diagnostic and intervention codes are required regardless of it is cardiac in nature, this is in order to understand the reasons for re-attendances and to ascertain if it is due to an inaccuracy in the index admission diagnosis.

For all the cohorts the data controller and processor is the University of Manchester. The contributing NHS sites will transfer the identifiable data to the university. NHSD will only receive and return data to the Data Safe Haven at the University of Manchester. Data will be returned with a study ID in place of the identifiable variables.

University of Manchester believe that these data sets are the minimum required to successfully gain the aforementioned data points

Once cross linked University of Manchester intend for the data to be pseudonymised. Identifiers will be passed to NHS Digital to facilitate the linkage - but all clinical/mortality data will be sent back to the University of Manchester with identifiers removed, and a unique study-ID in place of the identifiers.

The University of Manchester require 10 years of data retrospectively. 10 years is the standard for long term cardiovascular outcomes as mentioned in the literature.

The data has already been geographically restrained due to the cohort only being from Greater Manchester emergency department attenders.

The project group does not believe that there is an alternative method to gain outcome data for patients seen ten years ago.

The data controller and the data processor is the University of Manchester.

Funders

- National Institute of Health Research - Doctoral Research Fellowship

- Manchester University NHS Foundation Trust - Data Driven Healthcare award

All organisations party to this agreement must comply with the Data Sharing Framework Contract requirements, including those regarding the use (and purposes of that use) by "Personnel" (as defined within the Data Sharing Framework Contract i.e.: employees, agents and contractors of the Data Recipient who may have access to that data).

University of Manchester will send to NHS digital two cohorts of data, both will contain patient identifiers (NHS number, Date of Birth, Postcode and the Study ID) and the date and time of index event (attendance to an emergency department with the presenting complaint of chest pain).

The research only requires data since the patient was seen in an emergency department with chest pain, as such the University of Manchester have only requested data within this time frame.

The data has already been narrowed by geography as only the sites whom are using the troponin-only Manchester acute coronary syndrome diagnostic algorithm have been selected. This includes a selection of trusts within Greater Manchester.

In order for the results of the research to be generalisable University of Manchester are not seeking to narrow the demographics. This will ensure that the research and potential improvement in diagnostic accuracy is of benefit to the widest patient population.

University of Manchester are seeking to ascertain if any patients treated in emergency departments for chest pain re-attend and are subsequently diagnosed with a heart attack or undergo revascularisation.

The cohort has already been pre-selected as chest pain presenters to the emergency department. For the short term outcome at one year from index event the diagnostic and intervention codes are required regardless of it is cardiac in nature, this is in order to understand the reasons for re-attendances and to ascertain if it is due to an inaccuracy in the index admission diagnosis. For the long term outcome a cohort could be created that extracted only codes for myocardial infarction, revascularisation, and stroke.

NHS digital will link the cohort with the HES Admitted Patient Care Dataset. The data is then to be pseudonymised and transferred securely to the University of Manchester's Data Safe Haven.

Under V1 of this agreement, NHS Digital will additionally link the patient cohort (both cohort 1 and cohort 2) to the Civil Registration Secondary Care Cut dataset - so that data on the mortality outcomes of the relevant participants can be provided to University of Manchester. Only three variables are required - date of death, date of registration, and cause of death. A study ID will be provided alongside the requested data - no identifiers will accompany the mortality data. A

Data transferred to NHS Digital will contain identifiable information, data returned to the University of Manchester will be pseudonymised. NHS Digital will remove identifiers and replace with a study ID before release.

On receipt the data will be linked back to the clinical data set with unique study ID numbers, once linked the study id will be removed.

NHS trusts including Manchester University NHS Foundation Trust, Stockport NHS Foundation Trust, Wrightington Wigan and Leigh NHS Foundation Trust, and East Lancashire Hospitals NHS Foundation Trust will send the data to University of Manchester for the cohorts.

University of Manchester - will receive and store the data from NHS Digital. Data will be transferred to NHS Digital from The University of Manchester using Secure Electronic File Transfer (SEFT) and received from NHS Digital via SEFT also.

In summary the value of known prognostic factors will be evaluated for predicting long term cardiovascular outcomes, Short term cardiovascular outcomes will be used to update the current clinical prediction model and ascertain the optimum method for updating models.

The data will be pseudonymised prior to being returned to the University of Manchester. No individual level data will be published in finalised outputs - all data will be presented in aggregated results with small numbers suppressed in line with the HES Analysis Guide.

University of Manchester can confirm that there will be no attempt to re-identify individuals.

The data will only be processed by University of Manchester staff and doctoral students. The doctoral students are substantive employees of the University of Manchester.

All of the research team have been trained in data protection and confidentiality as per the University of Manchester's guidelines.

The data will be held and processed within the University of Manchester Data Safe Haven - a secure virtual machine environment with fully auditable logs.

The data will not be removed from the University of Manchester's Data Safe Haven, it will be securely deleted from it once it has been processed and the project finalised.

The outputs of the research will include:

(a) The results of the data processing will form part of the doctoral work of a National Institute for Health Research Doctoral Research Fellow. As part of the funding arrangements from the National institute of Health Research's funding requirements the results will also form part of regular reports to them.

(b) The findings will be submitted for publication in peer reviewed journals. This is likely to involve publication of (a) the refined statistical methodology for updating models, (b) updating and maintaining the diagnostic algorithm, (c) the prognostic value of routinely measured data for long term cardiovascular risk prediction in ED . The primary target audience for the clinical study will be emergency medicine physicians, GP's, acute medicine physicians, cardiologists, clinical biochemists, public health professionals and industry leaders in acute diagnostics. However, given the novelty of this work it is likely to appeal to a wide general medical readership. Therefore, submission will primarily strive to achieve publication in a high impact medical journal such as the New England Journal of Medicine or the Lancet.

(c) Presentations at international and national conferences with relevant target audiences will be sought (e.g. European Society for Emergency Medicine Annual Congress, European Society of Cardiology Annual Conference, Royal College of Emergency Medicine Annual Scientific Conference). In addition, a public engagement strategy will be developed in conjunction with Public Programmes and the patient groups, in order that the local population have the opportunity to learn about the work and to engage with future work.

The data will be reported upon in aggregate and summary statistics, it will not comment specifically on individuals patients. All outputs will contain aggregated results with small number suppression applied - in line with the HES Analysis guide.

This will include a press release to be generated with the host institution, Health Innovation Manchester and the NIHR, to be disseminated in conjunction with at least one patient story. A social media strategy will be developed to enhance the impact of the work, this will be aided by the research team being authors on a popular international blog and podcast. Twitter updates will also be given.

The project intends for the publications to be open access and for the outputs to be generated within 6 months from the end date, which is currently scheduled for 2022.

Promotion of health

It is hoped that the dissemination of the data will enable analysis which with safeguard and improve healthcare. This is through three main mechanisms;

(a) Developing a method for updating and maintain clinical prediction models

(b) Updating and improving T-MACS which is current in clinical practice across Greater Manchester

(c) Allowing emergency departments to prognosticate for long term cardiovascular outcomes in a high risk population.

Public interest

The public benefits greatly from clinical prediction models across all medical practice, but these are all slowly losing accuracy over time. It is hoped that the dissemination of NHS Digital data will allow a method to be validated for safeguarding and protecting these vital tools.

It is hoped that the data will allow for T-MACS to be updated, maintained and improved. This will hopefully enable its continued use to safely, accurately and rapidly diagnose myocardial infarctions for patients attending the emergency department with chest pain.

Identifying and potentially modifying long term cardiovascular outcomes in the emergency department has the obvious benefit of improving the populations health. Furthermore it makes efficient use of NHS services, for the greater benefit to the individual patient’s health.

Dissemination of findings

First, the findings will be submitted for publication in peer reviewed journals. This is likely to involve publication of (a) the validated method for updating clinical models, (b) the updated and improved T-MACS model, and (c) the prognostic ability of routinely collected ED data to predict long term cardiovascular outcomes. The primary target audience for the clinical study will be emergency medicine physicians, GPs, acute medicine physicians, cardiologists, clinical biochemists, public health professionals and industry leaders in acute diagnostics. However, given the novelty of this work it is likely to appeal to a wide general medical readership. Therefore, submission will primarily strive to achieve publication in a high impact medical journal such as the New England Journal of Medicine or the Lancet.

Presentations at international and national conferences with relevant target audiences will be sought (e.g. European Society for Emergency Medicine Annual Congress, European Society of Cardiology Annual Conference, Royal College of Emergency Medicine Annual Scientific Conference). In addition, a public engagement strategy will be developed in conjunction with Public Programmes and the patient groups, in order that the local population have the opportunity to learn about the work and to engage with future work.

This will include a press release to be generated with the host institution, Health Innovation Manchester and the NIHR, to be disseminated in conjunction with at least one patient story. A social media strategy will be developed to enhance the impact of the work.

As part of the planning for this piece of research the study team have already engaged a PPI group to gauge opinion on the research plan. This includes eight patient representatives from the Wythenshawe ëTicker Clubí. This is an organisation that actively represents patients with cardiac pathologies, and consists of patients who have had such pathologies themselves and associated interventions. The group were supportive of the idea, and were particularly pleased that it would be safeguarding the clinical decision rule for the future. There were concerns regarding the communication between the patient and the clinician in conveying the calculated risk of ACS/MACE, particularly now that more autonomy will be given to the AI. However the group were satisfied that with careful explanation and shared decision making models this would be minimised and countered.

The Study Team intend to hold another four meetings through the study period, one at the mid-point of phases one and two, and another at the end of each phase. This will enable the research to gain the most from patient and public involvement.

Impact

This project, if successful, will develop methodology that will protect all clinical prediction models from degradation. Given the widespread use of clinical prediction models this will have a large impact on all areas of clinical practice, ensuring that the accuracy and safety of these clinical tools maintained. This will inform the wider scientific community in how best to avoid calibration drift in other clinical prediction models. It is hoped there will also be the improvement of the TMACS algorithm increasing accuracy, benefiting patient care and safeguarding it against future calibration drift.

If the findings are positive implementation strategy will be developed in conjunction with Health Innovation Manchester, which provides direct access to the Joint Commissioning Board in Greater Manchester.

This study and the results will form part of a PhD thesis for a National Institute of Health Research doctoral research fellow.

Not stated in the register.