Development and external validation of a risk prediction model for falls in patients with an indication for antihypertensive treatment: retrospective cohort study

Abstract Objective To develop and externally validate the STRAtifying Treatments In the multi-morbid Frail elderlY (STRATIFY)-Falls clinical prediction model to identify the risk of hospital admission or death from a fall in patients with an indication for antihypertensive treatment. Design Retrospective cohort study. Setting Primary care data from electronic health records contained within the UK Clinical Practice Research Datalink (CPRD). Participants Patients aged 40 years or older with at least one blood pressure measurement between 130 mm Hg and 179 mm Hg. Main outcome measure First serious fall, defined as hospital admission or death with a primary diagnosis of a fall within 10 years of the index date (12 months after cohort entry). Model development was conducted using a Fine-Gray approach in data from CPRD GOLD, accounting for the competing risk of death from other causes, with subsequent recalibration at one, five, and 10 years using pseudo values. External validation was conducted using data from CPRD Aurum, with performance assessed through calibration curves and the observed to expected ratio, C statistic, and D statistic, pooled across general practices, and clinical utility using decision curve analysis at thresholds around 10%. Results Analysis included 1 772 600 patients (experiencing 62 691 serious falls) from CPRD GOLD used in model development, and 3 805 366 (experiencing 206 956 serious falls) from CPRD Aurum in the external validation. The final model consisted of 24 predictors, including age, sex, ethnicity, alcohol consumption, living in an area of high social deprivation, a history of falls, multiple sclerosis, and prescriptions of antihypertensives, antidepressants, hypnotics, and anxiolytics. Upon external validation, the recalibrated model showed good discrimination, with pooled C statistics of 0.833 (95% confidence interval 0.831 to 0.835) and 0.843 (0.841 to 0.844) at five and 10 years, respectively. Original model calibration was poor on visual inspection and although this was improved with recalibration, under-prediction of risk remained (observed to expected ratio at 10 years 1.839, 95% confidence interval 1.811 to 1.865). Nevertheless, decision curve analysis suggests potential clinical utility, with net benefit larger than other strategies. Conclusions This prediction model uses commonly recorded clinical characteristics and distinguishes well between patients at high and low risk of falls in the next 1-10 years. Although miscalibration was evident on external validation, the model still had potential clinical utility around risk thresholds of 10% and so could be useful in routine clinical practice to help identify those at high risk of falls who might benefit from closer monitoring or early intervention to prevent future falls. Further studies are needed to explore the appropriate thresholds that maximise the model’s clinical utility and cost effectiveness.

1 Extended methods

Sample size
Our pre-specified sample size calculation for model derivation assumed a falls rate of 88.3 per 10,000 person years (52), an expected median follow up of 7 years, an estimate of Nagelkerke's R 2 statistic of 0.15 and a maximum number of 40 predictor parameters in the model. Based on these inputs, a sample size of approximately 15,358 person-years (2,194 participants) was estimated to be required for the development of this prognostic model (23).
In our analysis, we in fact considered a total of 44 predictor parameters, which instead required a sample size of approximately 16,898 person-years (2,414 participants). Our development sample size in CPRD GOLD far exceeded this.
For the external validation, our sample size calculation assumed an approximate skewnormal distribution for the model's linear predictor with a mean of 0.41, a variance of 0.8, a skewness parameter of 1, and a kurtosis parameter of 4; an assumed exponential distribution of survival times, with baseline rate parameter 0.0042 to ensure 91% survival at 10 years, and a constant censoring rate with censoring times following an exponential distribution with a rate parameter of 0.096 (to a give a probability of censoring by 10 years of about 85%). This resulted in a sample size of 12,000 patients (experiencing approximately 708 falls) which would be sufficient to target a 95% confidence interval of width 0.2 around the estimate of the calibration slope (24). Once again the sample size in the CPRD Aurum validation cohort far exceeded this.

Assessment of model performance across GP surgeries
Heterogeneity in model performance across different GP surgeries was assessed using a random effects meta-analysis, using restricted maximum likelihood estimation (REML), given that the case mix and incidence of falls were expected to vary between practices (40).
Confidence intervals for pooled estimates were derived using the Hartung-Knapp-Sidik-Jonkman variance correction, to account for any uncertainty in the between-practice variance (53). Performance estimates were first calculated for each GP practice, within each imputed dataset. Then estimates were combined across imputations, within GP practice by applying Rubin's Rules to combine statistics (and corresponding standard errors) where appropriate. Finally, estimates were combined across practices using meta-analysis to gain estimates of average model performance across all populations (54).

External validation performance of the original model
Upon external validation, the original model showed excellent discriminative, with pooled Cstatistics of 0.843 (95% CI: 0.841 to 0.844, 95% PI: 0.789 to 0.881) and 0.833 (95% CI: 0.832 to 0.835, 95% PI: 0.789 to 0.870) at 5 and 10 years respectively ( The original model showed poor calibration at 5 and 10 years, where the model underpredicted the risk of a fall in those with lower predicted probabilities, while over-predicting the risk in those with higher predicted probabilities ( Figure S1, Table 3). The original model also showed overprediction of falls risk at 1 year, across the full range of predicted probabilities       People are living for longer, with more long-term physical and mental conditions which worsen their health. One example is high blood pressure, where people can take 3-4 drugs to prevent stroke. However, dozens of people have to be treated for at least a year to prevent a stroke in one person. This is because these drugs only reduce the possibility of stroke, they do not remove it altogether. Some of these patients may be prone to side effects such as falls and kidney problems which may be more common than any benefits.
This proposal aims to use information from the medical records from hundreds of thousands of patients to establish the link between drugs used to prevent heart attack and stroke and side effects. We will focus on drugs that lower blood pressure (known as antihypertensives), lower cholesterol (known as statins) or prevent blood clots (known as antiplatelets). This information will be used to develop a calculator which predicts a person's risk of experiencing side effects. This calculator will form part of a support tool which will help patients and doctors make better informed decisions about starting or continuing drugs.

Background
The population is ageing and consequently, the number of people living with age-related chronic conditions is increasing. Polypharmacy (five or more prescribed medications) is common in older people and is associated with an increased risk of adverse drug reactions. Preventative medications, such as those used to manage blood pressure and cholesterol, are common in polypharmacy and often require large numbers of people to be treated to prevent a small number of cardiovascular disease (CVD) events. This leaves many individuals on drugs of little benefit, some of whom may be susceptible to side effects such as falls, kidney problems and muscle pain.

Aims
This proposal aims to quantify the harms of cardiovascular prevention medication, and the characteristics of those people most likely to suffer them. This research is one part of a larger research programme to develop a clinical decision tool which estimates an individual's likelihood of benefiting or suffering harm from treatment.

Methods
Aim 1: Derive prognostic models for an individual's risk of adverse events associated with cardiovascular prevention treatment (antihypertensives, statins and antiplatelets) using data from the CPRD GOLD.
Aim 2: Externally validate each model using data from the CPRD Aurum. Specific outcomes will be examined in relation to exposure to either antihypertensives, statins or antiplatelets (details given in section N).

Objective
The overarching objective of this study is to better understand the harms of preventative treatments (antihypertensives, statins and antiplatelets) by quantifying an individual's baseline risk of harm and the modifying effect of treatment.

Specific aims
Aim 1: Derive prognostic models for an individual's risk of adverse events associated with cardiovascular prevention treatment (antihypertensives, statins and antiplatelets). Adverse event outcomes will include falls (antihypertensives), myopathy (statins) and bleeding (antiplatelets).
Aim 2: Externally validate each model. Aim 3: Use causal inference methods (propensity score matching, instrumental variable analysis) to examine whether modification of treatment could have an important impact on the risk of adverse events.

Rationale
The proposed work will develop and validate new prediction tools for an individual's risk of harms from cardiovascular prevention medication. This information will be combined with existing tools for the benefits of treatment and be used to form a new strategy that better targets preventative therapy at those with the most to gain. By empowering patients and clinicians to better understand the risks of preventative medications, this work will promote a more patient-centred, shared-decision making approach to cardiovascular disease prevention in primary care.

F. Study Background
Cardiovascular disease (CVD) is the leading cause of mortality worldwide. 1 As a result, much healthcare resource is assigned to preventing CVD through modification of risk factors, such as raised blood pressure and/or cholesterol. 2 This can be achieved through prescription of medications such as antihypertensives, statins and antiplatelets, which have been shown to be effective in reducing the risk of cardiovascular disease. [3][4][5] These medications are often started when patients are at low-to-moderate risk and continued for many years, despite a low likelihood of benefit. 6 Some patients who take these medications may suffer side effects such as falls, acute kidney injury (AKI), muscle pain and bleeding which can significantly reduce an individual's quality of life, particularly those who are old and frail. [7][8][9] Despite changes in an individual's risk/benefit profile, clinicians are often reluctant to stop prescribing them. 10 The association between cardiovascular preventative therapy and adverse events is not fully understood. This is due in part, to the completeness of reporting of side effects and adverse events in previous trials. 11 12 Observational studies do show an association between treatment and adverse outcomes such as falls, 9 13 14 AKI, 15 myopathy, 16 diabetes, 17 and intracranial bleeding, 18 but are limited in some cases due to small sample sizes and bias caused by unmeasured confounding. Observational data can however be useful since it allows multi-morbid populations to be examined, which are more representative of the general population than those enrolled into clinical trials. 19

Predicting the benefits and harms of treatment
A structured approach to weighing up the risks and benefits of treatment is common in atrial fibrillation, where combining risk scores from the CHA 2 DS 2 -VASc tool 20 and HAS-BLED tool 21 allows one to compare an individual's risk of stroke with their risk of a bleed. There are many other risk prediction tools which can be used to identify individuals who may benefit from CVD prevention treatment (i.e. are at high risk of CVD), 22-28 but very few which can be used to identify those at risk of adverse events in the community. Existing tools show moderate discrimination and focus on patients with specific conditions and/or use in an acute hospital setting. [29][30][31][32][33][34] None have been externally validated. [35][36][37] Recent studies have described prediction models for the benefits and harms of intensive blood pressure lowering treatment, 38 39 but these were derived exclusively from patients in the SPRINT trial, 7 limiting their generalisability.
Stratifying treatments for prevention of CVD is important for UK and international health. Such an approach could facilitate shared decision making between patients and doctors and allow better targeting of treatments in settings where resources and access to drugs may be limited.

How this project fits in
This is one project in a larger work programme aiming to personalise preventative treatments for cardiovascular disease. In this project, we will quantify risks due to cardiovascular preventive treatments. In other projects, including systematic reviews, we will quantify and combine estimates of the benefits and harms of preventive treatments. This will enable us to develop a clinical decision tool to assist patients and their clinicians in understanding the relative benefits and harms of proposed treatment changes.

G. Study Type
Hypothesis testing

H. Study Design
Longitudinal cohort study

I. Feasibility counts
There are approximately 5.5 million patients in England, aged 40+ and with up to standard registration for a period between 1/1/1998 to 31/12/2017 in the CPRD GOLD. Restricting to English practices (linked to HES and ONS) and only including patients fulfilling the eligibility criteria listed below, the expected population is approximately 3.3 million and out of these we expect 216,000 fall events, 230,000 myopathy/muscle pain events and 1,200 intracranial haemorrhage events.

J. Sample size considerations
We have based our sample size calculation on the least common of the primary outcomes listed above (intracranial haemorrhage), and used the method developed by Riley et al., 40 to minimise the potential for overfitting and ensure precise estimates of key parameters. A sample size of approximately 80,000 patients is required for the development of a clinical prediction rule for this outcome, assuming an event rate of 24.6 per 100,000 person years, 41 a median follow up of 7 years, 42 a conservative estimate of Nagelkerke's R 2 statistic of 0.15 and a maximum number of 40 parameters per model.
Feasibility counts showed that this sample size is adequate for the development of the prediction rules. Since Intracranial bleeding is the rarest of the primary outcomes to be studied, the sample size will be sufficient for all other outcomes of interest. For validation, Vergouwe et al., 43 consider at least 100 events per sample population to be "substantial" and we will comfortably exceed this in our validation cohorts for each outcome of interest.
For analyses using causal inference methods (propensity score matching and instrumental variable analysis), assuming clinically significant increased rate of each adverse event with treatment of 10%, and an event rate of at least 0.5% per year in the non-exposed group, approximately 88,380 patients (44,190 in each group) and 4,634 events will be required to accurately define the relationship between preventative treatment and adverse events, with 90% power and an alpha of 0.05.

K. Planned use of linked data (if applicable):
Data linkage to the ONS and Basic HES are required to define the primary and secondary outcomes in the study. The ONS mortality register will be used to define any outcomes which result in death, and also censor follow-up at death. Specific linkages required will include data and ICD-10 coded cause of death (see attached code lists). Linkages to Basic Inpatient HES will be combined with data from the ONS to define all other outcomes in the study. HES data will also be used to define patients' eligibility fo r the study (e.g. previous CVD), and define the study population (e.g. ethnicity where unavailable in Primary Care records). Data required from Basic Inpatient HES will include primary diagnosis, secondary diagnosis, patient characteristics (e.g. sex, ethnicity), date of admission and date of discharge. All deaths and hospital admissions occurring after a patient's index date will be included. A linkage to the Index of Multiple Deprivation is required to acquire patient level quintiles of multiple deprivation, to better define the sample population and use as a covariate in the prognostic modelling.

L. Definition of the Study population
Individual patient data will be extracted from the medical records of all patients registered at linked general practices contributing to the CPRD GOLD and Aurum in England.
Patients will be included if they meet the following criteria: -Patients over the age of 40 years -Registered to a CPRD 'up-to-standard' practice -Records available after the study start date (01/01/1998) -Sufficient data to define the index date (see table below) Patients will be excluded if they meet the following criteria: -Previous prescription of the exposure variable (antihypertensive, statin or antiplatelet) -Blood pressure >180 mm Hg or total cholesterol >7.5 mmol/L (in patients with high blood pressure or cholesterol, treatment is indicated regardless of risk) Study entry criteria -Patients will enter the cohort on the index date, defined as:

Exposure
Antihypertensive therapy Statin therapy Antiplatelet therapy Index date definition 12 months after the first systolic blood pressure reading ≥130 mm Hg 12 months after the first coded or calculated CVD risk of ≥5%

months after first CVD event
Study exit criteria -Last date at which the most recent linked data are available from the CPRD (study end date, July 2018) -Date of the most recent data upload from the practice to which a given patient is registered -Date at which a given patient transfers out of a registered CPRD practice -Date of death or specific outcome of interest M. Selection of comparison group(s) or controls Aim 1: Not applicable -single cohort in CPRD GOLD.
Aim 2: Not applicable -single cohort in CPRD AURUM. Aim 3 will use propensity score matching and instrumental variable analysis -see Methods section below.

N. Exposures, Outcomes and Covariates
This study will examine the overall risk of harm from three types of cardiovascular medication: antihypertensives, statins and antiplatelets. 'Harm' will be defined by the outcomes listed in the table below, which will be examined both individually and as a composite outcome. The association between treatment and a positive and negative control outcome will also be examined to check the validity of the treatment effect estimates. Positive controls are outcomes know to be affected by treatment (e.g. cardiovascular disease) and negative outcomes are those known not to be affected by treatment. All outcomes will be defined according to diagnostic and/or symptom codes, unless otherwise stated. Further sensitivity analyses will explore the impact of our definition of myopathy and muscle pain, focussing on formal diagnostic codes for myopathy only.
Our PPI engagement suggests some patients are concerned about side effects which will result in having to go to hospital, whilst others are more concerned about chronic conditions which affect their daily quality of life. The primary outcomes of this study will therefore be hospital attendance, defined according to ONS and Basic HES datasets. Sensitivity analyses will explore outcomes defined more broadly using primary care codes as well.
All published outputs from this work will be based on the outcomes as specified in this application. Further work with our patient and public involvement representatives will explore definitions of these outcomes and how they should be incorporated into the final calculators developed as part of the wider programme of work related to this project.
Each model will be derived pragmatically, using coded data routinely available in a primary care setting. Potential predictors will be selected based on previous literature and expert opinion. 30 31 44 Predictors will be defined by codes in the primary care records, using code lists from previous studies (via www.clinicalcodes.org). Each model will also incorporate basic patient characteristics, blood pressure level, co-morbidities and frailty, estimated from the electronic frailty index.

Aim 1 -prognostic model derivation (primary analysis)
Where feasible and appropriate, flexible parametric survival models will be derived in the CPRD GOLD for each outcome separately. Baseline survival over time will be estimated using the Royston-Parmar approach of restricted cubic splines on the cumulative log hazard scale. In contrast to Cox regression, this allows a smoothed estimate of the baseline survival function to be derived, and the time period of prediction to be varied, facilitating individualised risk prediction. For the primary analysis, a 10-year period of prediction will be considered. Further sensitivity analyses will be conducted examining a 5-year period. Where insufficient data are available, or modelling procedures become prohibitively complex (computationally), simpler approaches such as Cox regression will be considered.
Each drug class (prescribed in the 12 months prior to the index date) will be entered into the model as a binary variable to allow for multiple treatments to be considered in each model. Where appropriate, models will be constructed with all other variables entered as continuous (not categorised) variables, and potential non-linear trends will be examined. Transformations or fractional polynomials will be used if non-linear trends are detected. Where hazard ratios are not proportional, interactions of effects with time will be included. Further interaction terms, identified in an accompanying systematic review, accounting for treatment effects which differ across specific populations may be included. Where appropriate, models will be reduced and candidate predictors selected for inclusion based on p values of association. Time since cohort entry will be used as the underlying time function in each survival model.
In a further analysis, we will extend the modelling process as described above to account for competing risk of death from other causes. This will produce a model that predicts the risk of adverse events from cardiovascular treatment over time, in the real world where individuals might die from other causes and thus prevent cardiovascular treatment adverse events from occurring. We will again use a flexible parametric modelling framework to estimate the sub-distribution hazard and cause-specific cumulative incidence functions. 45 Missing data will be handled, under a missing at random assumption, using multiple imputation (see below) followed by Rubin's rules to combine parameter estimates across studies.
Where feasible and appropriate, each model will be internally validated using the bootstrap method, to allow optimism-adjusted estimates of calibration and discrimination performance to be obtained. Adjustment for overfitting will be undertaken using a uniform shrinkage factor identified from bootstrapping, followed by re-estimation of the cumulative baseline sub-distribution hazard function (and thus baseline cumulative incidence function) to maintain overall calibration between observed and predicted risks. Given the large sample size, overfitting is likely to be small (see sample size calculation above). If bootstrapping is computationally infeasible due to the large sample size, then a smaller random sample of the population will be used or alternative internal validation methods employed such as cross validation.

Aim 2 -prognostic model validation
External model validation will be undertaken using data from CPRD Aurum. The accuracy of risk predictions from each model will be assessed with calibration plots and 5 and 10 years, calibration measures (slope, E/O statistics), measures of overall fit (using pseudo-R 2 estimates) and discrimination measures (C-statistics and D-statistics).

Aim 3 -causal inference analysis
For most of the outcomes described in this proposal, the association and magnitude of effect with treatment is not well defined in randomised controlled trials. We will therefore undertake causal inference work to examine whether modification of treatment could have an important impact on the risk of adverse events using two causal inference methods: propensity score matching and instrumental variable analysis.
For the propensity score analysis, eligible patients will be matched 1:1 at the index date using propensity scores, which indicate the likelihood a patient will be prescribed treatment on the basis of their known (pre-treatment) characteristics and other known information which might influence the decision to treat. Predictors of preventative drug prescription will be explored in a logistic regression model. Matched patients will be compared using the Cox proportional hazards models, but hazard ratios will only be adjusted by factors unbalanced at baseline which are not already included in the propensity score model. Time since cohort entry will be used as the underlying time function in each model.
For the instrumental variable analysis, GP's previous antihypertensive prescribing preferences will be used as an instrument to predict the likelihood of actual treatment. Risk of adverse events will be estimated in patients attending a GP considered a high prescriber of preventative medication and compared to those attending a GP considered a low prescriber. If the results from the propensity score matching suffer from unmeasured confounding, they will differ from those of the instrumental analysis. If this is the case, these models will be refined with additional factors to achieve better adjustment and matching. Where all approaches provide similar results, we can be more confident that the findings are accurate and reliable. 46 A similar approach will be applied to secondary outcomes, for example, memory loss which has been hypothesised as being linked with statins 47 and is a concern of many patients in practice, but there is no evidence to support this association from trials. 48 We recognise that the evidence base may change for some outcomes over the next few years, and so the need to test the associations between treatment and certain outcomes may change.

Subgroup and sensitivity analyses
Where possible and appropriate, analyses of treatment associations will be examined in subgroups of the population. These will include age, sex, baseline blood pressure and cardiovascular risk, by drug type/dose/intensity and past medical history. Subgroup definitions will be agreed by expert opinion or taken from an accompanying systematic review of previous trials, examining the same topic.
Sensitivity analyses will be undertaken examining risk models for outcomes defined according to both hospital and primary care records. We will also explore the impact of our definition of myopathy and muscle pain, focussing on formal diagnostic codes for myopathy only. Models utilising a 5-year follow-up will be examined as a further sensitivity analysis. The primary analyses will examine the association between treatment and outcomes, assuming treatments are not altered or modified in the future. Where possible, sensitivity analyses will examine treatment entered as time varying covariates, and censor patients who have treatments removed or switched following the index date.
Two approaches will be taken to determine whether observed associations are potentially subject to residual confounding. 46 Where all approaches provide similar results, we can be more confident that the findings are accurate and reliable. Further steps will be taken to examine the validity of treatment effect estimates using positive and negative controls: the impact of treatment on an outcome known to be affected by treatment (e.g. cardiovascular disease; positive control) and outcomes not known to be affected by treatment (negative control). Here, we will combine two common but unrelated outcomes to be examined as negative controls: bowel cancer and chronic obstructive pulmonary disease (COPD). If treatment has a significant impact on these negative controls, it suggests that there is something missing in the analysis (i.e. an unmeasurable factor confounding the treatment effect such as being generally unwell or an unhealthy lifestyle) causing an imbalance between the treatment and control groups, rather than a true treatment effect.

Q. Plans for addressing missing data
There is potential for missing data in this study, particularly with variables such as ethnicity which are recorded with varying degrees of accuracy in routine practice. Because this analysis is focused on treatment for prevention of cardiovascular disease, accurate selection of the sample population is important and therefore any patients with insufficient data available to define their blood pressure or cardiovascular risk status at baseline (index date) will be excluded. Where there is no record of blood pressure lowering, statin or antiplatelet treatment, it will be assumed the patients were not exposed to blood pressure lowering, statin or antiplatelet treatment.
Patient eligibility includes only using 'acceptable patients' in the analysis, and therefore there is no need to impute age and sex. Where there is no record of smoking history or alcohol consumption, patients will be assumed to be non-smokers and non-drinkers. Likewise, those with no record of co-morbidities will be assumed to have no history of these conditions. All other covariates (including BMI, ethnicity and IMD) used in the prognostic or casual inference modelling will be imputed using multiple imputation. BMI and will be treated as a continuous variable.
All analyses of treatment associations with outcomes will be conducted by intention-to-treat, and patients moving practice after the index date (and therefore being lost to follow-up) will be censored at the point at which they are no longer active in the database.

R. Patient or user group involvement (if applicable)
This application has been discussed with patients (aged 67-86 years, taking 5-11 meds), a carer who also takes multiple meds takes 5 meds and the AgeUK Bakewell day centre (group discussion with 18 frail elderly persons).

Key issues raised
• Individuals vary widely on the amount of information they wish to know about their medications, and this is often influenced by the doctor-patient relationship. • Patients sometimes rely on unclear or unreliable information about the potential harms of their treatments from medication packets, the internet or friends. • Some patients are concerned about suffering events which will cause them to go to hospital, but most focus on the 'here-and-now'; chronic issues which affect their quality of life (e.g. pain limiting mobility). • Some do not trust their GPs understanding of medications and prefer to see a specialist pharmacist with more time to consider their needs. Others who see the same GP regularly are happy to 'do as they are told,' but assume the GP fully understands the benefits and harms of the treatments they take. • Patients are happy to be involved in research, but location and access is very important, and a multifaceted approach to engagement is required. Some are more comfortable with technology than others. • Older individuals said they were happier having (potentially sensitive) PPI discussions with someone they knew and trusted.

Project areas directly influenced by PPI
Overarching aim: Some individuals wish to know about the benefits and harms of medications whilst others expect the GP will fully understand the risks. A clinical decision support tool is needed to improve the understanding of both patients and their GPs.
Model outcomes: Patient's opinions vary in the outcomes that concern them the most and so this proposal focusses on both serious acute events and also chronic problems which can affect daily life. Advice will be sought as to how these outcomes should be defined in the final calculators.
PPI engagement: A variety of methods of PPI engagement are required to maximise access for patients. Engaging with relatives and familiar members of local churches and community groups will be important.

S.
Plans for disseminating and communicating study results, including the presence or absence of any restrictions on the extent and timing of publication All findings from the proposed research project will be published in peer-reviewed scientific journals. Findings will be presented at national and international conferences in Primary Care (e. An article discussing the issues raised by the research will be written for the online newspaper 'The Conversation', and this will also be posted on the Nuffield Department of Primary Care Health Sciences website as an online blog. Where appropriate, results of the research will be press-released in combination with their publication in scientific journals. Social media (twitter) will be used to draw further attention to the work.
The results of this work will be used to develop a calculator which estimates an individual's risk of suffering harm from preventative treatments for cardiovascular disease. The risk calculator will be made freely available online for use by patients and practitioners.
T. Limitations of the study design, data sources, and analytic methods This study focuses on potential harms from adverse events associated with preventative treatments for cardiovascular disease. Such outcomes may be less well recorded in routine electronic health records. For example, initial scoping work suggests the rate of falls documented in the CPRD is between 0.06-0.34% per 1000 person-months. These incidence rates may be less than those reported in smaller previous studies (0.15-6.40% per 1000 person-months) using different definitions, captured via detailed questionnaires and follow-up. [49][50][51][52] Incidence rates are likely to be higher when linked data from HES and ONS are available. 53 Missing data on certain outcomes may reduce the accuracy of the prognostic models, but should not bias the results provided the data are missing at random and not affected by whether patients are prescribed treatment. To minimise the risk of reporting bias, outcomes in the primary analysis will be defined according to hospital admissions documented in the HES or ONS.
The causal inference work will use an observational cohort design, and as such, there is an inherent selection bias of patients in both exposure and control groups. The impact of this bias will be limited by propensity score methods used to control for confounding, but this approach assumes that all confounding factors are measured and accounted for within the analysis. In the present study, we will examine the validity of this assumption by 1) using an instrumental variable (GP prescriber preference) and examining the association with the outcomes of interest and 2) by studying the impact of treatment on positive and negative controls: hospitalisation and/or death from cardiovascular disease and bowel cancer/COPD. We will conclude the assumption has been met if high GP prescribers have a similar association with outcomes and the treatment itself and treatment is shown to be associated with the positive control, but not the negative control.

Strategy for dealing with potential errors resulting from multiple testing
This study will examine multiple outcomes across 3 exposure variables. To avoid potential errors arising from multiple testing, a primary outcome has been clearly defined for each exposure variable. Secondary outcomes are also pre-specified. The primary outcome will be given priority in the final analysis write-up and any related reports and presentations.
Primary and secondary outcomes will be defined using data from Basic inpatient HES and ONS. Sensitivity analyses will explore analyses defining these outcomes using the above sources and read-coded primary care data. The accuracy of such outcome data has been examined previously in patients with acute myocardial infarction and whilst recording of risk factor and co-morbid information was consistent across primary care, hospital admissions and disease registry records, the crude incidence of acute myocardial infarction was underestimated by up to 50% if only one data source was used, compared with using all three sources. 53 The use of linked CPRD data will ensure outcome data are ascertained accurately.

U. References
For causal inference analyses (only), sensitivity analyses will explore outcomes defined more broadly using primary care codes as well.

O. Data/ Statistical Analysis
We no longer plan to conduct an internal validation of the prediction models. The following text from the protocol is therefore redundant: Where feasible and appropriate, each model will be internally validated using the bootstrap method, to allow optimism-adjusted estimates of calibration and discrimination performance to be obtained. Adjustment for overfitting will be undertaken using a uniform shrinkage factor identified from bootstrapping, followed by re-estimation of the cumulative baseline sub-distribution hazard function (and thus baseline cumulative incidence function) to maintain overall calibration between observed and predicted risks. Given the large sample size, overfitting is likely to be small (see sample size calculation above). If bootstrapping is computationally infeasible due to the large sample size, then a smaller random sample of the population will be used or alternative internal validation methods employed such as cross validation.

Aim 3 -causal inference analysis
For most of the outcomes described in this proposal, the association and magnitude of effect with treatment is not well defined in randomised controlled trials. We will therefore undertake causal inference work to examine whether modification of treatment could have an important impact on the risk of adverse events using two causal inference methods: multivariable adjustment, propensity score adjustment/matching and instrumental variable analysis.
For the analysis using multivariable adjustment, Cox proportional hazards models will be used to compare adverse event rates in each experimental group. Estimates will be adjusted for pre-specified variables thought to confound the relationship between treatment and adverse events.
For the analysis using propensity scores, we will conduct separate analyses adjusting for propensity score and matching via propensity score.

Subgroup and sensitivity analyses
Where possible and appropriate, analyses of treatment associations will be examined in subgroups of the population. These will include age, sex, baseline blood pressure and cardiovascular risk, by drug type/dose/intensity and past medical history. Subgroup definitions will be agreed by expert opinion or taken from an accompanying systematic review of previous trials, examining the same topic. For computational reasons (whereby only one imputation model is required for all subgroup analyses of each outcome), these analyses will only be undertaken using propensity score adjustment.
P. Plan for addressing confounding Two approaches will be taken to determine whether observed associations are potentially subject to residual confounding. 46 Where all approaches provide similar results, we can be more confident that the findings are accurate and reliable. Further steps will be taken to examine the validity of treatment effect estimates using positive and negative controls: the impact of treatment on an outcome known to be affected by treatment (e.g. cardiovascular disease and death; positive controls) Amendment 3 -25.11.2021 (approved 03/12/2021) We have added exploratory analyses using supervised machine learning approaches to address missing data and deal with confounding by indication in the causal inference analyses.

Aim 3 -causal inference analysis
For most of the outcomes described in this proposal, the association and magnitude of effect with treatment is not well defined in randomised controlled trials. We will therefore undertake causal inference work to examine whether modification of treatment could have an important impact on the risk of adverse events using two causal inference methods: propensity score matching and instrumental variable analysis.
For the propensity score analysis, eligible patients will be matched 1:1 at the index date using propensity scores, which indicate the likelihood a patient will be prescribed treatment on the basis of their known (pre-treatment) characteristics and other known information which might influence the decision to treat. Predictors of preventative drug prescription will be explored in a logistic regression model. Matched patients will be compared using the Cox proportional hazards models, but hazard ratios will only be adjusted by factors unbalanced at baseline which are not already included in the propensity score model. Time since cohort entry will be used as the underlying time function in each model.
For the instrumental variable analysis, GP's previous antihypertensive prescribing preferences will be used as an instrument to predict the likelihood of actual treatment. Risk of adverse events will be estimated in patients attending a GP considered a high prescriber of preventative medication and compared to those attending a GP considered a low prescriber. If the results from the propensity score matching suffer from unmeasured confounding, they will differ from those of the instrumental analysis. If this is the case, these models will be refined with additional factors to achieve better adjustment and matching. Where all approaches provide similar results, we can be more confident that the findings are accurate and reliable. 46 Generative Adversarial Networks (GAN) will also be used to impute missing counterfactual data (unobserved effects of untaken treatments) on an individual level. GAN models are a class of machine learning frameworks that involves the simultaneous training of a pair of deep neural networks in competition with each other. The idea is that the first network, the generator, generates synthetic data samples that are similar to the original samples, while the second network, the discriminator, evaluates the authenticity of the generated samples [1]. Data similarity and fidelity are usually evaluated using well-known metrics such as Maximum Mean Discrepancies [2] and Jensen-Shannon Divergence [3]. While GAN's initial uses were for imaging applications, its generative capabilities make it naturally suitable for generating missing values that can be used to impute the original samples.
In this work, we build on the intuition that treatment effects estimation and causal inference are a missing data problem, where counterfactuals (unobserved effects of untaken treatments) are missing [4]. Specifically, we plan to use GAN to impute missing counterfactual data, for the samples available in CPRD. By modifying the generator's goal to accurately impute missing data, and the discriminator's goal to distinguish between observed and imputed counterfactual values for the same patient record. The model can generate the missing data conditional on the observed values and assigned treatment for each patient as seen in [5,6]. While useful, the work of [5,6] are not compatible for time-series data nor for estimating multiple treatment effects, which limits its use in real-life treatment applications. Therefore, we plan to extend such previous works to make them compatible for multiple treatment options and for time-series data to allow for long-term treatment (antihypertensives, statins, antiplatelets) estimation effects for multi-morbid populations as initially proposed in this project.
The proposed GAN approach does not generate new patient samples, but rather generates counterfactual (missing component) conditioned on observed factual data, and outputs a complete record with both factual and counterfactual outcomes to allow for treatment effect estimation for each patient. This approach is naturally immune to re-identification attacks since no new samples are generated, but we will run privacy checks and measures within the data to verify that the model is not copying potential outcomes from patients in the training sample. Example privacy measures that we plan to apply include empirical evaluations such as mathematical privacy definitions such as identifiability [7] that quantifies the probability of re-identification to ensure that no combination of attributes could reveal the identity of a patient. Other privacy evaluation methods include testing that the model is robust against membership inference attacks, where an attacker attempts to determine if a specific patient was used in the training set of the model [8]. Lastly, we will experiment with strict theoretical guarantees such as differential privacy which allow for the model to learn almost nothing about an individual while learning useful information about a population [9].