While each individual data source captured stroke events that were uniquely recorded in that source, the overall concordance of stroke events between sources was very high (median 0 days between records) indicating both a good level of data linkage across sources and quality within sources. We additionally demonstrate, with a comprehensive ascertainment of stroke in all available sources in England, that during the study period (2020-2023) most ischaemic stroke patients were dispensed an anticoagulant or antiplatelet agent. However, there was under-use of guideline-recommended antihypertensives (40%), statins (80%), and anticoagulants in those with AF (77%), particularly in older people or those with more comorbidities [9]. There was no major change in the proportion of people treated during the COVID-19 period. Home-time in the first 6 months, an indicator of post-stroke disability, was lower in older people, those with a more severe stroke, or more comorbidities, with modest differences by region, sex and ethnicity. After the COVID-19 period, people spent more time at home.

Secondary prevention is key to reducing stroke incidence. Whilst population-wide prescriptions of antihypertensive and lipid-lowering drugs fell during the COVID-19 pandemic[8], in stroke patients we found no major difference in dispensed medicines by year or by patient characteristics. This indicates a longer-term and systemic problem not due to the pandemic [10]. The evidence is clear that targeting a systolic blood pressure of ≤125mmHg [11] in all patients with stroke (whether ischaemic or haemorrhagic) reduces recurrent stroke, and that most stroke patients have a blood pressure over the threshold for hypertension diagnosis at presentation[12]. It is most likely that antihypertensives have substantially lower rates of dispensing - particularly in older people who stand to gain the most - and that many more patients could benefit. While these findings suggest a potential gap in secondary prevention, the use of dispensing data alone cannot account for clinical contraindications, patient preferences, or end-of-life care goals that may appropriately preclude certain medications from being prescribed. Consequently, the observed lower rates of dispensing may reflect, in part, appropriate clinical decision-making that is not captured in routine EHR records.

Linked data for measuring quality of care are essential for a learning health system[13]. Here, we developed a large, linked dataset with ascertainment of quality beyond the hospital system to primary care and pharmacy. However, this work is unlikely to be replicable routinely beyond the pandemic unless the recommendations of the Sudlow review[14] - to allow national linkage of primary care data - are implemented. No timeline is available for providing access to such data, meaning that patients with stroke who do not become inpatients (for example discharged from emergency departments, or looked after in an ambulatory TIA/stroke clinic, or diagnosed in primary care) are missing from national measurement of quality. This is particularly important for pandemic preparedness as our work shows that using a single source for ascertaining stroke, e.g. only hospitalisation data, would underestimate the number of people with stroke in England.

Our analysis has a number of strengths. First, it ascertains stroke patients in the whole English population who are registered in primary care (almost the entire population[15]). Second, we ascertained stroke from multiple overlapping sources, including the national quality register of hospitalised stroke, hospital administrative data and primary care. The date of stroke was largely consistent between sources. Third, we used previously developed and validated code lists to identify strokes. Fourth, we were able to ascertain all relevant secondary prevention medicines dispensed to patients, where a submission had been made for reimbursement from NHS England, most prescriptions in the population and a more accurate proxy of compliance than prescribing data.

Our analysis has several weaknesses. First, we were unable to check the coded diagnosis of stroke against written records. The diagnosis of stroke in health systems data in the UK[16], compared with manual review of records, has good positive predictive value (PPV), when in the primary position in hospital records (94%), is reasonable in primary care (80%), although is less good in death records (57%), and on average is less good for haemorrhagic (52%) than ischaemic stroke (83%). This is important, because risk factors and case fatality differ by subtype. Accurate clinical coding is important for all health systems, and methods should be improved with greater use of clinicians or automated coding systems to aid clinical coders [17]. Second, we did not include recurrent stroke, because differentiating further consultations with the first stroke is difficult to separate from recurrent stroke in records. Third, we examined NHS dispensed rather than prescribed medicines. This will miss medicines that were prescribed but not dispensed, which in some cases is a substantial proportion[18]. A further limitation is that our data lacks granular information on clinical intent, medication intolerance, and contraindications. We cannot distinguish between a failure to prescribe and a deliberate clinical decision based on patient frailty or preferences. It will also miss prescriptions from the private sector which are not collected through NHS reimbursement schemes. Fourth, in this manuscript, we did not examine in detail differences by ethnicity or geography and so may have missed communities who were particularly disadvantaged. Fifth, whilst home-time does correlate with disability and symptoms in this analysis and others, it is influenced by other factors such as family needs and does not directly measure impairment [19–21]. Sixth, our estimates of stroke incidence were high, compared with population-based incidence studies with bespoke measurement of stroke. This may be because there is overuse of stroke codes in primary care data, or that differentiating first-ever from recurrent stroke is difficult. This emphasises the need for improved clinical coding throughout healthcare pathways. Seventh, we used data provided by SSNAP under the Notice under Regulation 3(4) of the Health Service (Control of Patient Information) Regulations 200, that under provisioned records because of truncated monthly uploads to the SDE. Planning mechanisms for complete flow of data would prepare for the next health emergency.

Our analysis has several implications for health policy makers and stroke clinicians. Secondary prevention should be a priority for healthcare quality improvement for patients with stroke, and investigation of the clinical, system, and patient related factors leading to underuse should be research priorities. Despite the disruption caused by the COVID-19 pandemic on secondary prevention, we have shown here that there is longer term underuse that must be addressed. Home-time is an indicator of disability which could be added to the SSNAP and the stroke audit in Scotland for all patients, if it is interpreted carefully. Lastly, the use of linked health data - with an awareness of its limitations - is a promising avenue for audits of care for people with stroke and other conditions, to maximise ascertainment, streamline efficiency of data collection and enhance utility.