Heart failure (HF) is one of the commonest reasons for hospitalisation in adults [1]. The burden of HF is rising globally because of ageing populations [2]. The Centers for Medicare and Medicaid Services’ (CMS) Hospital Readmission Reduction Program in the US currently penalises hospitals financially for higher than expected risk adjusted 30-day readmission rates for HF [3]. Emergency readmissions are a marker of poor patient health status, decline in quality of life as well as quality of care problems [3]. There is an ongoing discussion about the effectiveness of the 30-day readmission rate as a quality metric as many HF patients have multiple readmissions after 30 days [3]. For the general patient population, only a quarter of 30-day readmissions are potentially avoidable [4]. The concept of “high-impact users”, those with a high readmission rate, e.g. three unplanned admissions within a year, goes beyond the 30-day window [5]. Identification of these patients can help healthcare providers to formulate targeted policies in case-management programmes.

It is equally important to investigate the causes of emergency readmissions to assess if there is a repeated cycle of events that are common in these patients that can be potentially avoided in the community [6]. To avoid regression to the mean, which happens when patients have a crisis and high admission rates followed by a return to baseline rates, we have previously used trajectory modelling of rates over a five-year period [7].

International clinical trials have suggested variation in overall hospitalization and readmission due to exacerbation of HF in different regions of the world [8]. Hospitals in North America were shown to have the highest rate of hospitalizations [8]. Comparison of outcomes in HF patients among different countries showed disparity in prescription of medical therapy, patients undergoing diagnostic investigations, behavioural counselling and placement of implantable devices and other interventions [9]. Some countries have also documented inter-state variation in the outcomes of HF patients [10]; we previously compared admissions for HF in England and its regions with those in Lombardy in northern Italy and found that, despite key similarities in healthcare systems, English and especially London patients of a given age and gender spent on average much more time in hospital for HF but lived longer following their index admission than patients in Lombardy [11].

We investigate whether there is variation in the proportion and characteristics of high-impact users among HF patients by geographical region in England. Such variations could be partly due to differences in the medical management of these patients. Areas with higher proportions of high-impact users can then be identified and local policies implemented to improve outcomes.

Data were obtained from Clinical Practice Research Datalink (CPRD) linked to Hospital Episode Statistics (HES) and the Office for National Statistics (ONS) deaths database. CPRD is the largest national primary care database with over 13 million enrolled patient medical records across the country and contains 8.5% of the patient population in England [12]. It includes information on demographics, medical diagnosis and procedural information for each consultation with the general practitioner (GP). The fact and date of death are obtained from ONS (Office for National Statistics). Around 660 practices have volunteered in CRPD to share patient data. Around 60% of the general practices contributing to CPRD are linked to HES data.

The CPRD data include a field for region in England: North-East, Yorkshire and Humber, North-West, East of England, London, South-East, South-West, South-Central, and West-Midlands. The number of patients was very low in Yorkshire and Humber region; hence, they were grouped together with North-East region.

The assessed risk factors were identified from previous studies and categorised into patient-based, social and lifestyle related, and management-based factors. The patient-based factors consisted of age at diagnosis, sex, and past medical history recorded in the last 5 years preceding the diagnosis of HF. Age was grouped into categories for use in the model in the following brackets: 18–45, 45–54, 55–64, 65–74, 75–84, and 85+ − an adaption from previous studies [13]. The social and lifestyle factors included impact of bereavement, marital or relationship problems, history of smoking and heavy alcohol intake. These factors were chosen because they impact on the prognosis of heart failure [14]. The patients with ex-smoking and current smoking status in the preceding years were identified through medcodes; smoking was categorised as either former or current. Bereavement was defined as loss of an immediate family member.

The management-based factors consisted of GP visit coded for the monitoring of renal function, flu vaccination, measurement of blood pressure and exercise recommendation. Other factors included 3 or more emergency admissions for any reason other than HF in the year preceding the diagnosis of HF, the number of annual GP visits (including out-of-hours) and the number of out-of-hours GP visits in the year preceding the diagnosis of HF. Previous annual GP visits and annual out-of-hours GP visits did not have a linear relationship with the high-impact users. Hence, they were categorised according to percentile ranges (<25th, 26-50th, 51-75th and > 75th percentiles). The effect of HF diagnosed as an inpatient and history of use of medications for treatment of signs and symptoms of HF was also evaluated. Medication use for the management of signs and symptoms of heart failure included prescriptions for drugs like loop and thiazide diuretics, beta-blockers, angiotensin-converting-enzyme inhibitors, angiotensin II receptor antagonists, digoxin, and bumetanide [15]. The information on social factors and management-based factors was obtained for 5 years before the date of diagnosis of HF.

Group-based trajectory modelling (GBTM) was applied to the data to categorise patients into subgroups with different trends in readmission rates. The outcome was the annual number of emergency readmissions for each patient for each successive year during the follow-up period. The patients who died during their first hospital admission where they were diagnosed with HF were excluded from the analysis. They were part of the low-risk group and were used to compare the outcomes with other groups. In order to determine the optimum number of subgroups within a population, the choice of model was based on the following criteria: the smallest value of Bayesian Information Criteria (BIC), largest value for average posterior probability for each group, odds of correct classification (OCC) > 5 and each trajectory with significant parameter estimates (p <  0.05). These criteria are usually chosen to test the model with the best estimate of number of groups and predictors associated with them [16‐18]. The Statistical Analysis Software (SAS v9.4) was used to apply GBTM on the datasets via the macro ‘Proc Traj’ [19]. For each subgroup within the patient population, the average number of readmissions annually was measured and depicted on the graph. The group showing highest annual readmission rate persistently in the follow up period was termed as high risk group and those with consistently low average readmission rate were classed as low risk group. The group with mediocre annual readmission rate in the follow up period was categorised as intermediate group. Multinomial logistic regression model was used to assess co-variates associated with high-impact users and low-impact group was used a reference for comparison.

Sequence analysis was performed to identify common causes (primary diagnoses) and their pattern of emergency admissions among subgroups of HF population using the ‘TraMineR’ package in R [20]. The primary diagnosis codes were grouped together based on their common pathology.

Regional analysis of heart failure patients showed that all 8 English regions had more than two subgroups based on annual readmission rates during 5 years’ follow-up. The high-impact users usually had two subgroups: chronic high-impact with constantly high readmission rates and short-term high-impact with initial high readmission rates followed by a rapid immediate decline. A significant number of patients were intermediate users, some of whom had the potential of becoming high-impact users. Risk factors such as chronic respiratory disease, chronic renal disease, stroke, anaemia, mood disorder, and cardiac arrhythmia had a common association with high-impact users in most of the regions. A high proportion of patients among the high-impact group had multiple readmissions, with similar repeated common causes consisting of respiratory tract infection, urine infection, exacerbation of heart failure, ischaemic heart disease and external injuries.

The patients who died during the follow up period were included in the study to provide a pragmatic picture of the observational data following heart failure diagnosis. Short-term high impact users had very high readmission rate in the first year following heart failure and then rapid decline in the readmissions. These patients were also shown to have severe cardio-pulmonary past medical history. Hence, the decline in their readmission rate could have been the result of increased mortality among them after first year following the diagnosis of heart failure.

The proportion of high-impact users varied from 4 to 15% by region: the proportion of high-impact users and their annual readmission rate was lower in southern regions. There was large variation in the characteristics of the patients and the use of healthcare resources among different regions. GP visits for all types of HF medication review and prescriptions were lower in Northern, West Midlands and East of England regions. These regions had a higher number of patients with hospital admissions and out-of-hours GP visits preceding the diagnosis of HF. The recent national audit on hospitals in the UK showed marked variation in the discharge outcomes and in the proportion of patients with HF medication prescription, follow-up imaging, discharge planning and referral to a specialist nurse, cardiologist and cardiac rehabilitation [21].

In every region, intermediate users were a sizable group with the potential to become high-impact users. In each region, they were relatively young population and had similar risk factors. They were more likely than other groups to have hypertension, which was suggestive of the fact that they were prone to get further cardiovascular complications [22, 23]. They also had smaller odds of being associated with regular GP visits for HF medications [24, 25]. It may be the case that these patients complied poorly with their medication, resulting in vascular morbidities in the future [22, 24]. We need to explore other factors that prevent them from becoming high-impact users. It is possible that they may have less severe heart failure or the cause of heart failure is different from high-impact users. In any case, a combination of clinical data with administrative data can help answer these questions.

This study is an initial step towards the demonstration of regional variation in the readmission pathways using epidemiological data. Most of the previous data on regional variation is obtained from multi-centre clinical trials that have assessed the clinical effect of variation in the use of medical therapy for the treatment of heart failure [26, 27]. For most studies, the primary outcome has been the overall mortality and readmission for exacerbation of heart failure condition, whereas some studies have evaluated the overall readmission rate as a secondary measure [8, 27].

Despite variation in co-morbidities of the patients in different regions, the common causes of hospital admissions and their sequences were similar among high-impact patients with multiple hospital readmissions. These patients underwent a vicious cycle of admissions for cardiopulmonary signs and symptoms, chest and urine infections, ischaemic heart disease and external injuries, including falls. Exacerbation of HF is one of the common causes of readmissions. However, these patients get admitted to hospital for other reasons as well. Multiple causes of hospital admissions among HF patients show that the medical management of HF alone cannot yield better outcomes for them [1, 28]. Reducing the readmission rate will require improvement of holistic care of these patients, such as the promotion of secondary preventative measures for ischaemic heart disease, regular flu vaccination, exercise recommendation and prevention of falls and fractures [1, 28‐30].

It is important to recognise chronic high-impact users early and assess the option of different care pathways to allow easy transition of care. The predictive model could be used to identify potential high-impact users among heart failure patients by the clinical team. These patients may benefit from closer surveillance. For example, previous work suggests that more aggressive pre- or post-operative cardiopulmonary work-up or rehabilitation to avoid iatrogenic complications may be of benefit following AAA repair. An early and aggressive mobility and cardiopulmonary rehabilitation program for patients in ITU and colonic surgery patients has already been shown to reduce readmission rates [31, 32].

Understanding the structure and provisions of primary care setting in each region may play an important role in understanding regional variation in the trends of readmission rates. The variability in availability and involvement of specialist HF community nurses also need to be assessed in each region. GPs may also differ in their referral rates to specialist community nurses. If they are too late, then the condition may reach at uncompromised severity and it may become hard to prevent exacerbation of the condition. With the current training system, not all GPs spend time on specialist cardiology wards and may not receive hands-on training in the management of heart failure. As a consultant, they may be hesitant in treating the condition and would have low threshold for referral of these patients, as an elective or emergency setting, to specialist care [33]. Furthermore, GP practices have increased trends in recruiting GP with specialist interest. These GPs get additional training in the specialised area, such as cardiology, gynaecology, etc. The proportion of GPs for a population sample in each region and, among those, the proportion of GPs with specialist interest in cardiology is still to be evaluated in each region [34].

This study had certain limitations which need to be considered. Firstly, the analysis did not take into account use of other healthcare resources like outpatient visits or emergency department visits. The categorisation of groups was based on emergency hospital readmission rates alone to understand long-term morbidity and real-life events that impact quality of life. We also focused on emergency hospital admissions because they are a major contributor to health cost among high-impact users. The comparison of the study results with previous studies was limited because very few studies have conducted trajectory and sequence analysis to study long term hospital care use [35]. Secondly, despite coding errors, the Read and ICD coding of the conditions used to select patient cohort have high specificity, and primary diagnoses in HES data are accurate [36, 37]. We tried to use all possible codes that define the condition to include most cases. Thirdly, the use of primary care data linked to hospital data in CPRD does not include all patients suffering from the condition in focus. Six hundred seventy four practices in the UK are registered with CPRD and provide information for primary care data [12, 38]. Approximately half of these practices meet the quality criteria for data input [15]. Not all GP practices are linked with secondary data. However, it is the best available linked data in the country and indeed the largest such database in the world. It is much used for research and provides a great opportunity to assess various hospital- and primary care-based factors and to evaluate long-term outcomes [39]. Fourthly, the regional codes in the CPRD data can only demarcate the country in the broader regions as shown in the study. These regions consist of heterogeneous populations of patients which can present with intra-regional variation in the use of healthcare resources. Further studies are required to investigate differences in healthcare use at a smaller regional distribution and assessing other potentially contributory factors, such as, ethnicity, social support, socio-economic deprivation. This study provides an initial assessment and overview of the differences in the healthcare among regions.

We found regional variations in five-year rates of readmission and death and in the proportion of high-impact users in HF patients. Potential reasons include quality of care provided at primary and secondary care level. High-impact users comprise a minority but require special support in the community, and further studies are required to assess the factors associated with GP management of the condition.