Long-term outcomes and impact of diagnostic coding of patients with heart failure identified in primary care: insights from the OSCAR-HF cohort

Abstract

Background and aims

Misclassification of heart failure (HF) in primary care electronic health records (EHRs) may affect care quality and outcomes. The OSCAR-HF (Optimising Standards of Care of Heart Failure in general practice) study implemented audit and feedback to improve HF identification in Belgian general practices. We assessed long-term outcomes in the OSCAR-HF cohort and examined differences by preaudit HF registration status.

Methods and results

We followed 538 patients identified with HF after an EHR audit in eight primary care practices (2017, median follow-up 7.4 years). Overall survival at 1, 2 and 5 years was 93.2%, 84.3% and 62.4%. During follow-up, 1083 hospital admissions occurred; 11.3% were HF-related and 64.3% were non-cardiovascular. We compared patients with a preaudit registered HF diagnosis (n=177) versus those without (n=361). Preaudit registered patients had more HF admissions during follow-up (23% of patients vs 13%, p=0.002), yet lower all-cause mortality in multivariable analysis (adjusted HR 0.62; 95% CI 0.40 to 0.95). Nearly half of the HF with reduced ejection fraction (HFrEF) population had no preaudit HF registration (49%, n=54/110). Unregistered patients more often had HF with preserved ejection fraction (HFpEF) (61% of patients vs 49%, p<0.001).

Conclusion

In this primary care HF cohort, HF admissions were infrequent and most hospitalisations were non-cardiovascular. Patients with a registered preaudit HF diagnosis had more HF admissions but lower mortality. The high proportion of unregistered HFrEF and the increased range of therapeutic options for the HFpEF spectrum underscore the need for comprehensive case-finding and accurate coding across the HF spectrum in primary care.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Heart failure (HF) is frequently misclassified or undercoded in primary care electronic health records (EHRs), which can affect both care delivery and the interpretation of real-world outcomes.

WHAT THIS STUDY ADDS

• This is the first study to link HF registration status in primary care EHRs to long-term outcomes over more than 7 years.

• Almost half of all patients (54/110) with HF with reduced ejection fraction (HFrEF) did not have a registered HF diagnosis before the diagnostic audit, and unregistered patients more often displayed an HF with preserved ejection fraction (HFpEF) profile.

• HF admissions were a minority of all admissions (11%), while non-cardiovascular causes dominated hospitalisations (64%) and death (82%).

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Strategies focusing solely on patients with prior HF admissions or specialist diagnoses may risk missing large, treatment-eligible groups across the HF spectrum.

• Accurate case-finding and coding for both HFrEF and HFpEF should remain a priority for primary care.

• The low number of HF-related admissions and deaths challenges the use of HF hospitalisation as a primary effectiveness outcome in community-based interventions, as it may underestimate the true burden of disease and the impact of care improvements in this setting.

Introduction

Heart failure (HF) is a chronic condition with a large impact on patients’ health and quality of life. Since age is a major risk factor, it is a public health problem of increasing importance in ageing populations, caused by the substantial costs associated with hospital admissions for decompensated HF. Generally, patients have three or more other chronic conditions and are 75 years or older.^{1 2} This geriatric and multimorbid patient profile implies an important role for primary care health providers. General practitioners (GPs) are an essential part of the multidisciplinary HF team,^3–5 and their involvement in coordinating care across different healthcare settings for this complex patient population is crucial for holistic care.

However, diagnosing HF in primary care is difficult. Signs and symptoms (such as fatigue or shortness of breath) are often non-specific in this population. In addition to these disease-specific barriers for the diagnosis of HF, GPs report patient-related factors (patients’ reluctance of referral), physician-related factors (lack of confidence in diagnostic testing) and contextual factors (such as lack of availability of diagnostic testing).⁶ Unfortunately, this results in frequent misdiagnosis in primary care.^7–9

To address this issue and improve primary care HF management, we set up the OSCAR-HF (Optimising Standards of Care of Heart Failure in general practice) study, implementing a complex intervention consisting of audit and feedback, natriuretic peptide (NT-proBNP) testing at the point-of-care and a specialist HF nurse, nested within eight different primary care practices in Belgium.^{10 11} The cohort of patients identified as having HF by their GP after a diagnostic audit in the electronic health record (EHR) constituted the OSCAR-HF cohort (n=538). However, an important clinical question is whether patients who did not have a registered HF diagnosis before such a diagnostic audit have clinically different features and long-term outcomes compared with their counterparts with a registered HF diagnosis. Additionally, OSCAR-HF was unique in the sense that GPs had to conduct a comprehensive assessment of the accuracy of EHR registration, which is why we want to compare long-term outcomes with those seen in other primary care HF cohorts.

The aim of the present study is therefore to assess long-term outcomes of the OSCAR-HF cohort and to explore whether in this cohort, the subpopulation with unregistered HF in their EHR had different baseline characteristics and long-term outcomes than their counterparts with a registered diagnosis.

Methods

Study population

This is a follow-up analysis of the OSCAR-HF pilot study; a non-randomised, non-controlled prospective observational trial conducted for 6 months in 2017 in eight Belgian general practices (51 GPs).^10–12 Eligible patients were aged ≥40 years and were registered with participating practices. The intervention consisted of reimbursed natriuretic peptide point-of-care testing for participating GPs during the study period, a specialist HF nurse intervention where GPs could refer patients with HF for an educational intervention and an audit and feedback assessment in the EHR. With regard to the first two interventions, we refer to the published protocol and baseline results.^{10 11}

To understand the status of registered HF preaudit, we briefly summarise the audit procedure.¹² At the audit-date (T0), all participating GPs completed a standardised EHR audit and feedback workflow. The audit consisted of (1) a basic audit querying coded and free-text HF diagnoses and (2) an extended audit querying coded and free text terms for known HF risk factors such as signs, symptoms, comorbidities and combinations of typical HF medications. The audit resulted in a list of patients for case-review by the GP, who classified each patient as HF or non-HF. Patients classified as having HF constituted the OSCAR-HF study population. In addition, an expert panel assessed the validity of each 10th HF diagnosis, ruling diagnoses as either objectified or non-objectified. Based on these discussions, a diagnostic flowchart was constructed to facilitate a standardised approach for each case (online supplemental figure S1). The panel review of every 10th HF diagnosis served to calibrate and standardise adjudication rather than to classify a sample: the resulting flowchart was subsequently applied to every case, so that objectification status was determined for all 538 patients. Type of HF was assigned from echocardiographic data available in the EHR, using the thresholds HF with reduced ejection fraction (HFrEF) <40%, HF with mildly reduced ejection fraction (HFmrEF) 40–50% and HF with preserved ejection fraction (HFpEF) ≥50%. Ejection fraction was reported quantitatively in 58% of available reports and qualitatively in the remainder, and patients with no echocardiographic data on file were classified as non-classifiable. Classification was based on any documented echocardiogram rather than on the most recent value, so patients whose ejection fraction had subsequently improved retained their original classification. This procedure is described in full in the baseline report of the OSCAR-HF study.¹¹ The full audit classification sequence, from the 18 011 patients eligible for the audit to the 538 classified as having HF, is described in detail elsewhere.¹³

The study start or baseline (T1) occurred approximately 1 month after the audit. During this period, the principal investigator performed detailed data extraction, and coded and harmonised HF diagnoses. Consequently, at baseline (T1) all patient files contained a corrected HF diagnosis, and the 6-month OSCAR-HF intervention formally began.

Throughout this manuscript, we distinguish two exposure groups based on EHR status before the audit (T0): patients with a recorded HF diagnosis (preaudit registered HF) and patients without a recorded HF diagnosis (preaudit unregistered HF). All outcomes are measured after the audit, from this common corrected starting point. Importantly, during the 6-month study period, all patients had access to three components of care: reimbursed NT-proBNP point-of-care testing, an audit-and-feedback process in the EHR and referral to a specialist HF nurse. Uptake of these components could vary: differences in use between the two preaudit registration groups are reported in the Results.

Baseline variables

In OSCAR-HF, we extracted a number of relevant characteristics at baseline (T1, after the audit). To assess possible differences according to preaudit registration status, we compared sex, age, age group, objectified HF at baseline, type of HF, institutionalisation in a long-term care facility, smoking behaviour, body mass index, renal clearance, comorbidities, use of typical medication for HF, history of hospital admissions before and during the study, quality of life and the intervention of a HF nurse during the 6-month study period.

Study outcomes

All outcomes were measured from T0 forward. We investigated mortality rates at 1-year, 2-year and 5-year after inclusion in the OSCAR-HF cohort (study start). We also explored the number of hospital admissions with at least an overnight stay. We defined the cause of admission as the primary diagnosis in the discharge letter. Since we did not have access to death certificates, we identified the cause of death from the discharge letter or physician notes. Date of death was extracted through the central Belgian mortality registry. We classified causes of admission and death into 17 categories (online supplemental table S1) fitting within three larger groups of HF admissions, cardiovascular and all-cause (ie, non-cardiovascular) admissions. For a HF admission we used the European Society of Cardiology consensus statement¹⁴: an overnight stay in-hospital caused by substantive worsening of HF symptoms. For cardiovascular admissions: we specified admissions for myocardial infarction, unstable angina, stroke and transient ischaemic attack, percutaneous coronary intervention, peripheral vascular intervention or stent thrombosis.¹⁵ All other cardiovascular admissions were coded as ‘other cardiovascular diseases’.

Data collection

We previously described the data collection methods for the OSCAR-HF study.^{10 11} Briefly, all variables were collected manually from each individual’s electronic patient file in general practice. For the data collection on long-term outcomes, we created an electronic clinical research form (online supplemental table S2) in Research Electronic Data Capture (REDCap). REDCap is a secure, web-based software platform designed to support data capture for research studies.¹⁶ We collected data on all hospitalisations starting from the initial date of the OSCAR-HF practice audit (T0). Participants contributed person-time from the audit date (T0) until the earliest of:

1. Death (ascertained via the national mortality registry, which is linked to Belgian GPs’ EHR).

2. Transfer out of the practice or relocation outside the catchment area.

3. ≥12 months without any contact in the electronic patient file (we assumed that this meant patients moved to another practice and recorded the date of last contact as censoring date).

4. Administrative end of follow-up (database lock).

We did not formally categorise reasons for censoring beyond these operational rules; therefore, most censoring reflects administrative end of follow-up, practice transfer/relocation or long intervals without recorded contact.

Data were collected and entered in this platform by two researchers during a follow-up visit in the participating practices.

Data analysis

We present categorical data as counts (%). We summarise non-normally distributed continuous data with medians and IQRs, or mean (SD) for normally distributed continuous data. For group-comparisons by preaudit registered versus unregistered HF, we employed several statistical tests. The χ² test was used to assess the association between categorical variables. For continuous variables that did not follow a normal distribution, the Wilcoxon rank-sum test was applied to compare the medians between groups. Fisher’s exact test was used for categorical variables with small sample sizes to ensure accurate p value calculations. To address the issue of multiple testing and reduce the risk of type I errors, we calculated q-values using the false discovery rate method. We calculated crude mortality rates using the OSCAR-HF date as starting point. Because death precludes further admissions, we modelled mortality as a terminal event and analysed recurrent HF admissions jointly with death using a joint accelerated frailty model (reReg package).^{17 18} In addition, we used the reda package¹⁹ to compute mean cumulative function (MCF) estimates using the Nelson-Aalen estimator and fitting a gamma frailty model with a spline rate function. In these summaries, follow-up ends at the terminal event (death); therefore, the MCF estimates the cumulative admission burden per patient up to death or censoring. We obtained SEs for the joint model by non-parametric bootstrap (5000 resamples, fixed seed). 95% CIs are Wald intervals on the log-scale and p values the corresponding two-sided Wald tests. We did not impute missing data. For descriptive comparisons, missing values are reported as a separate ‘Unknown’ category in table 1. Covariates for the joint model were binary indicators derived from the EHR. Where a finding was not recorded, the indicator was coded as absent. All analyses were conducted in R V.4.4.2.

Table 1. Overview of baseline characteristics and differences between the OSCAR-HF subgroups according to registered HF in the electronic health record before audit and feedback.

			Preaudit registered HF
Baseline characteristics	N	Overall n=538*	No n=361*	Yes n=177*	P value†	q-value‡
Sex	538				0.3	0.6
Female		270 (50)	176 (49)	94 (53)
Male		268 (50)	185 (51)	83 (47)
Age (years)	538	81 (72, 87)	81 (73, 87)	81 (72, 88)	0.7	0.9
Age group	538				>0.9	>0.9
40–50		16 (3.0)	10 (2.8)	6 (3.4)
50–60		28 (5.2)	18 (5.0)	10 (5.6)
60–70		60 (11)	41 (11)	19 (11)
70–80		143 (27)	96 (27)	47 (27)
>80		291 (54)	196 (54)	95 (54)
Objectified HF at baseline	538	359 (67)	208 (58)	151 (85)	<0.001	<0.001
Objectified HF after 6 months follow-up	538				<0.001	<0.001
No		128 (24)	112 (31)	16 (9.0)
Unclear		54 (10)	42 (12)	12 (6.8)
Yes		356 (66)	207 (57)	149 (84)
Type of HF	538				<0.001	<0.001
HFmrEF		75 (14)	44 (12)	31 (18)
HFpEF		306 (57)	220 (61)	86 (49)
HFrEF		110 (20)	54 (15)	56 (32)
Unclassified		47 (8.7)	43 (12)	4 (2.3)
Long-term care facility	538	57 (11)	37 (10)	20 (11)	0.7	0.8
Active smoking	538	39 (7.2)	26 (7.2)	13 (7.3)	>0.9	>0.9
BMI	348	27.4 (24.2, 31.2)	27.1 (24.2, 30.8)	28.3 (24.4, 31.4)	0.3	0.5
Unknown		190	138	52
Obesity	538	112 (21)	67 (19)	45 (25)	0.065	0.2
Last baseline eGFR	531	60 (44, 76)	62 (46, 78)	57 (42, 73)	0.054	0.2
Unknown		7	5	2
Number of comorbidities	538	5.00 (4.00, 7.00)	5.00 (4.00, 7.00)	6.00 (4.00, 7.00)	0.007	0.030
History of valvular disease	489	196 (40)	124 (39)	72 (42)	0.5	0.6
Unknown		49	42	7
History of atrial fibrillation	538	266 (49)	165 (46)	101 (57)	0.013	0.049
History of ischaemic heart disease	538	239 (44)	163 (45)	76 (43)	0.6	0.8
Hypertension	538	345 (64)	236 (65)	109 (62)	0.4	0.6
Diabetes mellitus	538	164 (30)	111 (31)	53 (30)	0.8	>0.9
COPD	538	76 (14)	51 (14)	25 (14)	>0.9	>0.9
Dementia	538	60 (11)	44 (12)	16 (9.0)	0.3	0.5
Depression	538	113 (21)	81 (22)	32 (18)	0.2	0.5
Hypothyroidism	538	48 (8.9)	38 (11)	10 (5.6)	0.062	0.2
Hyperthyroidism	538	51 (9.5)	34 (9.4)	17 (9.6)	>0.9	>0.9
Pacemaker	538	63 (12)	36 (10.0)	27 (15)	0.073	0.2
Cancer	538	117 (22)	86 (24)	31 (18)	0.10	0.2
Target dose of RAAS at baseline	538	141 (26)	89 (25)	52 (29)	0.2	0.5
Target dose of betablocker at baseline	538	63 (12)	38 (11)	25 (14)	0.2	0.5
ARNI	7	2 (29)	0 (0)	2 (50)	0.4	0.6
Unknown		531	358	173
Loop diuretics	534	207 (39)	111 (31)	96 (55)	<0.001	<0.001
Unknown		4	3	1
Mineralocorticoid receptor antagonists	534	150 (28)	79 (22)	71 (40)	<0.001	<0.001
Unknown		4	3	1
Quality of life (MLWHFQ)	298	19 (8, 32)	20 (9, 32)	18 (6, 31)	0.2	0.5
Unknown		240	161	79
HF nurse intervention during the 6-month study period	538	139 (26)	77 (21)	62 (35)	<0.001	0.003
HF admission in 3 years before study start	538	139 (26)	77 (21)	62 (35)	<0.001	0.003
HF admission during long-term follow-up	538	87 (16)	46 (13)	41 (23)	0.002	0.009
Cardiovascular admission in 3 years before study start	538	162 (30)	118 (33)	44 (25)	0.063	0.2
Cardiovascular admission during long-term follow-up	538	77 (14)	54 (15)	23 (13)	0.5	0.7
All-cause admission in 3 years before study start	538	345 (64)	229 (63)	116 (66)	0.6	0.8
All-cause admission during long-term follow-up	538	348 (65)	230 (64)	118 (67)	0.5	0.7

Bold q-values indicate statistical significance (q<0.05) after false discovery rate correction for multiple testing.

^*n (%); Median (Q1, Q3).

^†Pearson’s χ² test; Wilcoxon rank-sum test; Fisher’s exact test.

^‡False discovery rate correction for multiple testing.

ARNI, angiotensin receptor–neprilysin inhibitor; BMI, body mass index; COPD, chronic obstructive pulmonary disease; eGFR, estimated glomerular filtration rate; HF, heart failure; HFmrEF, heart failure with mildly reduced ejection fraction; HFpEF, heart failure with preserved ejection fraction; HFrEF, heart failure with reduced ejection fraction; MLWHFQ, Minnesota Living with Heart Failure Questionnaire; OSCAR-HF, Optimising Standards of Care of Heart Failure in general practice; RAAS, renin–angiotensin–aldosterone system.

The study findings are reported in accordance with the Reporting of Studies Conducted Using Observational Routinely Collected Health Data recommendations.²⁰ The completed Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) checklist is provided in online supplemental table S5 and as online supplemental file 2.

Results

Median follow-up was 7.4 years (IQR 7.1–7.6 years), with a maximum follow-up of 7.7 years. At T1, the mean age was 79 years (SD 11). In this period, there were 1083 admissions in total among 378 patients (70.3%) (figure 1). HF admissions accounted for 11.3% of all admissions (n=122). Cardiovascular (non-HF) diseases accounted for 24.5% of admissions (n=265). 64.3% of all admissions (n=696) were for non-cardiovascular reasons. We provide both the number of patients with ≥1 admission and the event counts per cause category in online supplemental table S3, as well as the mean (SD) admissions per patient in each category.

Figure 1. Overview of causes of admissions during long-term follow-up for the OSCAR-HF cohort. HF admissions are coloured blue. Cardiovascular admissions are coloured orange. All other admissions are coloured green. HF, heart failure; MI, myocardial infarction; OSCAR-HF, Optimising Standards of Care of Heart Failure in general practice; PTCA, percutaneous transluminal coronary angioplasty.

1-year survival (with respect to all-cause mortality) for the OSCAR-HF cohort was 93.2% (95% CI 91.1% to 95.4%), 2-year survival 84.3% (95% CI 81.2% to 87.5%), 5-year survival 62.4% (95% CI 58.2% to 66.9%) and 7-year survival 52.9% (95% CI 48.6% to 57.7%) (figure 2).

Figure 2. Kaplan-Meier survival curves (all-cause mortality) comparing patients with a preaudit registered HF diagnosis in their electronic health record (blue) and those without (red). Shaded areas represent 95% CIs. Numbers at risk are shown. HF, heart failure.

At the end of follow-up, there were 236 recorded deaths. Cardiovascular causes accounted for 18.2% of deaths (n=43), while the majority were due to non-cardiovascular causes, including infections, neoplasms and chronic respiratory diseases (online supplemental figure S2).

Patients with a preaudit registered diagnosis were more likely to have objectified HF (85% vs 58%), a history of HF admission (35% vs 21%), atrial fibrillation (57% vs 47%) and to receive mineralocorticoid receptor antagonists (MRAs) (40% vs 22%) or loop diuretics (55% vs 31%) (table 1). During the 6-month study period they also had more contact with the specialist HF nurse (35% vs 21%). In contrast, patients without a preaudit registered diagnosis were more likely to have HFpEF (61% vs 49%). Patients with preaudit registered HF were also significantly more likely to have an admission for HF during follow-up (23% vs 13%, p=0.002). The MCF estimate for HF admissions in patients with preaudit registered HF was significantly higher than their counterparts without a registered diagnosis (1.88 times as high, 95% CI (1.09 to 3.26), p=0.024) (figure 3). There was no significant difference in cardiovascular (non-HF) or all-cause (non-cardiovascular) admissions. We report specialist care contact and HF medication by registration status in online supplemental table S6. Patients with a preaudit registered diagnosis more often had a specialist HF diagnosis (68% vs 29%) and a cardiology consultation in the previous 5 years (93% vs 82%), whereas prescription of renin-angiotensin system inhibitors (67% vs 61%, p=0.18) and beta-blockers (73% vs 72%, p=0.84) did not differ. Looking only at the patient population with HFrEF (n=110), there was no evidence of a difference between registered and unregistered patients in cardiology consultation before baseline (100% vs 96%), renin-angiotensin system inhibitors (80% vs 89%), beta-blockers (86% vs 87%) or MRAs (54% vs 47%) at baseline, and target doses were reached in only a minority of both groups.

Figure 3. Unadjusted MCF plots for cumulative admissions per patient over time: (A) HF admissions, (B) CV (non-HF) admissions and (C) AC (non-cardiovascular) admissions. The blue line represents patients with a registered HF preaudit; the red line represents patients without registration. Shaded areas indicate 95% CIs. AC, all-cause; CV, cardiovascular; HF, heart failure; MCF, mean cumulative function.

In a multivariable analysis, registered HF was associated with a modestly higher rate of HF admissions (HR 1.022, 95% CI 1.001 to 1.043, p=0.042), although the magnitude of this effect was small (figure 4). As for mortality, somewhat surprisingly, patients who had preaudit registered HF had a significantly lower mortality risk (HR 0.62, 95% CI 0.40 to 0.95), which was also true for patients taking loop diuretics (HR 0.54, 95% CI 0.33 to 0.87). Full univariable and multivariable model estimates for recurrent HF admissions and mortality are provided in online supplemental table S4. A graphical summary of the study design and principal findings is provided in online supplemental file 3.

Figure 4. Forest plots of multivariable associations with (A) HF admissions and (B) all-cause mortality in the OSCAR-HF cohort. Points represent HRs on a log scale; horizontal bars indicate 95% CIs. Orange markers denote statistically significant associations (p<0.05). ‘ns’ indicates non-significant associations. Variables are ordered by effect size (distance from 1) within each panel. HF, heart failure; LVEF, left ventricular ejection fraction; MRA, mineralocorticoid receptor antagonist; NP, natriuretic peptide; OSCAR-HF, Optimising Standards of Care of Heart Failure in general practice.

Discussion

In this community HF cohort more than half of all patients (52.9%) survived after 7 years of follow-up. The majority of hospitalisations were for non-cardiovascular causes (64.3%), with only a minority due to HF (11.3%). We observed significant differences in characteristics and long-term outcomes across preaudit registration groups. Patients without registered HF before the audit were more likely to have HFpEF, and less likely to be prescribed loop diuretics or MRAs. Patients with a registered diagnosis were significantly more likely to be admitted for HF during follow-up (23% vs 13%), yet exhibited a significantly lower risk of all-cause mortality in multivariable analysis (HR 0.62; 95% CI 0.40 to 0.95). We hypothesise that this pattern reflects a difference in disease recognition before the audit, where patients with more advanced or specialist-acknowledged HF were more likely to be registered and thus remained at higher risk of HF admissions, whereas those identified through the audit more often had HFpEF, leading to fewer HF admissions but a heavier non-cardiovascular burden and worse survival. Importantly, almost half of the HFrEF patient population did not have preaudit registered HF, indicating that accurate identification and coding across the HF spectrum (HFrEF and HFpEF) should remain a primary care priority.

Comparing our findings to long-term outcomes in other primary care HF cohorts two observations stand out. First, survival rates at one and 5 years (93.2% and 62.4%) were higher than those reported in primary care cohorts in the UK (80.8% and 48.2% last follow-up 2017), Sweden (79.7% at 1 year, 2017)^{21 22} and Belgium (74% at 1 year, 2006).²³ A possible reason could be survivor bias since these latter cohorts used the incident HF diagnosis date as index (rather than an audit index date such as in our study). About a third of our population had a registered HF diagnosis previous to the study start, implicating they already survived the high-risk early phase of HF and this could have led to inflated survival estimates compared with incident cohorts. Other potential explanations are the OSCAR-HF intervention itself which improved diagnostic accuracy and GPs’ awareness of HF, and could have resulted in a higher baseline of care quality, as well as improvements in therapeutic management of HF in the last decade and finally the fact that GPs could have wrongly identified patients as having HF, as acknowledged in the limitations. Second, the low proportion of HF admissions. This is in line with findings from a large cohort in the UK, where admissions for HF accounted for 13% of hospitalisations within a year of diagnosis and decreased over time, and cardiovascular admissions represented fewer than half of all admissions,²⁴ although our cohort consisted largely of stable patients managed in the community. Indeed, in our population non-cardiovascular causes dominated admissions, with infectious disease being the single largest category, as well as deaths, where only 18.2% were cardiovascular. Set against a mean age of 79 years (SD11) and a median of five chronic conditions (IQR 4–7), this is best understood as a reflection of the burden of multimorbidity in an older community HF population in which infection, frailty and non-cardiac comorbidity drive hospitalisation at least as much as the index cardiac diagnosis. This low number challenges the use of HF hospitalisations as a primary effectiveness outcome in community-based interventions, since it addresses only a minority of the acute care burden these patients actually experience, and may substantially underestimate the true burden of disease and the impact of care improvements in this setting.

To our knowledge, this is the first study to assess the impact of unregistered HF in the EHR through a full case revision by GPs and link this status to long-term mortality or admission outcomes. As stated previously, the unregistered HF population was more likely to have HFpEF, and less likely to have objectified HF, or take loop diuretics (31% vs 55%) or MRAs (22% vs 40%). As such they more closely resembled the typical community HF phenotype with HFpEF and a high burden of comorbidity.^{25 26} It is therefore not surprising that over time, patients with a registered HF diagnosis accumulated 1.88 times as many HF admissions per patient as those without a registered diagnosis (MCF ratio 1.88; 95% CI 1.09 to 3.26). In the joint frailty model, registered HF (HR 1.022, 95% CI 1.001 to 1.043, p=0.042), reduced ejection fraction (HR 1.017, 95% CI 1.000 to 1.035, p=0.047) and a recorded history of clinically relevant valvular disease (HR 1.037, 95% CI 1.009 to 1.066, p=0.010) showed statistically significant but small independent associations with recurrent HF admissions. This is interesting since our findings also raise the question as to how patients with unregistered HFrEF were managed. Here it is important to note that registration status describes documentation in the GP’s record rather than the true underlying care. As we reported, specialist contact and guideline-directed treatment within HFrEF were comparable regardless of registration status. Absence of a registered diagnosis in the HFrEF population therefore likely did not imply absence of specialist care or appropriate therapy, which offers a plausible explanation for why it was not accompanied by an excess of HF admissions, although target doses were reached in just a minority of patients with HFrEF in both groups (renin-angiotensin system inhibitors 33% vs 40%, beta-blockers 23% vs 9%), so neither group could be considered optimally treated. The only previous study that addressed this question was conducted in a US primary care network two decades ago, and found that patients whose problem list recorded HF were more likely to receive an ACE inhibitor or angiotensin receptor blocker (ARB), with no difference for beta-blockers. The authors attributed this difference to initiation by cardiologists without access to the primary care problem list.²⁷ A large more recent cohort study suggested that the specificity of the recorded HF diagnosis rather than the presence of a code predicted care.²⁸ Whether improving the record improves outcomes is unclear, since a pay-for-performance incentive for GPs to deliver structured HF management was not associated with better outcomes.²⁹ However, it should be noted that the HFrEF subgroup in our study was small, and that it is in HFpEF, where the GP is often the primary manager of care and where therapeutic options have significantly increased in years since our study concluded, where early recognition and coding may matter most.

Our results thus show we are missing patients at both ends of the HF spectrum: almost half of the HFrEF population and more than two thirds of the HFpEF population lacked a registered diagnosis preaudit. Accurate identification and coding should remain a primary care priority in order to timely initiate guideline-recommended care, although our findings indicate that coding by itself should not be expected to change treatment where specialist care is already established.

Strengths and limitations

This study has several strengths. First, it was conducted in a real-world primary care setting. Second, the OSCAR-HF cohort was a primary care cohort comprehensively identified by GPs, which allowed us to address substantial misclassification¹³ and offers a unique perspective, distinct from other registry-based primary care studies that identified patients based on diagnostic codes. Third, we report long-term outcomes over more than 5 years, including mortality and recurrent hospitalisations. These data are rarely available for primary care HF cohorts. In addition, we conducted a detailed manual data extraction from patients’ individual EHRs, which was linked to the national mortality registry, and made use of advanced statistical methods such as joint frailty models and MCF analysis to account for recurrent events and competing risks.

However, there are some limitations. First, this is an observational non-randomised study design, which limits any causal inference regarding the effects of registration and interventions on outcomes. Second, as stated previously, inclusion of previously selected patients might have introduced survivor bias and possible inflated survival estimates. Third, data collection was conducted by GPs in training rather than specialist HF consultants, which may have introduced variability in the interpretation of clinical details, although we specifically designed a structured case form and admission classification algorithm and conducted rigorous testing and training before data collection. Fourth, this study was conducted in a Belgian primary care setting with a strong GP infrastructure, which might limit generalisability to other health systems. Fifth, a key limitation was that the OSCAR-HF cohort was identified by GPs through a structured EHR audit and case review rather than by universal, external adjudication. This approach likely introduced some degree of misclassification (both missed and wrongly assigned HF cases). We mitigated this by employing an extended audit (coded and free-text queries), an expert panel that validated every 10th HF diagnosis and developed a diagnostic flowchart, and by harmonising and recoding diagnoses before baseline (T1). Nevertheless, not all patients had fully objectified HF at baseline (67% overall, higher in the preaudit registered subgroup), which could bias estimates and may partly contribute to comparatively favourable survival relative to incident cohorts. This GP-driven approach was a deliberate design choice to maximise external validity and to evaluate what a scalable diagnostic audit can achieve in real-world primary care, where GPs will perform patient selection if implemented at scale. Sixth, HF phenotyping was limited to the variables captured in the prespecified EHR extraction (ejection fraction category, diastolic dysfunction, clinically relevant valvular disease, left ventricular hypertrophy and left atrial enlargement). More recent relevant measures such as quantified valvular regurgitation severity and right ventricular function were not collected and could not be reported, so within-category heterogeneity of HF phenotypes were not captured. Seventh, patients during the study period had access to reimbursed point-of-care natriuretic peptide testing and to a specialist HF nurse, neither of which reflects standard primary care. Uptake of the nurse intervention was moreover higher in patients with a preaudit registered diagnosis (35% vs 21%) so the enhanced care environment could have been a potential confounder of the comparisons reported here as well as a constraint on generalisability. Finally, echocardiographic findings were recorded less often for patients without a preaudit registered diagnosis (12.2% vs 4.0% missing), consistent with their lower rate of objectified HF (58% vs 85%). Because unrecorded echocardiographic findings were coded as absent, these covariates potentially reflect what was documented rather than what was present, and their associations should be interpreted accordingly.

Research implications

Our study has several implications for policy and future research. First, our results suggest that the systematic identification of HF in primary care remains important, although registration in the GP’s record might have been an indicator for patients’ recognition as patients with HF by their GP, rather than their treatment: guideline-recommended therapy was similar regardless of registration status. We also found that previous HF admissions and a specialist HF diagnosis are strong unadjusted predictors for future HF admissions (HR 6.63, 95% CI 2.57 to 17.12 and HR 3.24, 95% CI 1.32 to 7.97, respectively), although neither remained independently associated after adjustment. From a public health perspective, it therefore remains important to invest in diagnostic infrastructure such as audit and feedback integrated in the EHR or natriuretic peptides. Second, future studies should explore how electronic health data can be leveraged for risk stratification and population health management in the community, including the development of decision-support tools to improve diagnostic accuracy and treatment optimisation. These scalable interventions (audit and feedback, natriuretic peptides and decision support) can consequently be investigated in an implementation science framework in larger and more diverse primary care settings.

Conclusion

In this long-term follow-up of a primary care HF cohort, we found that patients with a previous registered diagnosis in the EHR had more HF admissions but significantly lower mortality compared with those without registration. Almost half of all patients with HFrEF did not have a registered HF diagnosis, although guideline-directed therapy was similar. The majority of admissions in this cohort were for non-cardiovascular causes. These findings highlight the importance of accurate coding to correctly identify HF populations in primary care, as well as the need for comprehensive strategies that address the full spectrum of HF, including HFpEF and multimorbid patients, to improve outcomes in the community.

Data availability statement

Data are available upon reasonable request.