Post-mortem examination of high mortality in patients with heart failure and atrial fibrillation

Patients were included if they presented with HF and AF to the cardiology department or transferred from another department at SCJUO or other local hospitals. Heart failure was categorized using common clinical classification into HFrEF (LVEF < 40%), mildly-reduced/intermediate (LVEF 40–49%) and HFpEF (LVEF ≥ 50%; see Additional file 1: Table S1 for definitions) [11]. LVEF was based on echocardiography performed by accredited cardiologists; the most recently available data were used, either performed during the admission or within the previous 6 months. The diagnoses of AF from the clinical team were retrieved and categorized using accepted definitions (Additional file 1: Table S2) [12]. To avoid missing data, patients without a complete follow-up available were excluded (for example, four patients who lived outside the region) and also patients whose death was due to a violent cause and required investigation by forensic specialists.

Clinical factors were determined at the time of admission from the patient’s electronic medical record. Information on participants was collected within a database system separate from any clinical databases. Data collected included demographic information, associated clinical evaluations and comorbidities and treatments received during the hospital admission and at discharge. Laboratory data and drug information were used to confirm specific comorbidities, in addition to complementary investigations, such as cardiac and non-cardiac imaging. The diagnosis of ischemic heart disease was based on the patient’s history of significant coronary heart disease by coronary angiography or based on chest pain associated with an increased level of changes in cardiac markers (troponin I or highly sensitive troponin I) and/or echocardiographic changes consistent with ischemia or a positive non-invasive stress test. Hypertension was defined based on clinical history, use of antihypertensive therapy, a systolic blood pressure ≥ 140 mmHg or a diastolic blood pressure ≥ 90 mmHg. Diabetes was defined based on serum glucose levels greater than 126 mg/dL, glycosylated haemoglobin values greater than 6.5% or the use of oral antidiabetics or insulin. Ischaemic or haemorrhagic stroke was certified by a computed tomography performed during hospitalization in patients with neurological deficits. The CHA₂DS₂-VASc thromboembolic risk score was calculated with points for heart failure, hypertension, age ≥ 75 years [double], diabetes mellitus, previous thromboembolism [double], vascular disease, age 65–74 years and female gender. The HAS-BLED bleeding risk score was calculated with points for uncontrolled hypertension, creatinine > 2.2 mg/dL or dialysis, cirrhosis or elevated liver tests, prior stroke, prior bleeding, labile control of warfarin, age > 65 years, medication predisposing to bleeding and excess alcohol intake. N-terminal pro-B-type natriuretic peptide (NTproBNP) levels were analysed using an accredited and calibrated Pathfast assay (Pathfast devices, Mitsubishi Chemical Europe), with a detection range of 15–30,000 pg/mL.

Deaths were extracted from the hospital database electronic medical record, as well as from the Unique Integrated Computer System of the Social Health Insurance in Romania (SIUI). The primary outcome was mortality in-hospital or during the follow-up period, categorized into heart failure-related death, vascular death and non-cardiovascular death.

Confirmation of events required documentary evidence supporting the diagnosis (for example, a death certificate, pathologist notes and autopsy report). Post-mortem autopsies were performed where death occurred within 24 h of admission, when the cause of death was not immediately clear or when requested by the attending physician, unless the next of kin refused consent. Post-mortems were performed according to a standardized protocol (Additional file 1: Online Methods). At the time that autopsies were performed, pathologists had full access to the patient charts, the clinical medical history and the results of any additional investigations, in order to fully correlate post-mortem findings (see example in Additional file 1: Fig. S1).

Values are presented as median ± interquartile range (IQR; 25th to 75th centiles) or percentage. Group comparisons were assessed with the Kruskal–Wallis non-parametric analysis of variance test. Mortality was analysed with Cox regression models, presented as hazard ratios (HR) with associated 95% confidence intervals (CI). Multivariate models were prespecified to include age, gender, New York Heart Association (NYHA) class, LVEF, type of AF (paroxysmal versus non-paroxysmal), obesity (body mass index ≥ 30 kg/m²), coronary artery disease, hypertension, chronic obstructive pulmonary disease and diabetes. For multivariate analysis, patients with new-onset AF were included in the paroxysmal category, regardless of therapy, with sensitivity analyses excluding these patients finding no material difference in results. As natriuretic peptides are not specific to HF and are also prognostic in patients with coronary disease [13], NTproBNP was kept out of multivariate models so as not to obscure other associations. All models were also adjusted for medical therapy at baseline (statins, renin–angiotensin–aldosterone antagonists, beta-blockers, diuretics, amiodarone and digoxin therapy). Interactions were assessed with likelihood ratio testing, and the proportional hazards assumption in Cox models was confirmed using Schoenfeld residuals. In those that died, separate Cox models were generated to compare the differences in associations between heart failure-related death, vascular death and non-cardiovascular death. Kaplan–Meier plots were used to present the pooled, unadjusted data and number at risk, with the log-rank test of equality used to compare the groups. LVEF was modelled using a restricted cubic spline analysis in the Cox model.

A two-tailed p-value of < 0.05 was considered statistically significant. Where multiple comparisons were performed, an adjusted p-value was used. Analyses used complete case data as the amount of missing data was small (no imputation performed). Statistical analysis was performed with Stata (version 14.2, StataCorp LP, TX).

Over a mean follow-up of 1.5 ± 0.9 years, 291 patients (28.9%) died, of which 186 had a post-mortem performed (63.9% of patients who died; see Additional file 1: Table S3 for the characteristics of those undergoing autopsy). Baseline characteristics and comorbidities, comparing those alive and dead at follow-up, are presented in Table 1. There was no difference in either the CHA₂DS₂-VASc thromboembolism score or the HAS-BLED bleeding risk score. The median NTproBNP value on admission for those that died was 8710 pg/mL (IQR 4871–20,430); Additional file 1: Fig. S2. In-patient ventricular arrhythmias were documented in 57 patients (5.7%); 21 subsequently died (36.8%) which was not significantly different to those without ventricular arrhythmias (28.4%; p = 0.18).

Factors independently associated with all-cause mortality in the multivariate model were limited to the presence of coronary artery disease (HR 2.34, 95% CI 1.77–3.08; p < 0.001) and hypertension (HR 1.45, 95% CI 1.11–1.88; p = 0.006); Additional file 1: Table S4. Category of HF (according to LVEF) and type of AF (according to temporal pattern) were not significantly associated with all-cause mortality; Fig. 1. LVEF was not related to all-cause mortality overall (HR 1.00 per 1% increase, 95% CI 0.98–1.01; p = 0.44). Figure 2 graphically depicts this lack of relationship between LVEF and all-cause mortality across the range of LVEF observed in study participants.

A full list of causes of death is presented in Table 2, including the differences between patients that underwent post-mortem examination. The 291 deaths were classified as HF-related in 136 patients (46.7%), vascular death in 75 patients (25.8%) and non-cardiovascular death in 80 patients (27.5%). The mode of death according to the HF category and AF type is presented in Fig. 3. Baseline variables associated with the different modes of death were unique and contrary, particularly with regard to LVEF (Table 3).

Comparing HF-related death with the other causes of death, multivariate analysis identified higher NYHA class (HR 2.45 per one class increase, 95% CI 1.73–3.46; p < 0.001) and lower LVEF (HR 0.95 per 1% increase, 0.93–0.97; p < 0.001) as independently associated with HF-related death. The median NTproBNP level in this group of 11,869 pg/mL (IQR 5498–25,410) was significantly higher than for other causes of death, but with a broad range (p < 0.0001; Additional file 1: Fig. S2).

Factors associated with vascular death compared to other causes were hypertension (HR 2.83, 1.36–5.90; p = 0.005) and more preserved LVEF (HR 1.08 per 1% increase, 1.05–1.11; p < 0.001). The median NTproBNP level in this group was 7098 pg/mL (IQR 5054–9841).

Non-cardiovascular death compared to other causes identified obesity (HR 2.20, 1.21–4.00; p = 0.010) and more preserved LVEF (HR 1.10 per 1% increase, 1.06–1.13; p < 0.001) as significant predictors. The median NTproBNP level in this group of 6618 pg/mL (IQR 2029–9767) was not significantly different compared to patients with vascular death (p = 0.08).

Each mode of death had recognizable comorbidities and characteristics associated with them (Additional file 1: Table S5). However, the overall multi-morbidity burden (as estimated by clinical risk scores) was similar across those alive and the various modes of death. The mean baseline CHA₂DS₂-VASC risk score for stroke and thromboembolism was 5.2 ± 1.3 for patients subsequently alive at the end of follow-up, 5.0 ± 1.2 for HF-related death, 5.4 ± 1.4 for vascular death, and 5.1 ± 1.3 for non-CV death (p = 0.15). Similarly, HAS-BLED was similar across the groups: 3.1 ± 1.4, 2.9 ± 1.6, 3.4 ± 1.6 and 2.8 ± 1.4.

The patients included in this study were predominantly of European descent, with highly symptomatic HF and AF at presentation. Patients came from a wide geographic area including around half from rural communities. Due to the study’s aim of assessing post-mortem evidence where possible, a retrospective design was required. However, this design, and the inherent observational nature of the study, can lead to the potential selection and ascertainment biases. To reduce bias, we enrolled consecutive patients and ensured that data completion rates were high with no need for imputation and no patients had missing information on their vital status. We did not assess follow-up therapy, so it is possible that some patients withdrew or were withdrawn from their anticoagulant therapy after discharge. Details on cardiac resynchronization therapy were not available, and only 49 (4.9%) had an implanted cardiac defibrillator as device implantation required referral to other hospitals.

A key strength of our analysis is the availability of post-mortem examinations for a large proportion of cases. The need for an autopsy was principally determined by the timing of death (within 24 h of admission). However, attending physicians were also able to request a post-mortem if the cause of death was not clear and the family did not object. We used the recorded cause of death to avoid any issues of competing diagnoses; however, in clinical practice, it is likely that other underlying conditions would also have played a role in the patient’s demise. We do not report on sudden cardiac deaths as these are rarely documented as causes of death within Romania. Pathologists instead record any underlying diseases contributing to mortality, which is a limitation to the generalizability of our study to non-hospitalized populations. Autopsies may have misrepresented deaths due to ventricular arrhythmia in cases with other incidental findings (e.g. occlusive coronary disease), although the clinical staff had full access to hospital records when ascertaining the cause of death. There is a possibility of a miscoded cause of death in those not undergoing post-mortem examination; however, we saw only small differences in the underlying aetiology, with fewer vascular and more HF-related deaths in those with autopsy information. Finally, any observations are limited by the number of death events; however, this was large in comparison with other published data with 3 in 10 patients dying during the study period.