Abstract
The Proximity Extension Assay proteomics chip provides a large-scale analysis of 92 biomarkers linked to cardiovascular disease or inflammation. We aimed to identify the biomarkers that best predicted long-term all-cause mortality in patients with acute myocardial infarction.
In this prospective cohort study, 92 biomarkers were analysed in 847 consecutive patients from the Västmanland Myocardial Infarction Study with a median follow-up of 6.9 years.
The mean (± standard deviation) age of the patients was 70 (11.8) years and 32.7% were female. Two hundred and seven patients had died after follow-up. The biomarkers most strongly linked to all-cause mortality were growth differentiation factor 15 (GDF-15) and tumour necrosis factor-related apoptosis-inducing ligand receptor 2 (TRAIL-R2). Cox regression analysis showed that GDF-15 (hazard ratio 1.25 per unit change, 95% confidence interval, 1.02–1.53, p = 0.031) and TRAIL-R2 (hazard ratio 1.37 per unit change, 95% confidence interval 1.12–1.67, p = 0.002) were independent predictors of long-term all-cause mortality after adjusting for age, gender, diabetes, previous myocardial infarction, stroke, heart failure, hypertension, smoking, hypercholesterolaemia, body mass index, ST-elevation myocardial infarction, left ventricular ejection fraction, troponin I, estimated glomerular filtration rate, N-terminal pro-brain natriuretic peptide and C-reactive protein. The combination of GDF-15 and TRAIL-R2 with established risk factors and biomarkers showed a discriminating accuracy of separating survivors from non-survivors with a cross-validated area under the receiving operating characteristics curve of 0.88 within five years.
GDF-15 and TRAIL-R2 were the most powerful Proximity Extension Assay chip biomarkers in predicting long-term all-cause mortality in patients with acute myocardial infarction.
Introduction
Acute myocardial infarction (AMI) is a major cause of morbidity and mortality worldwide. Advances in early reperfusion strategies limiting the extent of AMI have significantly reduced mortality, whereas late complications have become more relevant.1 Several proteins are involved in both the immediate biological response following an acute ischaemic event and the process of injury repair and remodelling. Therefore, it is no surprise that several studies have shown an association between numerous biomarkers and prognosis.2–5 Biomarkers with prognostic power might be a potential tool for risk stratification as well as a future guide for the appropriate use of resources and therapies following an AMI.6 As the multiplex Proximity Extension Assay (PEA) proteomics chip provides a large-scale analysis of 92 biomarkers linked to cardiovascular disease (CVD) and inflammation, our objective here was to identify the most prevailing biomarkers in the PEA chip that added most to the prediction of long-term all-cause mortality in patients with AMI.7
Methods
Study population
In the Västmanland Myocardial Infarction Study (VaMIS; ClinicalTrials.gov identifier: NCT 01452178) consecutive patients ≥18 years of age admitted to the Coronary Care Unit, Hospital of Västmanland Västerås, Sweden from November 2005 to May 2011 for suspected AMI were candidates for inclusion.8 AMI was diagnosed according to the guidelines at the time of inclusion: a typical rise and fall in cardiac troponin I level of ≥0.4 µg/l in combination with ischaemic symptoms, new pathological Q waves, ST-segment elevation or depression seen by electrocardiography, or the need for coronary intervention.9 Among 1459 patients diagnosed with an AMI, 201 were excluded because the blood samples for biomarker analyses were obtained >72 h after admission. Furthermore, a total of 250 patients were excluded because of dementia or confusion (n = 81), other severe disease (n = 62), linguistic difficulties (n = 57) or declining participation (n = 50). In addition, all patients where protocol data were missing (n = 161) were excluded. Finally, 847 patients remained for analysis in the present study. Written informed consent was obtained from all participants. The Ethics Committee of Uppsala, Sweden approved the study (Protocol number 2005:169).
Data collection
The baseline data collected at the index examination included demographics and medical history. Self-reported medication at inclusion was confirmed from medical records. Medical history was defined by patient-reported diagnoses, information from medical records and the use of anti-diabetes, anti-hypertensive medications or statin treatments. Blood pressure was measured manually with the subject in a seated, resting position. Smokers were identified from self-reported information and defined as previous or current smokers. The left ventricular ejection fraction (LVEF) was assessed using the biplane Simpson’s formula.10 In subjects for whom it was not possible to obtain the Simpson’s LVEF (n = 175, 20.7%), a visual estimation of LVEF was made. Reproducibility data on echocardiographic measurements have been presented elsewhere.11 The estimated glomerular filtration rate (eGFR) was calculated according to the CKD-EPI formula.12
Blood sampling procedure
Blood samples were taken in 5 ml lithium heparin-coated vacuum tubes. The tubes were centrifuged at 2000 g for 10 min (Becton Dickinson and Co., Franklin Lakes, New Jersey, USA) or 2200 g for 10 min (Vacuette, Greiner Bio-One GmbH, Austria). Plasma was then reallocated to 5 ml plastic tubes and frozen at −70℃ within 2 h. The plasma samples were stored at −70℃. Before analysis, the samples were thawed at room temperature, mixed and centrifuged at 3470 g at 4℃ for 15 min and aliquoted into a microtitre plate using a pipetting robot, the Tecan Freedom Evolyze.
Proteomics
Measurement of protein biomarkers in plasma was performed using the PEA chip, Proseek Multiplex CVD I96×96 (Olink Bioscience, Uppsala, Sweden) at the Clinical Biomarkers Facility, Science for Life Laboratory, Uppsala University, Uppsala, Sweden. This assay requires less than 10 µl of sample to measure a panel of 92 biomarkers associated with CVD or inflammation and four internal control samples. The proteins were measured simultaneously by the binding of paired oligonucleotide-labelled antibodies to the target proteins and the subsequent formation of new polymerase chain reaction (PCR) amplification targets detected and quantified by high-throughput real-time PCR.13 The measures are specified as Normalized Protein Expression (NPX). The NPX values are generated from quantitative PCR quantification cycles (Cq) where a higher value corresponds to a lower protein profusion. The Cq values were corrected for technical variation by an inter-plate control. The lower limits of detection (LOD) were assessed as NPX = Olink negative control – (ϪCqsample – Ϫinter-plate control). Values below the LOD were imputed as being LOD normalized for the plate. These values were rescaled to a distribution with a mean of 0 and a standard deviation of 1 for each plate. Of the 92 biomarkers, 11 proteins gave a call rate of < 80% and so were excluded from further analysis (heat shock 27 kDa protein, pappalysin-1, pentraxin-related protein PTX3, beta-nerve growth factor, magnetosome protein, P-selectin glycoprotein ligand 1, melusin, SIR2-like protein 2, interleukin-4, caspase 8 and natriuretic peptide B).
Validation of the assay including 90 proteins and seven samples analysed in nine separate runs showed a mean intra-assay coefficient of variation of 8% (range 4–13%) and an inter-assay coefficient of variation of 15% (range 11–39%). Further information is presented on the Olink webpage (http://www.olink.com).
End-point and follow-up
The primary end-point was all-cause mortality. All-cause mortality was chosen as the end-point primarily for two reasons. First, the well-known uncertainty concerning the cause of death, especially in the elderly, and second, even though the PEA chip encloses biomarkers associated with CVD, several of the most considered biomarkers in the chip are not specific for CVD. The patients were followed until 31 December 2015. The median (1st to 3rd quartile) follow-up was 6.9 years (5.2–8.5 years). Two patients were lost during follow-up.
Statistics
Differences between baseline characteristics are presented as mean and standard deviation, median and range, or percentage and frequency (Table 1). Variables were compared using Mann–Whitney nonparametric U and Chi-squared tests. All biomarkers were analysed separately in Cox regression models adjusted for age and sex. p-values were adjusted for multiplicity using the Hochberg method. Age was included in the models using a restricted cubic spline (RCS) with three knots. Selection of the biomarkers in the PEA that yielded the best prediction of all-cause mortality within five years was performed using L1 penalized logistic regression analyses with series of different penalties and evaluated using 10-fold cross validation to estimate the area under the receiver operating characteristics curve (ROC AUC) adjusted for overfitting. To evaluate whether the selected biomarkers in the PEA chip added prognostic information above and beyond established risk factors and biomarkers, four different logistic regression models and corresponding cross-validation ROC AUC were studied. Model I included age and gender. Model II included established risk factors and biomarkers: age, gender, diabetes, previous AMI, stroke, heart failure, hypertension, hypercholesterolaemia, smoking, body mass index, ST-elevation myocardial infarction (STEMI), LVEF, eGFR, C-reactive protein (CRP), troponin I and N-terminal pro-brain natriuretic peptide (NT-proBNP). Model III included age, gender and the selected PEA-biomarkers. Model IV incorporated established risk factors and biomarkers used in model II and the selected PEA-biomarkers. In all models, age was included using RCS as previously described. Models IV and II were further compared using Net Reclassification Improvement (NRI) over tertiles of risk predicted using established risk factors. The selected PEA-biomarkers and risk factors were also evaluated separately and combined in Cox regression models to calculate crude and adjusted hazard ratios. In this analysis, PEA-biomarkers, age, NT-proBNP, CRP, eGFR and troponin I were initially included in the model using a three-knot RCS. Before inclusion, NT-proBNP, CRP, eGFR and troponin I were log-transformed to avoid numerical problems with the RCS and are listed in Table 2 with the prefix ‘log’. All variables were tested for possible non-linear effects using the likelihood ratio test. Variables that did not reach statistical significance entered the model as linear terms. Univariable and multivariable linear regression analyses were used to determine the association between the PEA-selected biomarkers, established risk factors and biomarkers. For statistical analysis, we used IBM SPSS Statistics (v. 24.0, IBM Corp., Armonk, NY, USA) and R (R Foundation for Statistical Computing; 2016, Vienna, Austria; https://www.r-project.org). Two-sided p-values < 0.05 were considered to be statistically significant.
Characteristics of the 847 patients with acute myocardial infarction.
| Variables | All patients | Survivors | Non-survivorsa | p-value |
|---|---|---|---|---|
| Patients, n | 847 | 640 | 207 | |
| Age, years, mean (±SD) | 70 (11.8) | 67 (10.8) | 78 (10.4) | <0.001 |
| Female, n (%) | 277 (32.7) | 197 (30.7) | 80 (38.6) | 0.036 |
| STEMI, n (%) | 256 (30.2) | 212 (33.1) | 44 (21.3) | 0.001 |
| Body mass index > 30 kg/m2, n (%) | 178 (21.0) | 143 (22.3) | 35 (16.9) | <0.001 |
| Systolic blood pressure, mmHg, mean (± SD) | 127 (21.8) | 127 (20.6) | 129 (24.6) | 0.388 |
| LVEF < 45%, n (%) | 196 (23.1) | 112 (17.5) | 84 (40.6) | <0.001 |
| LVEF < 35%, n (%) | 80 (9.4) | 36 (5.6) | 44 (21.3) | <0.001 |
| Current or previous smokers, n (%) | 550 (64.9) | 423 (66.1) | 127 (61.4) | 0.460 |
| Hypertension, n (%) | 466 (55.0) | 336 (52.5) | 130 (62.8) | 0.010 |
| Hypercholesterolaemia, n (%) | 265 (31.3) | 195 (30.5) | 70 (33.8) | 0.367 |
| Diabetes mellitus, n (%) | 152 (17.9) | 92 (14.4) | 60 (29.0) | <0.001 |
| Angina pectoris, n (%) | 221 (26.1) | 121 (18.9) | 100 (48.3) | <0.001 |
| AMI, n (%) | 187 (22.1) | 103 (16.1) | 84 (40.6) | <0.001 |
| Stroke, n (%) | 64 (7.6) | 32 (5.0) | 32 (15.5) | <0.001 |
| Heart failure, n (%) | 75 (8.9) | 30 (4.7) | 45 (21.7) | <0.001 |
| HbA1c (Mono-S) ≥ 5.0%, n (%) | 282 (33.2) | 187 (29.2) | 95 (45.8) | <0.001 |
| eGFR, ml/min per 1.73 m2, mean (± SD) | 70.9 (21.8) | 76.2 (19.0) | 54.5 (21.8) | <0.001 |
| Troponin I, mean (± SD) | 15.7 (20.6) | 16.2 (21.8) | 14.0 (16.5) | 0.442 |
| Medication at baseline, n (%) | ||||
| Beta blockers | 331 (39.1) | 205 (32.0) | 126 (60.9) | <0.001 |
| ACEi/ARB | 289 (34.1) | 182 (28.4) | 107 (51.7) | <0.001 |
| ASA/Plavix | 344 (40.6) | 204 (31.9) | 140 (67.6) | <0.001 |
| Diuretics | 165 (19.5) | 80 (12.5) | 85 (41.1) | <0.001 |
| Statins | 275 (32.5) | 176 (27.5) | 99 (47.8) | <0.001 |
| PCI, n (%) | 421 (49.7) | 360 (56.2) | 61 (29.4) | <0.001 |
| CABG, n (%) | 50 (5.9) | 27 (4.2) | 23 (11.1) | <0.001 |
| Variables | All patients | Survivors | Non-survivorsa | p-value |
|---|---|---|---|---|
| Patients, n | 847 | 640 | 207 | |
| Age, years, mean (±SD) | 70 (11.8) | 67 (10.8) | 78 (10.4) | <0.001 |
| Female, n (%) | 277 (32.7) | 197 (30.7) | 80 (38.6) | 0.036 |
| STEMI, n (%) | 256 (30.2) | 212 (33.1) | 44 (21.3) | 0.001 |
| Body mass index > 30 kg/m2, n (%) | 178 (21.0) | 143 (22.3) | 35 (16.9) | <0.001 |
| Systolic blood pressure, mmHg, mean (± SD) | 127 (21.8) | 127 (20.6) | 129 (24.6) | 0.388 |
| LVEF < 45%, n (%) | 196 (23.1) | 112 (17.5) | 84 (40.6) | <0.001 |
| LVEF < 35%, n (%) | 80 (9.4) | 36 (5.6) | 44 (21.3) | <0.001 |
| Current or previous smokers, n (%) | 550 (64.9) | 423 (66.1) | 127 (61.4) | 0.460 |
| Hypertension, n (%) | 466 (55.0) | 336 (52.5) | 130 (62.8) | 0.010 |
| Hypercholesterolaemia, n (%) | 265 (31.3) | 195 (30.5) | 70 (33.8) | 0.367 |
| Diabetes mellitus, n (%) | 152 (17.9) | 92 (14.4) | 60 (29.0) | <0.001 |
| Angina pectoris, n (%) | 221 (26.1) | 121 (18.9) | 100 (48.3) | <0.001 |
| AMI, n (%) | 187 (22.1) | 103 (16.1) | 84 (40.6) | <0.001 |
| Stroke, n (%) | 64 (7.6) | 32 (5.0) | 32 (15.5) | <0.001 |
| Heart failure, n (%) | 75 (8.9) | 30 (4.7) | 45 (21.7) | <0.001 |
| HbA1c (Mono-S) ≥ 5.0%, n (%) | 282 (33.2) | 187 (29.2) | 95 (45.8) | <0.001 |
| eGFR, ml/min per 1.73 m2, mean (± SD) | 70.9 (21.8) | 76.2 (19.0) | 54.5 (21.8) | <0.001 |
| Troponin I, mean (± SD) | 15.7 (20.6) | 16.2 (21.8) | 14.0 (16.5) | 0.442 |
| Medication at baseline, n (%) | ||||
| Beta blockers | 331 (39.1) | 205 (32.0) | 126 (60.9) | <0.001 |
| ACEi/ARB | 289 (34.1) | 182 (28.4) | 107 (51.7) | <0.001 |
| ASA/Plavix | 344 (40.6) | 204 (31.9) | 140 (67.6) | <0.001 |
| Diuretics | 165 (19.5) | 80 (12.5) | 85 (41.1) | <0.001 |
| Statins | 275 (32.5) | 176 (27.5) | 99 (47.8) | <0.001 |
| PCI, n (%) | 421 (49.7) | 360 (56.2) | 61 (29.4) | <0.001 |
| CABG, n (%) | 50 (5.9) | 27 (4.2) | 23 (11.1) | <0.001 |
Non-survivors after five years.
SD: standard deviation; STEMI: ST-elevation myocardial infarction; LVEF: left ventricular ejection fraction; AMI: acute myocardial infarction; HbA1c: glycated haemoglobin; eGFR: estimated glomerular filtration rate; ACEi: angiotensin-converting enzyme inhibitor; ARB: angiotensin receptor blocker; ASA: acetylsalicylic acid; PCI: percutaneous coronary intervention; CABG: coronary artery bypass grafting
Characteristics of the 847 patients with acute myocardial infarction.
| Variables | All patients | Survivors | Non-survivorsa | p-value |
|---|---|---|---|---|
| Patients, n | 847 | 640 | 207 | |
| Age, years, mean (±SD) | 70 (11.8) | 67 (10.8) | 78 (10.4) | <0.001 |
| Female, n (%) | 277 (32.7) | 197 (30.7) | 80 (38.6) | 0.036 |
| STEMI, n (%) | 256 (30.2) | 212 (33.1) | 44 (21.3) | 0.001 |
| Body mass index > 30 kg/m2, n (%) | 178 (21.0) | 143 (22.3) | 35 (16.9) | <0.001 |
| Systolic blood pressure, mmHg, mean (± SD) | 127 (21.8) | 127 (20.6) | 129 (24.6) | 0.388 |
| LVEF < 45%, n (%) | 196 (23.1) | 112 (17.5) | 84 (40.6) | <0.001 |
| LVEF < 35%, n (%) | 80 (9.4) | 36 (5.6) | 44 (21.3) | <0.001 |
| Current or previous smokers, n (%) | 550 (64.9) | 423 (66.1) | 127 (61.4) | 0.460 |
| Hypertension, n (%) | 466 (55.0) | 336 (52.5) | 130 (62.8) | 0.010 |
| Hypercholesterolaemia, n (%) | 265 (31.3) | 195 (30.5) | 70 (33.8) | 0.367 |
| Diabetes mellitus, n (%) | 152 (17.9) | 92 (14.4) | 60 (29.0) | <0.001 |
| Angina pectoris, n (%) | 221 (26.1) | 121 (18.9) | 100 (48.3) | <0.001 |
| AMI, n (%) | 187 (22.1) | 103 (16.1) | 84 (40.6) | <0.001 |
| Stroke, n (%) | 64 (7.6) | 32 (5.0) | 32 (15.5) | <0.001 |
| Heart failure, n (%) | 75 (8.9) | 30 (4.7) | 45 (21.7) | <0.001 |
| HbA1c (Mono-S) ≥ 5.0%, n (%) | 282 (33.2) | 187 (29.2) | 95 (45.8) | <0.001 |
| eGFR, ml/min per 1.73 m2, mean (± SD) | 70.9 (21.8) | 76.2 (19.0) | 54.5 (21.8) | <0.001 |
| Troponin I, mean (± SD) | 15.7 (20.6) | 16.2 (21.8) | 14.0 (16.5) | 0.442 |
| Medication at baseline, n (%) | ||||
| Beta blockers | 331 (39.1) | 205 (32.0) | 126 (60.9) | <0.001 |
| ACEi/ARB | 289 (34.1) | 182 (28.4) | 107 (51.7) | <0.001 |
| ASA/Plavix | 344 (40.6) | 204 (31.9) | 140 (67.6) | <0.001 |
| Diuretics | 165 (19.5) | 80 (12.5) | 85 (41.1) | <0.001 |
| Statins | 275 (32.5) | 176 (27.5) | 99 (47.8) | <0.001 |
| PCI, n (%) | 421 (49.7) | 360 (56.2) | 61 (29.4) | <0.001 |
| CABG, n (%) | 50 (5.9) | 27 (4.2) | 23 (11.1) | <0.001 |
| Variables | All patients | Survivors | Non-survivorsa | p-value |
|---|---|---|---|---|
| Patients, n | 847 | 640 | 207 | |
| Age, years, mean (±SD) | 70 (11.8) | 67 (10.8) | 78 (10.4) | <0.001 |
| Female, n (%) | 277 (32.7) | 197 (30.7) | 80 (38.6) | 0.036 |
| STEMI, n (%) | 256 (30.2) | 212 (33.1) | 44 (21.3) | 0.001 |
| Body mass index > 30 kg/m2, n (%) | 178 (21.0) | 143 (22.3) | 35 (16.9) | <0.001 |
| Systolic blood pressure, mmHg, mean (± SD) | 127 (21.8) | 127 (20.6) | 129 (24.6) | 0.388 |
| LVEF < 45%, n (%) | 196 (23.1) | 112 (17.5) | 84 (40.6) | <0.001 |
| LVEF < 35%, n (%) | 80 (9.4) | 36 (5.6) | 44 (21.3) | <0.001 |
| Current or previous smokers, n (%) | 550 (64.9) | 423 (66.1) | 127 (61.4) | 0.460 |
| Hypertension, n (%) | 466 (55.0) | 336 (52.5) | 130 (62.8) | 0.010 |
| Hypercholesterolaemia, n (%) | 265 (31.3) | 195 (30.5) | 70 (33.8) | 0.367 |
| Diabetes mellitus, n (%) | 152 (17.9) | 92 (14.4) | 60 (29.0) | <0.001 |
| Angina pectoris, n (%) | 221 (26.1) | 121 (18.9) | 100 (48.3) | <0.001 |
| AMI, n (%) | 187 (22.1) | 103 (16.1) | 84 (40.6) | <0.001 |
| Stroke, n (%) | 64 (7.6) | 32 (5.0) | 32 (15.5) | <0.001 |
| Heart failure, n (%) | 75 (8.9) | 30 (4.7) | 45 (21.7) | <0.001 |
| HbA1c (Mono-S) ≥ 5.0%, n (%) | 282 (33.2) | 187 (29.2) | 95 (45.8) | <0.001 |
| eGFR, ml/min per 1.73 m2, mean (± SD) | 70.9 (21.8) | 76.2 (19.0) | 54.5 (21.8) | <0.001 |
| Troponin I, mean (± SD) | 15.7 (20.6) | 16.2 (21.8) | 14.0 (16.5) | 0.442 |
| Medication at baseline, n (%) | ||||
| Beta blockers | 331 (39.1) | 205 (32.0) | 126 (60.9) | <0.001 |
| ACEi/ARB | 289 (34.1) | 182 (28.4) | 107 (51.7) | <0.001 |
| ASA/Plavix | 344 (40.6) | 204 (31.9) | 140 (67.6) | <0.001 |
| Diuretics | 165 (19.5) | 80 (12.5) | 85 (41.1) | <0.001 |
| Statins | 275 (32.5) | 176 (27.5) | 99 (47.8) | <0.001 |
| PCI, n (%) | 421 (49.7) | 360 (56.2) | 61 (29.4) | <0.001 |
| CABG, n (%) | 50 (5.9) | 27 (4.2) | 23 (11.1) | <0.001 |
Non-survivors after five years.
SD: standard deviation; STEMI: ST-elevation myocardial infarction; LVEF: left ventricular ejection fraction; AMI: acute myocardial infarction; HbA1c: glycated haemoglobin; eGFR: estimated glomerular filtration rate; ACEi: angiotensin-converting enzyme inhibitor; ARB: angiotensin receptor blocker; ASA: acetylsalicylic acid; PCI: percutaneous coronary intervention; CABG: coronary artery bypass grafting
Uni- and multivariable Cox regression analysis showing the hazard ratios per unit change for GDF-15 and TRAIL-R2 biomarkers in predicting long-term all-cause mortality in 847 patients with acute myocardial infarction.
| Univariable analysis | Multivariable analysisa | |||
|---|---|---|---|---|
| HR (95% CI) | p-value | HR (95% CI) | p-value | |
| GDF-15 | 2.57 (2.31–2.85) | <0.001 | 1.25 (1.02–1.53) | 0.031 |
| TRAIL-R2 | 2.10 (1.94–2.27) | <0.001 | 1.37 (1.12–1.67) | 0.002 |
| Univariable analysis | Multivariable analysisa | |||
|---|---|---|---|---|
| HR (95% CI) | p-value | HR (95% CI) | p-value | |
| GDF-15 | 2.57 (2.31–2.85) | <0.001 | 1.25 (1.02–1.53) | 0.031 |
| TRAIL-R2 | 2.10 (1.94–2.27) | <0.001 | 1.37 (1.12–1.67) | 0.002 |
Adjusted for GDF-15, TRAIL-R2, age, gender, diabetes, previous myocardial infarction, stroke, heart failure, hypertension, hypercholesterolaemia, smoking, body mass index, ST-elevation myocardial infarction, left ventricular ejection fraction, logtroponin I, logeGFR, logNT-proBNP and logCRP.
GDF-15: growth differentiation factor 15; TRAIL-R2: tumour necrosis factor-related apoptosis-inducing ligand receptor 2; HR: hazard ratio per unit change; CI: confidence interval; eGFR: estimated glomerular filtration rate; NT-proBNP: N-terminal pro-brain natriuretic peptide; CRP: C-reactive protein
Uni- and multivariable Cox regression analysis showing the hazard ratios per unit change for GDF-15 and TRAIL-R2 biomarkers in predicting long-term all-cause mortality in 847 patients with acute myocardial infarction.
| Univariable analysis | Multivariable analysisa | |||
|---|---|---|---|---|
| HR (95% CI) | p-value | HR (95% CI) | p-value | |
| GDF-15 | 2.57 (2.31–2.85) | <0.001 | 1.25 (1.02–1.53) | 0.031 |
| TRAIL-R2 | 2.10 (1.94–2.27) | <0.001 | 1.37 (1.12–1.67) | 0.002 |
| Univariable analysis | Multivariable analysisa | |||
|---|---|---|---|---|
| HR (95% CI) | p-value | HR (95% CI) | p-value | |
| GDF-15 | 2.57 (2.31–2.85) | <0.001 | 1.25 (1.02–1.53) | 0.031 |
| TRAIL-R2 | 2.10 (1.94–2.27) | <0.001 | 1.37 (1.12–1.67) | 0.002 |
Adjusted for GDF-15, TRAIL-R2, age, gender, diabetes, previous myocardial infarction, stroke, heart failure, hypertension, hypercholesterolaemia, smoking, body mass index, ST-elevation myocardial infarction, left ventricular ejection fraction, logtroponin I, logeGFR, logNT-proBNP and logCRP.
GDF-15: growth differentiation factor 15; TRAIL-R2: tumour necrosis factor-related apoptosis-inducing ligand receptor 2; HR: hazard ratio per unit change; CI: confidence interval; eGFR: estimated glomerular filtration rate; NT-proBNP: N-terminal pro-brain natriuretic peptide; CRP: C-reactive protein
Results
Baseline characteristics are shown in Table 1.
Biomarkers
The penalized regression analysis yielded a large best predictor set of 32 biomarkers: adrenomedullin, growth differentiation factor 15 (GDF-15), E-selectin, interleukin-1 receptor antagonist protein, cystatin-B, galectin-3, proteinase-activated receptor 1, angiopoietin-1 receptor, interleukin-27 subunit alpha, tumour necrosis factor-related apoptosis-inducing ligand receptor 2 (TRAIL-R2), fibroblast growth factor 23, tumour necrosis factor ligand superfamily member 14, myeloperoxidase, growth hormone, renin, chitinase-3-like protein 1, T-cell immunoglobulin and mucin domain, matrix metalloproteinase-10, urokinase plasminogen activator surface receptor, cathepsin D, C-X-C motif chemokine 16, galanin peptides, endothelial cell-specific molecule 1, matrix metalloproteinase-12, spondin-1, cathepsin L1, fatty acid-binding protein 4, leptin, C-C motif chemokine 20, ovarian cancer-related tumour marker CA 125, platelet endothelial cell adhesion molecule, and NT-proBNP. This set of biomarkers showed an accuracy of predicting long-term mortality with a ROC AUC of 0.85. The penalized regression analysis indicated that we could reduce this set of biomarkers to only two without a loss in accuracy, with a ROC AUC of 0.85. The two most powerful biomarkers in predicting long-term all-cause mortality were GDF-15 and TRAIL-R2.
The use of age and sex in the predicting equation (Model I) yielded a discriminatory accuracy for differentiating survivors from non-survivors of 0.79 in terms of a cross-validation ROC AUC. Model II, using conventional risk factors and biomarkers, yielded a ROC AUC of 0.86. Combining GDF-15 and TRAIL-R2 with age and sex (Model III) generated a ROC AUC of 0.86. Merging conventional risk factors and biomarkers with the selected PEA-biomarkers GDF-15 and TRAIL-R2 (Model IV) into the prediction equation provided a ROC AUC of 0.88. NRI was 0.09, 95% confidence interval (0.04–0.15) (p 0.001). The likelihood ratio test showed a significant increase in model fit by inclusion of the selected PEA-biomarkers in addition to conventional risk factors and biomarkers (p < 0.001). Using Model IV, the multivariable Cox regression analysis showed that GDF-15 and TRAIL-R2 were independent predictors of all-cause mortality (Table 2). Figure 1 illustrates the relationship between the observed and the predicted risk of death using Model IV.
This bar chart shows the associations between the observed and predicted all-cause five-year mortality in 847 patients with an acute myocardial infarction (AMI) using the biomarkers growth differentiation factor 15, tumour necrosis factor-related apoptosis-inducing ligand receptor 2, age, gender, diabetes, previous AMI, stroke, heart failure, hypertension, hypercholesterolaemia, smoking, body mass index, ST-elevation myocardial infarction, left ventricular ejection fraction, estimated glomerular filtration rate, C-reactive protein, troponin I and N-terminal pro-brain natriuretic peptide (model IV) as risk predictors. The horizontal axis shows the predicted risk of death after five years as a percentage. The number of patients in each stratum is shown on each bar. The vertical axis and bars show the observed risk of death after five years as a percentage.
In a multivariable linear regression analysis including risk factors and biomarkers, GDF-15 was independently associated with age, diabetes, smoking, LVEF, hypertension, troponin I, NT-proBNP, CRP and eGFR (R2 = 0.563). TRAIL-R2 levels were independently associated with age, diabetes, heart failure, stroke, smoking, STEMI, LVEF, NT-proBNP, CRP and eGFR (R2 = 0.556).
The hazard ratios for the 81 analysed biomarkers adjusted for age are given in Supplementary Material Table 1 online.
Discussion
The multiplex PEA proteomics chip assay was applied to 847 patients with an AMI to uncover which of 92 biomarkers linked to CVD and inflammation best predicted long-term all-cause mortality. GDF-15 and TRAIL-R2 were the two most powerful biomarkers predicting all-cause long-term mortality. GDF-15 and TRAIL-R2 in combination with established risk factors and biomarkers showed a discriminating accuracy of separating survivors from non-survivors with a cross-validation ROC AUC of 0.88.
To our knowledge, this is the first study to evaluate the prognostic significance of a large set of biomarkers associated with cardiovascular or inflammatory disease, and to explore which biomarkers best predict long-term all-cause mortality after an AMI. The biomarkers contained in the proteomics multiplex PEA chip are based on numerous human proteins or mRNA expression levels that in the literature have been proposed to be linked to the inflammatory response in patients with CVD or neoplastic disorders. The PEA technique has been shown to be highly sensitive and specific for detection of the target specific proteins, mainly because of the requirement for dual and proximal binding of the PEA probes.7
GDF-15 belongs to the transforming growth factor-beta superfamily proteins and is secreted as a 28 kDa disulphide-linked dimer. Under normal conditions, GDF-15 is expressed in low concentrations in most organs and upregulated because of tissue injury or remodelling. GDF-15 appears to be involved in regulating apoptosis, cell repair and cell growth, which are biological processes observed in cardiovascular and neoplastic disorders.14–17 These data confirm previous studies showing that GDF-15 is a powerful predictor of all-cause mortality.14,18–20 Moreover, our results extend previous findings showing that GDF-15 is one of the most robust biomarkers among the 81 studied in predicting all-cause mortality. The reason that GDF-15 is such a strong predictor of mortality is poorly understood, but as it is expressed in almost all organs, it is conceivable that it carries integrated information about various disease processes despite its nature and organ locations. It has also been demonstrated that an increased GDF-15 expression level might not be simply a consequence of a disease process but might play an active role in the pathobiology of both CVD and neoplastic diseases.17,21,22 This is supported by our present and previous findings showing that GDF-15 is associated with several established clinical and biochemical CVD risk markers.18,20,23 Therefore, these associations might explain why a single measure of GDF-15, taken within 72 h after hospitalization for an AMI, provided incremental long-term prognostic information.
TRAIL-R2 is one of several known tumour necrosis factor receptors (TNF-Rs). TNF-Rs are found in most tissues and present as trans-cell membrane proteins. Binding of tumour necrosis factor-related apoptosis-inducing ligands (TRAILs) to their receptors might induce pleiotropic cellular responses including activation, proliferation, differentiation and apoptotic signalling cascades leading to cell death.24,25 As members of the TNF-R family have an ability to induce intracellular death signals, the receptors are closely involved in neoplastic cell growth.26 The finding that TRAIL-R2 was one of the two most powerful predictors in the PEA chip in predicting long-term mortality in patients with an AMI is an intriguing but unexpected finding that has not been described previously. However, similar to GDF-15, TRAIL-R2 was associated with several cardiovascular risk factors, and because both TRAIL-R2 and GDF-15 are expressed in most tissues,27 this suggests that TRAIL-R2, as well as GDF-15, is an unspecific marker that is unregulated in several disease states reflecting an overall disease burden more than a specific CVD biomarker.
Recent studies have shown that low levels of soluble TRAIL (TRAIL-R ligand) in patients with AMI are associated with a worse prognosis. It has been suggested that low TRAIL levels might represent reduced inhibition of tissue level apoptosis.28,29 As apoptosis is a silent progressive cell death far beyond an acute event, it might contribute to slow cardiac remodelling post-AMI. Our data confirmed that low levels of soluble TRAIL were associated with increased all-cause mortality. However, low levels of TRAIL might be attributable to an upregulation of the TRAIL receptors as a potential increased receptor–ligand interaction reducing the availability of TRAIL-R. Although hypothetical, the significant biomarker expressions found here might represent intensified apoptotic activity or signify parts of complex molecular pathways such as inflammatory regulating activity in patients with AMI. This hypothesis is supported by our findings showing that upregulation of several TNF receptors such as TNFR1, TNFR2 and OPG were associated with a worse prognosis.
As both the TNF ligand family and the transforming growth factor-beta family constitute cytokines involved in multifaceted biological pathways in regulating cell death, cell proliferation and the inflammatory response to cell injury the present data propose that both cytokine families are involved in maintaining a balance between cell replication and cell death after an ischaemic or reperfusion injury.30 A future understanding of these interactions might have both a prognostic and a therapeutic objective.
Limitations
There were limitations to this study that need consideration. The study sample was limited to elderly Caucasian patients with AMI, and care should be used in drawing conclusions concerning other age and ethnic groups. Among the 1459 eligible patients with AMI, 13.8% (n = 201) did not fulfil the criteria of the 72-h limit to blood sampling. The local coronary intervention laboratory was only operational from 08:00 to 17:00 on Mondays to Fridays. Patients presenting with an AMI outside those hours were treated in nearby hospitals, so the blood sampling limit was frequently exceeded. The 201 patients not included should represent a random sample of patients, as they were admitted to other hospitals primarily because of the time of their AMI, so an extensive selection bias is not to be expected. Furthermore, 161 patients were excluded because of missing protocol data, of which a majority could be explained by the abovementioned reason of nearby hospital treatment or by a reduced echocardiographic image quality. Among the 1097 remaining patients with an AMI, 250 were excluded for other reasons (such as dementia or other severe diseases), which may have introduced a selection bias towards less critically ill patients in the study population.
The PEA chip technique does not permit an absolute quantification of the target proteins, so translation into clinically relevant cut-off values is not possible. Moreover, mainly because of the availability and quality of antibodies, there are several biomarkers associated with CVD or inflammation that are not incorporated in the present multiplex PEA chip. The plasma samples were stored for up to nine years before being analysed, so we cannot exclude the possibility that different protein stabilities might have influenced the analysis. However, the plasma samples were collected and stored consistently, which should have minimized any pre-analytical bias. GDF-15 is a well-characterized protein known for its plasma stability in vivo, and the plasma concentration remains steady far beyond an acute event. On the other hand, the finding indicating TRAIL-R2 as a predictor of all-cause mortality preceding an AMI needs to be prospectively validated, as protein co-variations might have generated statistical biases.
As the blood samples were taken within 72 h after an AMI, we do not know to what extent, or which, biomarkers exhibited an acute phase expression. Therefore, the expression of biomarkers found here is only valid for samples taken within 72 h after an AMI.
Author contribution
All the authors contributed to the conception or design of the work. ES, PW, EH, PH and JL contributed to the acquisition, analysis or interpretation of the data for the work. ES, PW and EH drafted the manuscript. PH, AS and JL critically revised the manuscript. All authors gave final approval and agreed to be accountable for all aspects of work ensuring integrity and accuracy.
Acknowledgements
The authors thank the participants and the staff of the VaMIS study, particularly Marja-Leena Ojutkangas, Annika Kärnsund, Göran Nilsson and Mathias Rehn for their contributions.
Declaration of conflicting interests
The author(s) declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: Agneta Siegbahn has received consulting fees from Olink Bioscience, Uppsala, Sweden. Olink Bioscience did not have any involvement in the study design, the data collection or analysis, or the publication process. All the other authors have declared no conflict of interest.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was supported by grants from Sparbanksstiftelsen Nya, the county of Västmanland, Selanders Stiftelse and the Swedish Medical Association.
Comments