Circulatory factors associated with function and prognosis in patients with severe heart failure

Patients were enrolled prospectively from the outpatient clinic at Karolinska University Hospital (Stockholm, Sweden). Patients with moderate to severe and stable chronic HF, defined as functional New York Heart Association (NYHA) class III disease, with no acute hospital admission within the last 8 weeks and with a left ventricular ejection fraction (LVEF) < 35% were eligible for inclusion. NYHA III was defined as a self-reported maximum continuous walking distance of no more than 200 m with dyspnea as a limiting factor. Patients fulfilling the inclusion criteria were invited to participate during scheduled visits to the attending cardiologist. After receiving written and verbal information about the study design, objectives, and potential risks, all participants signed a written informed consent form. Sixty-six patients were recruited between May 2009 and June 2013. To access a group of control subjects as closely matched as possible to patients with HF in terms of age, comorbidities, and physical activity, control subjects were enrolled among patients admitted to the outpatient clinic with symptoms suggestive of HF, but where the presence of HF was ruled out based on LVEF > 50% and a level of N-terminal pro-brain natriuretic peptide (NT-proBNP) < 300 ng/L. Echocardiographic signs of diastolic dysfunction were not an exclusion criterion in the control group, as long as the NT-proBNP level was within the normal range. The validity of the study group was checked by comparing baseline characteristics of the study group with the underlying cohort of outpatients enlisted at the hospital. The study was approved by the Regional Ethical Review Board in Stockholm, Sweden (no. 2007/1410-31/3) and carried out in accordance with the International Code of Medical Ethics of the World Medical Association (Declaration of Helsinki, 5th revision).

All participants (patients and controls) underwent echocardiography, and measurements of mean arterial pressure (MAP) and heart rate (HR) at rest (Table 1). Blood samples were collected with the subject in a fasting state in the morning in EDTA-coated tubes, and then centrifuged; plasma was aliquoted and stored at − 70 °C until analysis (Table 1). All subjects performed a symptom-limited cardiopulmonary exercise test (CPX) to assess VO_2peak (Table 1). The CPX consisted of maximum symptom-limited exercise either on a cycle ergometer (increments of 10 W every 60 s) or on a treadmill (1 m/s with a stepwise increase in the angle of 0.5°/min). In every CPX performed, continuous assessment of gas-exchange data (Vmax, SensorMedics, Anaheim, CA, USA) was performed. The exercise was terminated due to volitional exhaustion and/or the patient’s inability to maintain the speed of 1 m/s (treadmill) or a cadence of 60 rpm (cycle ergometer) despite strong verbal encouragement. Echocardiographic measurements were carried out in accordance with clinical guidelines (Vivid 7, General Electrics Healthcare, Little Chalfont, United Kingdom) and was analyzed by an echocardiographer blinded to the specific clinical history of the patient. Left ventricular end-diastolic diameter (LVEDD) was measured and left ventricular ejection fraction (LVEF) was calculated using the biplane Simpson's rule. Diastolic function was estimated based on mitral inflow deceleration time (DT) and for patients with sinus-rhythm mitral in flow E/A-wave. Daily activity was assessed in all participants by accelerometers (GT3X; Actigraph, Pensacola, FL, USA), which were mailed to all patients within 6 weeks of the CPX measurements. The patients were instructed to attach the accelerometer to their waist belt upon rising in the morning and to remove it only for showering, bathing, and sleeping. The monitors were set to begin collecting data 1 day before the delivery date, as estimated by the postal service, and to continue recording data until they were downloaded. The patients were asked to return the monitor by mail using a prepaid return envelope after having worn it for seven consecutive days. Raw data collected by the accelerometer were integrated into 60-s epochs using ActiLife software with the normal filter option and expressed as counts per minutes (cpm). Wear time was estimated using the algorithm described by Troiano et al. [78]. Nonwear time was defined as 60 consecutive minutes of 0 cpm, with allowance for 1–2 min of 0–99 cpm during this time. Patients with an estimated wear time of < 3 days were eliminated from further analysis (n = 3). In addition to analysis of time spent active, skewness was analyzed as a potentially important prognostic measure of variance in the level of physical activity [56]. Data on mortality and cause of death were obtained from the Swedish national cause-of-death registry in June 2015.

Factors were quantified from single samples using a multiplex immunoassay (Proseek Multiplex 96 × 96 CVD III; Olink Bioscience, Uppsala, Sweden), which is a 92-plex immunoassay based on a proximity ligation extension assay. Proximity extension assays use target-specific antibody pairs that are linked to DNA strands that, upon simultaneous binding to the target analyte, create a real-time polymerase chain reaction amplicon in a proximity-dependent manner enabled by the action of a DNA polymerase. The intra-assay coefficient of variation (CV) ranges between 5 and 11% (mean 6%), and the inter-assay CV ranges between 9 and 39% (mean 15%) [8]. All factors in the assay have been validated and relevant spike-in experiments have been performed to ensure that there is no cross-reactivity between the different biomarkers [51]. Further information about reproducibility and validation can be found at https://​www.​olink.​com. NT-proBNP levels were analyzed by electrochemiluminescence immunoassay using cobase immunoassay analyzers (Roche Diagnostics, Rotkreuz, Switzerland). Creatinine clearance was calculated according to the Cockcroft–Gault formula.

Given the large set of variables, the reduction of the dimension of our covariate list was performed through variable clustering and dimensionality reduction through PCA Bi plots was used to characterize the variance of the individual variables and to identify collinearities. Variables derived from echocardiographic assessment [LVEF, left ventricular end-diastolic diameter (LVEDD), peak systolic velocity, septal, and lateral E′ values, LA-area, estimated PA-pressure, and tricuspid annular plane systolic excursion (TAPSE)], clinical examination and exercise test (MAP, heart rate, and VO_2peak), and daily physical activity (time spent active, time spent inactive, and skewness) were analyzed. All variables were scaled to unit variance and mean-centered. The concentrations of the Proseek biomarkers are expressed as arbitrary units and were mean-centered prior to analysis because of the lack of specific calibrators for the assays. Principal components were calculated using the ‘prcomp’ function in the R statistics package (R Foundation for Statistical Computing, Vienna, Austria), and biplots were generated using the ‘factominer’ and ‘ggplot2’ libraries. Ontological enrichment analysis was performed using the web-based Database for Annotation, Visualization, and Integrated Discovery (DAVID).

For the case–control analysis, we used OPLS, which is similar to PCA, but was developed to handle classification rather than correlation. OPLS regression is particularly suited when the matrix of predictors has more variables than observations and when there is multicollinearity among X values [77]. An OPLS-DA classification model was constructed using the rOPLS-library in R [74], in which the contribution of each variable is represented by a loading value compared with what is predicted, in this case, patient or control. An OPLS model finds the multidimensional direction in the X space that explains the maximal variance in the Y space. The validity of the model (Q₂ value) was assessed by bootstrapped cross validation, yielding 95% confidence intervals for the contribution of each of the variables in the group classification.

Given the high covariance of the physiological measures, the dimensions of our covariate list were reduced utilizing summarized measures of each category derived from the principal components [79]. This approach retains variance/information that is shared between several of the variables and discards less systematic variance that is more likely to contain noise [84]. This process results in a summary score for each category, as a linear combination of variables within the category summarized for each patient. The two first principal components were kept from each of the categories, physical capacity, daily physical activity, and myocardial function. To develop networks representing the interaction between plasma factors and HF pathophysiology, a global interaction network was constructed using ARACNE, an algorithm for the reconstruction of gene regulatory networks in a mammalian cellular context. This is a validated network inference technique that infers the interaction between pairs of variables using a measure of correlation called mutual information (MI) [45]. MI between each pair of variables was tested using a permutation-based statistical test. ARACNE offers numerous advantages over more traditional measures of correlation, including the ability to spot nonlinear correlations, which are very effective in identifying biologically relevant connections [26]. Significant connections were identified by applying a P value threshold for significant MI values corresponding to a false discovery rate of 5%. To avoid potential overfitting in the ARACNE network, a correlation matrix was constructed, with confidence intervals estimated for each pairwise correlation through bootstrapping. Only pairwise correlations with an absolute correlation of > 0.2 and a false discovery rate (FDR) of < 5% were retained. The resulting network was visualized using the software application Cytoscape v. 3.5 (https://​www.​cytoscape.​org) [70].

Sixty-six patients with HF and 28 controls were enrolled in the study over a period of 3 years. The demographic and clinical characteristics of the subjects included are presented in Table 1. Within the HF patient cohort, VO_2peak was on average 13.4 mL/kg (range 6.1–22.5 mL/kg), and 23.8 mL/kg/min (17.7–25.4 mL/kg) for controls. Among the patients, time spent physically active (>100 counts/min) was 950 (IQR 541–1366) min/week, and 2182 (IQR 1592–2790) min/week for controls. The follow-up time for surviving patients with HF was 3 years (range 1–5). Of the 66 monitored patients, 42 died during follow-up (median time 1.8 years). All mortality events were categorized as cardiovascular in nature. To confirm validity of the study group, baseline characteristics of the patients were compared with data on all patients enlisted as heart failure outpatients at the hospital at the time of enrollment (n = 1467). The patients enrolled in the study where similar to the average patient with regards to age, body composition, comorbidities, and underlying conditions but hade worse cardiac function and functional status, fewer females and had more aggressive medical treatment than the average patient (Supplemental Table).

There was a high overall correlation across the different factors with a mean Pearson correlation coefficient of R² = 0.25. For these, PCA captured ~ 48% of the total variance explained by the first three principal components. Many of the factors assessed in the Proseek assay belong to the same family (e.g., tumor necrosis factor [TNF]-alpha, leukocyte adhesion molecules) or share similar functions. Indeed, ontological enrichment analysis showed that the assay as a whole is highly enriched for the cellular factors inflammation, cellular adhesion, and migration (Fig. 1). Despite this innate bias, clusters of highly correlated factors were tested for functional enrichment. A correlation matrix was constructed, and cluster analysis was performed based on pairwise correlations between variables. This revealed that the measured factors formed clearly distinguishable clusters with a high degree of correlation across observations, in which factors with similar biological functions cluster more closely together. That factors correlate according to their biological functions serves as a biological quality control; if data contained mainly noise or were caused by methodological bias, the correlations would have been random and/or unrelated to biological functions. Ontological enrichment analysis using DAVID identified three distinguishable clusters with different biological functions (Fig. 1). The first cluster was highly overrepresented by factors with described biological functions related to the regulation of cell growth and cell adhesion. The second cluster was most enriched for extracellular matrix organization factors and angiogenesis. The third cluster was enriched for inflammatory response markers. As an additional control for our measurements, NT-proBNP was analyzed using an immunofluorescence assay, in addition to the measurements performed using the Proseek platform. The agreement between the two methods was “very good”, with an R² correlation coefficient of 0.91, and a difference within 0.5 standard deviations for all observations shown by Bland–Altman analysis.

Based on the high covariance among the investigated circulatory factors, the differences in the concentrations analyzed in plasma between patients with HF and controls were explored through a multivariate approach. The OPLS model differentiating patients from controls correctly classified 84% (R²Y = 0.84) of the observations in the data set, with a predictive (Q₂ value) of 0.71 after cross validation (Fig. 2). The model identified 39 factors that contributed significantly to the model (18 were higher and 21 were lower in patients vs controls, respectively). The results are summarized in Table 2, together with the corresponding parametrically tested FDRs. As implied in the original methodology paper [74], there was a strong correlation between Variable Importance in Projection from the OPLS-DA classifier and P values based on parametric testing. A VIP threshold from the OPLS-DA of 0.8 corresponds to a parametric FDR of 6% (Table 2).

The variables derived from the clinical characterization of the patients were investigated with a similar approach to that described above for factors measured in the serum. The mean Pearson correlation coefficient across variables was 0.25, and PCA on all physiological variables combined showed high overall correlation between all measured variables; 75% of the overall variance could be captured with the first three principal components. Thus, while keeping most of the information intact, it was possible to summarize the clinical examination data into three principal components. This scaled down the number of dimensions and considerably reduced the background noise. The variables with the highest loading on principal component 1 were VO_2peak, MAP, heart rate, and daily physical activity; on principal component 2, the measures of both systolic (LVEF and peak systolic velocity) and diastolic dysfunction (E′, estimated PA-pressure, E-max and LA-area) had the highest loading. Physical capacity and heart rate contributed to both principal components 1 and 2, illustrating that they share some variance with both daily physical activity and measures of systolic function (Fig. 3). To validate and visualize prognostic information, a bi-plot was constructed in which data on mortality and time to mortality events were added to the PCA (Fig. 3) and analyzed with a Cox regression model. As expected, the prognostic utility of the investigated physiological variables was retained when summarized on PCA.

To identify associations between plasma factors and pathophysiological changes in HF, the circulating factors in the patients with HF were associated with the measured physiological variables with MI network inference, and covariance was calculated; that is, how large a portion of the variance in one variable can be accounted for from the variance of a second variable. The resulting network contained 76 unique factors with 141 significant edges relating the investigated physiological categories of variables (Fig. 4a). Of the physiological categories, physical capacity and daily physical activity were the most centrally placed hubs, and 48 of the 76 factors were connected to both physical capacity and daily physical activity. Apart from having the largest number of significant connections, physical capacity also had the highest number of connections to unique factors (Fig. 4b). The components derived from echocardiography, such as myocardial function, had fewer connections to factors, with only 20 significant edges, and these were shared with the other physiological categories (Fig. 4b). Of these 20 factors, 10 were among the 39 factors that differed significantly between patients and controls (Tables 2, 3). Some of these factors are considered to reflect cardiac stretching (ST2 and NT-proBNP), and others are linked to inflammatory responses (members of the TNF-alpha family; TNF-R2 and TNF-R1). The network analysis also identified some relatively or completely novel factors in the field of HF: growth differentiation factor 15 (GDF15), insulin-like growth factor-binding protein 7 (IGFBP7), transferrin receptor protein 1 (TfR1), and galactin-4. However, GDF15 and IGFBP7 have already been reported as potentially important markers of disease severity in HF, and are induced by hemodynamic and inflammatory stress [6, 30].

The ten factors that were expressed differently between patients with HF and controls were analyzed further to test whether they were linked with HF prognosis. This was tested by both a multivariate approach and through Cox regression. The PCA connecting physiological components to prognosis were recalculated substituting the physiological variables with the ten factors (Fig. 5a, b). As expected, because the factors were selected based on covariance and MI with prognostically important clinical variables, the combined factors reflected this prognostic value. The central network factors displayed a similar distribution of the variance as did the physiological variables (65% of variance retained within the first two components), and the HR of mortality events was 4.2 (P < 0.001; first component) and 1.2 [not significant (NS); second component]. In addition to this multivariate approach, the TEN network factors were also tested for prognostic value one-by-one in a Cox regression analysis, both as single independent variables and controlling for important confounders (age, estimated glomerular filtration rate [eGFR], VO_2peak, and LVEF; Table 3). All of the network factors except for PON3 and TR-AP were significantly associated with all-cause mortality (FDR < 5%) in univariate analysis (Table 3). Controlling for established clinical risk factors, TfR1, GDF15, and galectin remained significantly associated with all-cause mortality at FDR < 5% (Table 3).

Despite the effort made to design and conduct an integrative investigation analyzing factors related to function and prognosis of HF, this study had some limitations that need to be acknowledged. In contrast to other studies with nontargeted approaches, the analytical strategy used here did not identify possible pathophysiological and prognostic markers that are unrelated to physical capacity, physical activity, and echocardiographic estimates of disease severity. Moreover, the samples analyzed in this study originated from a homogeneous cohort of elderly patients of European origin with severe HF from a single center, which might limit the external validity of the results. In addition, several potentially confounding features differed between the patient and the control group. Because the proximity extension assay used here to quantify the biomarkers does not allow an absolute quantification of the target proteins, translation into clinically relevant cutoff values is not possible. In addition, mainly because of the availability and quality of antibodies, there are several biomarkers associated with cardiovascular disease or inflammation that are not incorporated in the current multiplex PEA chip. Because the PEA is a form of widespan-targeted analysis, consisting of a curated panel of factors selected for potential relevance in the context of cardiovascular disease, bias is introduced into the analysis, in which inflammatory factors and markers known to be sensitive to cardiac stretch are highly enriched. Therefore, the networks generated could be biased toward these biological functions. Notwithstanding these limitations, we believe that our investigation contributes new understanding of the molecular networks linking the failing heart and the loss of peripheral function in patients with HF.