Harnessing the Plasma Proteome to Predict Mortality in Heart Failure Subpopulations.
Mar 7, 2025DOI: 10.1161/circheartfailure.123.011208Journal ArticleAHA
Written by: Jessica ChadwickMichael A HinterbergFolkert W AsselbergsHannah BiegelEric BoersmaThomas P CappolaJulio A ChirinosJoseph CoreshPeter GanzDavid A GordonNatasha KureshiKelsey M LoupeyAlena OrlenkoRachel OstroffLaura SampsonSama ShresthaNancy K SweitzerStephen A WilliamsLei ZhaoIsabella KardysDavid E Lanfear
Pre-existing data access policies for each of the 4 cohort studies used in these analyses specify that research data requests can be submitted to their respective steering committees, will be promptly reviewed for consent, confidentiality, intellectual property restrictions, and privacy laws considerations, and will not unreasonably be refused.Left ventricular EF cutoffs for patient samples used in HFrEF and HFpEF model training and validation were based on the European Society of Cardiology and American Heart Association guidelines.3,16 Patients with mid-range left ventricular EF were not included in model training or validation.The training and validation of HF models were conducted in 3 separate HF cohorts: the PHFS (Penn HF Study),17 the ARIC study (Atherosclerosis Risk in Communities),18 and the HFGR (Henry Ford HF Genomic Registry).19 In addition, the longitudinal performance of the HFrEF and HFpEF models was tested in the Bio-SHiFT cohort (The Role of Biomarkers and Echocardiography in Prediction of Prognosis of Chronic HF Patients)20 and TOPCAT (Treatment of Preserved Cardiac Function HF With an Aldosterone Antagonist Trial).21 The flow diagram for training and validation of the HFrEF and HFpEF models from these cohorts is shown in Figure 1. A full description of each study can be found in the Supplemental Methods. Each study was approved by their respective institutional review boards, and informed consent was obtained from each patient. Demographics for the patients from each study included in the HFrEF and HFpEF model training and validation are described in Tables 1 and 2.HFrEF Cohort Description: Demographic and End Point Table for HFrEF Model Training and Validation Data SetsHFpEF Cohort Description: Demographic and End Point Table for HFpEF Model Training and Validation Data SetsHeart failure with reduced ejection fraction (HFrEF) and heart failure with preserved ejection fraction (HFpEF) proteomic risk score overview. Schematic overview of the HFrEF (A) and HFpEF (B) model training and validation program. Full demographic details of participants in each study are shown in Table 1. Study fractions outlined in red represent independent validation data sets, which were analyzed independently and merged and analyzed as a single data set. ARIC indicates Atherosclerosis Risk in Communities; Bio-SHiFT, The Role of Biomarkers and Echocardiography in Predication of Prognosis of Chronic Heart Failure Patients; HFGR, Henry Ford Heart Failure Genomic Registry; PHFS, Penn Heart Failure Study; and TOPCAT, Treatment of Preserved Cardiac Function Heart Failure with an Aldosterone Antagonist Trial.The SomaScan Assay and its performance characteristics have been described previously.9 The assay uses modified aptamers to quantify plasma levels of ≈5000 proteins (v4.0) with high specificity and limits of detection comparable to antibody-based assays. A full description of the SomaScan Assay, data normalization, calibration, and data quality control processes is provided in the Supplemental Methods. Specificity confirmation and precision of aptamer target binding for the protein features included in the HFrEF and HFpEF models are included in Tables S1 and S2.Fractions of data from multiple data sets were used as a common machine learning technique to develop models that mitigate overfitting while capturing disease heterogeneity and improving the generalizability of predictions22. Data sets were split into training, tuning (feature selection, parameter tuning, and cross-validation), and holdout validation (a holdout set that is used only to assess the final model and is not used for model development) data sets. Stratification of the data sets for HFrEF and HFpEF model training and validation is shown in Table S3. Proportions were chosen based on data availability (80% used for training, with additional data sets becoming available for tuning and validation) and ensured that demographics and mortality rate distributions were similar between the data set splits.The HFrEF and HFpEF models are accelerated failure time survival models, using a Weibull distribution, which predicts the risk of all-cause death at 1 year reported as a probability. As opposed to Cox proportional hazard models, accelerated failure time models allow for parameterization such that covariates either accelerate or decelerate the risk of HF mortality and provide a resulting predictive probability that can be more easily interpreted. Model features were chosen and fit using the least absolute shrinkage and selection operator. Both the HFrEF and HFpEF models were evaluated for discrimination using the Harrel C-index (the ability of a model to discriminate between random pairs of individuals with different times to event) and the 1-year area under the curve (AUC; the ability of the model to dichotomize the population into 2 classes at a defined timepoint). Models were required to exceed predefined performance criteria in a regulatory environment; a C-index and 1-year AUC ≥0.7 to ensure the proteomic models were at least equivalent to BNP discrimination in predicting death. A full description of statistical approaches and parameterization used to develop the HFrEF and HFpEF models and model equations can be found in the Supplemental Methods.Repeated assessment of the HFrEF model over time was tested on longitudinal samples from the Bio-SHiFT study (Table S4). One-year HFrEF risk predictions were calculated on all serially collected plasma samples in the study. Linear mixed models were used to characterize the baseline risk and how it changes over time in the full data set. The analysis was stratified by event status: participants who had a fatal event after the third timepoint (closest blood sample collected to the event) versus participants who did not have a fatal event during follow-up (2.5 years). Estimated marginal means were calculated for each timepoint by event status. Post hoc comparisons of mean risk at each timepoint and longitudinal changes in mean risk were conducted between the event groups. The Tukey and Dunnett adjusted P values were reported for each of the comparisons, respectively.The HFpEF model was assessed on longitudinal samples from the TOPCAT study (Table S5). One-year HFpEF risk predictions were calculated, and change in risk predictions was assessed for paired samples from baseline to 1 year in participants approaching an event (all-cause death) versus participants who remain event-free. Mean HFpEF risks were calculated for each timepoint by event status. A 1-tailed t test was used to compare mean risk at each timepoint and the change in mean risk from baseline between the event groups. A false discovery rate correction was applied to the resulting P values to correct for multiple comparisons at each timepoint.HFrEF and HFpEF model performance metrics (C-index and 1-year AUC) were calculated on all baseline samples in the Bio-SHiFT and TOPCAT study, respectively. The CIs on model performance metrics were calculated by bootstrapping the data set 200 times, calculating the metric each time, and then taking the 2.5% and 97.5% quantiles of the 200 values.Functional analyses were performed with the Database for Annotation, Visualization, and Integrated Discovery (v. 6.8),23 using UniProt accession IDs for gene labeling and Homo sapiens as species. Features chosen for model training (i.e., top 50 univariate features for HFrEF and the 39 overlapping features between univariate analyses conducted on the training and tuning data sets for HFpEF) were utilized for HFrEF and HFpEF enrichment analyses, respectively, while the entire v4.0 SomaScan Assay menu was used as background.Correlation between the NT-proBNP SOMAmer reagent (seq 7655.11) and clinically measured NT-proBNP (ELISA) was assessed using the Pearson correlation coefficient in participants from the HFGR cohort (with both measurements available). NT-proBNP values were measured using stored EDTA plasma samples, collected at enrollment, and stored at −80 °C, using an immunoelectrochemiluminescence assay on the Modular Analytics E 170 system (Roche Diagnostics), performed by ACS Biomarker (Amsterdam, Netherlands). This assay has <0.001% cross-reactivity with bioactive BNP. The Pearson correlation coefficient was calculated on the log10-transformed clinically measured NT-proBNP and SOMAmer reagent measured NT-proBNP. To evaluate the prognostic performance of NT-proBNP in HF mortality, the C-index and 1-year AUC between the log10-transformed relative fluorescence unit measurement of NT-proBNP and the days until death were calculated for patients with HFrEF and HFpEF, separately.Clinical data to calculate MAGGIC scores were only available for the HFGR cohort. The 80% holdout validation portion of this data set was used for comparative analyses against the HFrEF and HFpEF proteomic risk scores (PRSs). The C-index and 1-year AUC between the continuous MAGGIC score and the days until death were calculated for patients with HFrEF and HFpEF, separately. To test for differences in AUC between the PRS (HFrEF or HFpEF) and the clinical prognostic marker (MAGGIC score or NT-proBNP), a 1-sided DeLong test was applied using a significance level of 0.1. To test for differences in C-indices between the PRS (HFrEF or HFpEF) and the clinical prognostic marker (MAGGIC score or NT-proBNP), a 1-tailed Z-test (for right-censored survival data) was applied using a significance level of 0.05.The combined performance of the clinical prognostic marker (MAGGIC score or NT-proBNP) and PRS (HFrEF or HFpEF) was calculated using a linear combination: β1PRS +β2clinical prognostic marker, where β1 and β2 are coefficients from fitting a Cox proportional hazard regression model with the PRS (HFrEF or HFpEF) predictive probability and the clinical prognostic marker (MAGGIC score or NT-proBNP) as predictors on the corresponding validation data set. Performance was assessed using C-index and 1-year AUC, and significance was tested using a 1-sided Z-test and a 1-sided DeLong test, respectively. Analysis used R, version 3.4.3, in Linux on x86 64-bit architecture, with packages used for model building and assessment including caret (6.0.88), survival (3.2.11), and pROC (1.17.0.1).
RESULTS
Separate accelerated failure time models with Weibull distribution were trained and validated for patients with HFrEF and HFpEF. The HFrEF and HFpEF PRSs were trained in 1247 patients with HFrEF and 725 patients with HFpEF and validated in 762 patients with HFrEF and 785 patients with HFpEF, respectively, and predict the absolute likelihood of all-cause death within 1 year. There were 152 deaths (12.2%) within 1 year in the HFrEF training data and 72 deaths (9.4%) within 1 year in the HFrEF validation data. The event rate was lower in the HFpEF population with 46 deaths (6.3%) within 1 year in the HFpEF training data and 41 deaths (5.2%) within 1 year in the HFpEF validation data. The HFrEF and HFpEF models include 17- and 14-protein features (protein-binding aptamers), respectively, and both models use samples from across 3 independent studies (Figure 1; Tables 1 and 2). The resulting models are a balance of novel and established biomarkers; the proteins included in the HFrEF and HFpEF models are listed in Table S6, with their proportionate contribution to the model output, along with the coefficients of variation for each analyte. The linear combination of the aptamers for each model can be found in the Supplemental Methods.The HFrEF and HFpEF models had a C-index of 0.75 and 0.76 and a 1-year AUC of 0.78 and 0.81, respectively, in the validation data sets (Table 3). The median time to event was 4.9 months in the HFrEF validation data set and 4.2 months in the HFpEF validation data set. Model performance across all training and validation data sets can be found in Table 3. In addition to providing a numeric estimate of the probability of death within 1 year, relative risk bins were generated based on quartiles for both models. Survival curves stratified by relative risk bin for baseline HFrEF and HFpEF protein-risk are depicted in Figure 2. Observed 1-year fatal event rates across the 3 risk bins (low, medium, and high) were 1.3%, 9.3%, and 27.4% in HFrEF and 0%, 2.5%, and 13.3% in HFpEF, respectively.HFrEF and HFpEF PRS Performances: Performance Metrics for the HFrEF and HFpEF Training, Tuning, and Validation Data SetsHeart failure with reduced ejection fraction (HFrEF) and heart failure with preserved ejection fraction (HFpEF) proteomic risk score Kaplan-Meier. Observed survival probability and average time to event (death) in the combined HFrEF (A) and HFpEF (B) validation data set when stratified by relative risk bins. The Kaplan-Meier point estimate of observed events up to 12 months, stratified by risk bin: low risk, bottom quartile (absolute risk score 0≥x≤0.057 for HFrEF and 0≥x≤0.017 for HFpEF); medium risk, middle 2 quartiles (absolute risk score 0.057>x≤0.187 for HFrEF and 0.017>x≤0.081 for HFpEF); and high risk, top quartile (absolute risk score 0.187>x≤1.00 for HFrEF and 0.081>x≤1.00 for HFpEF).The baseline 1-year HFrEF PRS was 9.1% higher (P<0.001) in patients with a fatal event compared with those with no fatal event and was even higher in available follow-up samples for those who had fatal events subsequent to the follow-up sample (Table S7; Figure 3). HFrEF 1-year risk score for all-cause death increased by 3.0% more from baseline to timepoint 2 and 3.2% more from baseline to timepoint 3 in available samples for those who had fatal events compared with the no fatal event group (P<0.05). The performance (based on C-index and AUC) of the HFrEF model was also assessed in the Bio-SHiFT samples from the baseline timepoint and was shown to be comparable in this data set to the HFrEF training and validation data sets with a 1-year AUC of 0.78 and C-index of 0.77 (Table 3).Heart failure with reduced ejection fraction (HFrEF) and heart failure with preserved ejection fraction (HFpEF) proteomic risk score change in risk. A, Plot of estimated marginal means of predicted HFrEF (1-year) risk in participants who have a fatal event after timepoint 3 (event group) vs participants with no fatal event during follow-up (no event group). B, Plot of mean predicted HFpEF (1-year) risk in event group vs no event group. The error bars represent the 95% CI for each timepoint. The dashed horizontal lines represent the baseline risk prediction for each group. Bio-SHiFT indicates The Role of Biomarkers and Echocardiography in Predication of Prognosis of Chronic Heart Failure Patients; and TOPCAT, Treatment of Preserved Cardiac Function Heart Failure with an Aldosterone Antagonist Trial.Similarly, the 1-year HFpEF PRS was 8.6% higher (P<0.001) at baseline in patients with a fatal event compared with those with no fatal event and 10.9% higher at 1 year (P<0.001; Table S8; Figure 3). HFpEF PRS increased by 3.4% more from baseline to 1 year in participants with a fatal event versus no fatal event although this change was not statistically significant (P=0.81). HFpEF model performance was assessed in the baseline TOPCAT samples and shown to be comparable in this data set to the HFpEF training and validation data sets with a 1-year AUC of 0.75 and C-index of 0.74 (Table 3).A formal statistical functional enrichment analysis of the final protein features in the HFrEF and HFpEF models would not be valid because of intentional biases in machine learning, which limits the inclusion of correlated features. To understand biological pathways underlying the proteomic signal detected by our models, we explored gene ontology analyses of the subset of proteins utilized in the training of the HFrEF and HFpEF PRSs. Subsets of proteins selected for HFrEF model training were enriched for kidney failure disease pathology, while proteins selected for HFpEF model training were enriched for more heart-specific diseases, including cardiomyopathy, HF, heart decompensation, and tachycardia. All associations between individual model proteins and HF can be found in Tables S9 and S10.Because clinically measured NT-proBNP values were not available for many data sets, the prognostic performance of the NT-proBNP SOMAmer reagent (seq 7655.11) was used as the comparator. The NT-proBNP SOMAmer reagent has been previously shown to be highly correlated with clinically measured (enzyme-linked immunosorbent assay) NT-proBNP.10 NT-proBNP measurements by a standard clinical laboratory assay were only available for 1 study, HFGR. In the participants with both measurements available (n=494 paired measurements), the SOMAmer reagent NT-proBNP results were highly correlated with the clinically measured NT-proBNP. The Pearson correlation coefficient between these 2 measures was 0.92, indicating a strong linear correlation (Figure S1).To compare the performance of the NT-proBNP SOMAmer reagent as a predictor for all-cause death in validation data sets to that of the HF PRS signals, we calculated the C-index and 1-year AUCs separately for patients with HFrEF and HFpEF. The HFrEF and HFpEF PRSs had a significantly higher C-index than the NT-proBNP SOMAmer reagent alone (Table 4). NT-proBNP had a C-index of 0.73 for HFrEF patients and 0.72 for patients with HFpEF, whereas the HFrEF model had a C-index of 0.75 and the HFpEF model had a C-index of 0.76. Similarly, the 1-year survival AUCs for NT-proBNP were lower than the HFrEF and HFpEF PRSs, but only the HFpEF PRS was significant. Combining the PRS (HFrEF or HFpEF) with NT-proBNP did not significantly improve performance over PRS alone (Table 4).Comparative Analyses: HFrEF and HFpEF PRSs vs NT-proBNPClinical data required to calculate the MAGGIC score were also only available in the HFGR cohort. To compare the performance of MAGGIC score as a predictor for all-cause death in the 80% HFGR holdout validation data set to the HFrEF and HFpEF PRSs, we calculated the C-index and 1-year AUCs (for patients with HFrEF and HFpEF separately). The HFpEF PRS performed significantly better than the MAGGIC score (Table 5). MAGGIC score had a C-index of 0.71 and a 1-year AUC of 0.74, whereas the HFpEF PRS had a C-index of 0.79 and a 1-year AUC of 0.82. The HFrEF PRS performed comparably to the MAGGIC score, the MAGGIC score had a C-index of 0.75 and 1-year AUC of 0.84, and the HFrEF PRS had a C-index of 0.76 and 1-year AUC of 0.83. Combining the PRS (HFrEF or HFpEF) with the MAGGIC score did not significantly improve performance over the PRS alone (Table 5). However, for both HFrEF and HFpEF, the PRS score remains significantly associated with outcomes even when adjusted for MAGGIC score (HFrEF P<0.0001; HFpEF P<0.0001), suggesting that the PRS provides additional information not captured by MAGGIC score alone.Comparative Analyses: HFrEF and HFpEF PRSs vs MAGGICThe residual effects of age on the HFpEF and HFrEF models were also assessed because age is known to be a risk factor for HF mortality.24 During the modeling process, deviance residuals were assessed on the training data results, which represent a wide age range of 18.4 to 88.2 years of the HF population. While there is some positive correlation with age (more so in HFpEF), the low r2 values suggest that there is not a strong correlation with age (r2 of HFrEF residuals with events=0.064; P<0.0001; r2 of HFpEF residuals with events=0.157; P<0.0001).
DISCUSSION
In this study, we developed proteomic prognostic models that use a total of 27 unique circulating protein markers to predict 1-year mortality risk in patients with HFrEF and HFpEF using solely protein measurements. HFrEF and HFpEF PRSs were trained and validated across multiple large and diverse cohorts to better encompass the heterogeneity within HF populations and maximize generalizability. Stratification into distinct baseline risk groups, with the 2 middle-risk quartiles grouped together (to create the medium-risk bin), showed a wide dynamic range in 1-year observed fatal event rates between participants in the low versus high-risk bins. Longitudinal assessment of the PRS in the Bio-SHiFT study and TOPCAT trial showed fidelity to detecting changes in mortality risk over time. The HFpEF model performed significantly better than both NT-proBNP and the validated clinical risk score (MAGGIC), while the HFrEF model performed comparably to both NT-proBNP and MAGGIC scores. Improvements in AUC and C-index for the HFpEF PRS compared with NT-proBNP and MAGGIC score were comparable or greater than those seen in previous studies, for example, studies adding calcium score to traditional models for predicting the risk of incident coronary artery disease (C-index improved by 0.04–0.05), suggesting that these are clinically meaningful differences.25,26Our findings have strong external validity in the context of existing literature on risk assessment in HF and circulating proteins. Of the 27 unique proteins utilized in our HF prognosis models, 25 (93%) are associated with known HF pathophysiology and 16 (59%) have been previously identified in HF populations using the SomaScan Assay (Tables S9 and S10). Previously described biomarkers of HF prognosis incorporated in our models include NT-proBNP, GDF-15 (growth/differentiation factor 15), WFDC2 (WAP 4-disulfide core domain protein 2), ST2 (interleukin-1 receptor-like 1), and TnTc2 (cardiac muscle troponin T).3 We also identified novel proteins for the prediction of mortality in HFpEF: ALP (antileukoproteinase) and chordin-like protein 1. Although we did not find previously published evidence for an association between these 2 proteins and HF, these proteins are linked to other cardiovascular pathophysiology (eg, cardiac hypertrophy; Tables S9 and S10), suggesting the biological plausibility of our novel findings. While NT-proBNP was selected as a feature in both the HFrEF and HFpEF models, the proteomic approach was better able to predict HF mortality compared with NT-proBNP measurements alone, suggesting that they capture additional important features of the pathophysiology of HF with meaningful gains in C-index and AUC.One inconsistency between these findings and previous studies was that there was no overlap between proteins in our models compared with a recently developed proteomic score by Gui et al.12 The differences in these proteomic findings are most likely indicative of differing approaches used for model training. The current study utilized a proteome-only approach, while Gui et al12 forced adjustment for NT-proBNP and MAGGIC score.Finally, these models improve upon many of the limitations of individual marker measurements or clinical risk scores currently used for HF prognostication. For example, larger protein panels have an advantage over NT-proBNP alone, as the levels of the latter can misinform risk in individual patients, such as those with atrial fibrillation or obesity. In addition, the multi-protein scores change with risk over time as they are not reliant on static, immutable clinical factors. The MAGGIC clinical score contains predictors, many of which are not readily modifiable including age, time since diagnosis, chronic obstructive pulmonary disease, male sex, diabetes, β-blocker status, angiotensin-converting enzyme inhibitor or angiotensin-receptor status, lower systolic blood pressure, and body mass index.5We have addressed the many strengths of these models; however, there are also limitations that should be discussed. We did not study the effects of including clinical covariates in our models; however, we did assess whether combining our models with NT-proBNP and MAGGIC scores improved performance. The intent of this study was to evaluate the potential of the plasma proteome as the sole information source for developing prognostic HF models. This approach enabled machine learning algorithms to include proteins that likely capture the biology of key clinical and demographic factors.9 Furthermore, requiring multiple sources of data would make these models less convenient in routine clinical care, has the potential for increasing costs if further testing were required, and would not allow for scaling of multiple proteomic-only tests to be applied to a single blood sample. The data sets used to develop the PRS may not reflect the impact of newer therapies, and some observational cohorts are missing data about the adoption of guideline therapies. Finally, neither model has been assessed for changes in response to medication, which, if successful, could confirm that a desired treatment effect has occurred and allow for more personalized therapies.Overall, these data support that the plasma proteome could contain prognostic information and provide a powerful tool for understanding HF physiology. Our report provides a key step forward in precision medicine, demonstrating that proteomic models discovered and validated on holdout sets with an aptamer platform (SomaScan) can accurately predict HF mortality at 1 timepoint, as well as serially over time. Whether this approach has the potential to improve longitudinal monitoring of patients and facilitate personalized care remains to be proven.
ARTICLE INFORMATION
The authors are grateful to the participants in all the cohorts and to the study teams, physicians, and medical staff. The authors also thank the SomaLogic assay team, the platform bioinformatics team for quality control, the data management team for clinical data organization, and J. Williams for the agreements with the study institutions.The ARIC study (Atherosclerosis Risk in Communities) has been funded in whole or in part with federal funds from the National Heart, Lung, and Blood Institute (NIHR), National Institutes of Health (NIH), Department of Health and Human Services (contracts HHSN268201700001I, HHSN268201700002I, HHSN268201700003I, HHSN268201700004I, and HHSN268201700005I; grants R01HL087641 and R01HL086694), National Human Genome Research Institute contract U01HG004402, and NIH contract HHSN268200625226C. The Bio-SHiFT study (The Role of Biomarkers and Echocardiography in Prediction of Prognosis of Chronic Heart Failure Patients) was supported by the Jaap Schouten Foundation,Noordwest Academie, and European Union/European Federation of Pharmaceutical Industries and Associations Innovative Medicines Initiative 2 Joint Undertaking BigData@Heart grant 116074. Dr Ganz’s proteomic research is funded by NIH grants 1R01HL159081, R01 HL153499, 5U01DK108809, 5R01HL129856, R01HL148218, R01HL059367, 5R01AG052964, and R01 HL155209. Dr Asselbergs is supported by UCL Hospital's NIHR Biomedical Research Centre.SomaLogic Inc funded the costs of the proteomic assays, and Drs Chadwick, Loupy, Hinterberg, Sampson, and Ostroff and N. Kureshi are employees of the company. Dr Ganz is a member of SomaLogic’s Medical Advisory Board for which he accepts no financial remuneration of any kind. Dr Kardys received travel reimbursement from SomaLogic and Olink.Supplemental MethodsTables S1–S10Figure S1References 27–50
NS
Dr. Nancy Sweitzer
Cardiology · Advanced Heart Failure & TransplantationWashington University in St. Louisnancys
