Proteomic risk model superior to clinical risk model in predicting recurrent ASCVD events
Targeted proteomics improves cardiovascular risk prediction in secondary prevention
Introduction and methods
Background
In the last years, more and more novel therapeutic agents have become available that offer an opportunity to further reduce residual risk of ASCVD [1-9]. Traditional risk scores perform poorly in terms of discrimination [10]. More accurate methods are needed for the identification of patients who are at highest risk of recurrent events to guide risk-based therapeutic decisions.
Aim of the study
This study investigated the predictive value of a targeted plasma proteomics approach, analyzed with machine learning techniques in a secondary prevention setting.
Methods
The derivation cohort consisted of 870 individuals from the SMART cohort and the validation cohort consisted of 700 individuals from the Athero-Express cohort [11,12]. A total of 276 proteins were measured in blood samples from panels that were selected based on known associations with ASCVD.
Three models were constructed using machine learning techniques in the derivation cohort:
- A proteomic risk model comprising 50 proteins with the highest predictive value
- A clinical risk model comprising age, sex, BMI, SBP, total cholesterol, HDL-c, CRP, smoking status, diabetes, antihypertensive medication, and family history of CVD as parameters
- A combined model, comprising clinical risk parameters and protein parameters
All three models were recalibrated for use in the validation cohort and tested for discrimination, calibration and reclassification.
In an exploratory analysis, the predictive value of the proteomic model was assessed in patients with low and high CRP levels. For this analysis, individuals in the derivation cohort were categorized according to CRP levels: high CRP (>2 mg/L, n=463) and low CRP (≤2 mg/L, n=373). Those with CRP >20 mg/L were excluded.
Outcomes
The primary outcome was first recurrent ASCVD event, defined as acute MI, ischemic stroke or CV death.
Main results
Discrimination
- The proteomic risk model performed significantly better in terms of discriminatory value compared with the clinical risk model in the derivation cohort (ROC AUC 0.810 vs. 0.750; delta AUC 0.060, 95% CI 0.009-0.019, P<0.001), and in the validation cohort (ROC AUC 0.801 vs. 0.765; delta AUC 0.036, 95% CI 0.020-0.051, P<0.001).
- The combined model was superior to the proteomic risk model in the derivation cohort, but not in the validation cohort.
- The top 3 proteins with the strongest predictive value for recuring ASCVD events were NT-proBNP, kidney injury molecule-1 (KIM-1) and matrix metalloproteinase 7 (MMP-7).
Calibration and reclassification
- All models were well calibrated. However, risk was slightly underestimated in the highest-risk categories.
- The net reclassification improvement (NRI) and integrated discrimination index (IDI) were calculated by comparing the proteomic risk model with the clinical risk model. The NRI was 0.173 and IDI was 0.085 in the validation cohort.
High vs low CRP levels
- IL-6 levels were much higher in the group with high CRP levels, compared to the group with low CRP levels.
- In the group with low CRP levels, the top 10 proteins with highest predictive value comprised four proteins not represented in the initial model nor in the high CRP model. These proteins were α1-microglobulin-bikunin precursor (AMBP), nidogen-1 (NID1), tissue factor (TF), and vasorin (VASN). These proteins are related to neutrophil signaling. This points toward a residual inflammatory risk in individuals in the low CRP group, independent from the IL6-CRP pathway.
Conclusion
This study shows that a proteomic risk model comprising a panel of 50 proteins performed significantly better in terms of discrimination than a clinical risk model in predicting recurrent ASCVD events. The proteomics model was well calibrated and provided significant net reclassification improvement.
Share this page with your colleagues and friends: