Biomarkers in Medicine Publication

Application of a Machine Learning-Driven Multibiomarker Panel for Prediction of Incident Cardiovascular Events in Patients with Suspected Myocardial Infarction

June, 2020

Highlights

  • In a moderate- to high-risk population undergoing diagnostic coronary angiography, a multiple biomarker score with high accuracy for predicting incident major adverse cardiac events (MACE) to at least one year of follow up was evaluated.

  • The final score is comprised of four biomarkers: Kidney Injury Molecule-1 (KIM-1), NT-proBNP, osteopontin and Tissue Inhibitor of Metalloproteinase-1 (TIMP-1).

  • Lower risk patients (score 0 - 4.5) had a NPV of 97% for predicting MACE at 3-365 days, while higher risk patients (score 6-10) had a PPV of 36% for MACE at 3-365 days.

  • The present risk score may be particularly useful for those with suspected or known stable coronary artery disease, for which there are very few prognostic risk models and also has a role in enriching clinical trials of therapies to prevent MACE.

  • In such a cohort, the score could both identify higher-risk patients who merit consideration for more aggressive medical therapy while avoiding such measures in lower-risk patients.

Study Overview

BACKGROUND

With a prevalence of cardiovascular disease as high as 30% in American adults, prevention of major adverse cardiac events (MACE) is a major public health objective. Central to this preventative process is risk assessment.  Clinical risk factors are commonly used to predict MACE, however, they often lack sufficient positive and negative predictive value.

 

GOAL

In an at-risk population, to develop and validate a robust, non-invasive, multi-biomarker based score for predicting incident MACE within one year of follow up.

 

HYPOTHESIS

A multiple biomarker approach would provide accurate, biologically-based predictive information regarding risk for MACE, additive to or superior to clinical variables and individual biomarkers alone in and at-risk population referred for coronary angiography.

 

METHODS

The CASABLANCA study was a prospective, observational cohort study of 1251 subjects undergoing coronary and peripheral angiography ± intervention between 2008 and 2011 at Massachusetts General Hospital.

 

The initial 927 patients undergoing coronary angiography (with or without peripheral angiography) between 2008 and 2011 were included in this study and were randomly split into a training set (n=649) and a holdout validation set (n=278)

 

MACHINE LEARNING & BIG DATA

Candidate panels of proteins and clinical features were generated via least-angle regression (LARs), a method similar to forward stepwise regression. In this method, factors are included in the model one at a time, with their coefficients determined by their correlation with the outcome.  This is repeated until all factors are included in the model, and the step at which the performance plateaued resulted in an initial panel of interest.

 

With this panel of interest, predictive analyses were run on the training set using least absolute shrinkage and selection operator (LASSO) with logistic regression, predicting the outcome of one-year MACE using only the variables in the panel of interest.  The ultimate result of this process was a final panel of four proteins.

 

Age- and score-adjusted Cox proportional hazards analyses were performed.  Hazard ratios (HR) for an elevated prognostic score as well as per-unit score increase with 95% CI were estimated.  Time-to-first MACE event as a function of elevated prognostic score was calculated, displayed as Kaplan-Meier survival curves, and compared using log-rank testing.  Lastly, to facilitate interpretation, raw output of the prognostic model was linearly transformed into a scaled score of 0-10.

Key Findings

  • During the model-building process, no clinical variables were retained and the final score is comprised of four biomarkers: KIM-1, NT-proBNP, osteopontin and TIMP-1.

 

  • Lower risk patients (score 0 - 4.5) had a NPV of 97% for predicting MACE at 3-365 days, while higher risk patients (score 6-10) had a PPV of 36% for MACE at 3-365 days.

Conclusions

  • The present risk score may be particularly useful for those with suspected or known stable coronary artery disease, for which there are very few prognostic risk models and also has a role in enriching clinical trials of therapies to prevent MACE.
     

  • In such a cohort, the score could both identify higher-risk patients who merit consideration for more aggressive medical therapy while avoiding such measures in lower-risk patients.
     

  • In a moderate- to high-risk population undergoing diagnostic coronary angiography, a multiple biomarker score has been developed with high accuracy for predicting incident MACE to at least one year of follow up.

Scroll Down for ACC Publication

Highlights

  • Predicting future cardiovascular events (CVE) in patients presenting with suspected myocardial infarction (MI) is an unmet clinical need. In a population of patients with suspected MI presenting to the emergency department, we sought to externally validate a multiple biomarker panel for prediction of incident CVE.

  • 748 patients presenting with suspected MI to the emergency department were prospectively recruited. A multiplex assay including 4 biomarkers (NT-proBNP, KIM-1, osteopontin, TIMP-1) was measured in blood samples. A predictive score ranging from 0 to 10 had been previously derived in 649 stable and unstable patients with suspected coronary artery disease (CAD) or MI in the CASABLANCA trial.

  • All patients were followed for one year to assess incident CVE (composite of CV death and incident MI).

  • Application of the prognostic score resulted in an area under receiver operating characteristic curve (AUC) of 0.86 (p<0.001). Application of the optimal cutoff resulted in a high negative predictive value of 99.4% (95%CI 98.7%,100%) and a sensitivity of 86.4% (95%CI 72.0%,100%) for incident CVE. Patients with a prognostic score above the cutoff were more likely to develop a CVE, compared to patients below the cutoff.

 

  • In this external validation study of patients with suspected MI, the Prevencio HART CVE panel provided high accuracy for predicting incident CVE within one year of follow up.

Study Overview

BACKGROUND

With a prevalence of cardiovascular disease as high as 30% in American adults, prevention of major adverse cardiac events (MACE) is a major public health objective. There is a clinical need for alternative strategies to tailor preventive medical care to improve outcomes in high-risk patients.  

GOAL

The aim was to apply a previously developed multi-biomarker panel consisting of four proteins with novel risk-prediction in a cohort of patients presenting with symptoms suggestive of MI to the emergency department.  Additionally, a comparison of the predictive value to high-sensitivity cardiac troponin I (hs-cTnI) for predicting incident MACE within one year was the intent.

HYPOTHESIS

External validation of a multiple biomarker approach would provide accurate, biologically-based predictive information regarding risk for MACE, previously developed to be superior to clinical variables and individual biomarkers alone in a population presenting to the ED with symptoms suggestive of a myocardial infarction. 

METHODS

The CASABLANCA study was a prospective, observational cohort study of 1251 subjects undergoing coronary and peripheral angiography ± intervention between 2008 and 2011 at Massachusetts General Hospital and provided the internal validation for the development of the multiple biomarker panel and algorithm.  The samples were randomly divided into a training set (n=649) and a holdout validation set (n=278).
 

​The external validation was conducted on the initial 750 consecutive patients with available blood samples from the Biomarkers in Acute Cardiac Care (BACC) study (NCT02355457).  These patients presented to the ED with symptoms suggestive of myocardial infarction. All patients were contacted by direct telephone calls and a standardized questionnaire was used to assess cardiovascular mortality and incident MI within 1 year.

MACHINE LEARNING & BIG DATA

The same biomarker protein panel, scoring system and algorithm were used as the internal validation model predicting the outcome of one-year MACE.  The model is comprised of four biomarkers:  KIM-1, NT-proBNP, osteopontin and TIMP-1.

Time-to-first MACE event as a function of elevated prognostic score was calculated, displayed as Kaplan-Meier survival curve, and compared using log-rank testing. Lastly, to facilitate interpretation, raw output of the prognostic model was linearly transformed into a scaled score of 0-10.

Key Findings

  • In the internal validation set, using the same multi-protein panel and algorithm, HART CVE had an AUC of 0.79 with 97% PPV for low score patients.  In this external validation set, the AUC was 0.86 with a NPV of 99.4% for incident MACE.

  • Lower risk patients (score 0 - 5.5) had a NPV of 99.4% and sensitivity of 86.4% for predicting MACE at 3-365 days, while higher risk patients (score >5.5) had a PPV of 7.9% with a specificity of 69.7% for MACE at 3-365 days. Positive likelihood ratio was 2.85.

  • In comparison, hs-cTnI <6 ng/l had a sensitivity of 84.2% with a NPV of 99.1%.  Those with hs-cTnI ≥6 ng/l had a specificity of 47.3% and a PPV 4.1% for MACE at 3-365 days. Positive likelihood ratio was 1.60.

Conclusions

  • The present risk score suggest that this multi-protein panel might be useful to better prognosticate short-to-intermediate-term outcomes in patients presenting with chest discomfort and possible MI.  

  • Application of this biomarker panel in the ED could improve decision-making and help to identify high-risk individuals and trigger further diagnostic and therapeutic workup to improve outcome.

  • In a moderate- to high-risk population undergoing diagnostic coronary angiography, a multiple biomarker panel has been developed with high accuracy for predicting incident MACE to at least one year of follow up. The multiple biomarker panel externally validated with high accuracy in patients presenting with chest discomfort and possible MI. 

Scroll Down for AJC Publication

The AJC Publication

The ACC Publication