Prognostic proteomic models for low event rates: A case study with myocardial infarction
Background
Large-scale clinical proteomics provide increasing opportunities for patient risk stratification, especially with multi-marker models derived using machine learning techniques. Prognostic models can be developed as binary risk classifiers, or by using time-to-event data. Survival modeling is ubiquitous in statistical literature, but support for machine learning optimization is more limited in comparison to other regression techniques. We have developed and assessed a novel prognostic model development method combining two statistical techniques – survival analysis and subsampling – using existing machine learning tools in R. These methods were applied to a clinical dataset to identify a highly predictive proteomic model for myocardial infarction (MI) despite a low observed event rate.
Methods
Cox elastic net with subsampling tools were developed in R. Simulations were used to demonstrate the utility and accuracy of subsampling in a survival data context, with comparisons made to logistic regression, Cox elastic net (Coxnet), and SVM models. Following the validation of the approach via simulations, models were developed and assessed on the HUNT3 data set (n = 756), which had 61 (8.1%) MI events within four years of blood draw. Proteomic measurements were performed using SomaScan® v4.0
technology.
Results
Simulation results and analysis of HUNT3 data set show improved performance metrics using the subsampled survival method.
Conclusions
Survival analysis with subsampling is a novel combination of techniques that can be applied to proteomic data to improve biomarker discovery and predictive modeling in the context of relatively low incidence rates.Using these methods, sensitivity and specificity metrics were more balanced on real-world hold-out test sets, and simulation results showed improved discrimination metrics using subsampled survival analysis. Additionally, simulations showed that the proteins that were most highly correlated with MI were selected for final models, indicating that this method is a promising tool for clinical discovery and prognostic/diagnostic development.
Authors
Y. Hagar
L.E. Alexander
J. Chadwick
G. Datta
M.A. Hinterberg
SomaLogic Operating Co., Inc., Boulder, CO USA
More posters
PosterAptamer-based analysis of plasma proteome of growing tumors
With proteins, the presence of a tumor is more often accompanied with changes in the levels of endogenous, unmutated proteins in circulation. In this context, knowing which proteins represent the earliest markers or tumor presence would be enormously useful.
PosterProteomic Models to Predict Pre-Analytical Variation
In biomarker discovery, it is critical to assess any pre-analytical variation (PAV) in order to avoid artificial bias in the intended measurements. PAV may arise from both avoidable and unavoidable factors, resulting in misleading data and incorrect conclusions. Proteins, in particular, are vulnerable to variation in collection methods, storage temperatures, and processing protocols. It is vitally important to understand this PAV when analyzing samples using protein assays.
PosterEfficient development of certified diagnostic laboratory developed tests using proteomic data
We demonstrate the utility of combining pipeline tools, statistical learning techniques, and a knowledge base of in-silico proteomic datasets into a reproducible workflow that allows for efficient development of LDT-certifiable tests using SomaScan® technology.