Harnessing the Power of Artificial Intelligence to Compare and Classify the Predictive Powers of Several Distinct Proteins

Abstract

Several proteins have demonstrated prognostic potential in patients with peripheral arterial disease (PAD). In this study, we sought to harness the power of artificial intelligence and machine learning techniques to compare the prognostic potential of several proteins with respect to adverse cardiovascular events in patients with PAD. The predictive potential of five different plasma proteins (NT- pro BNP, CK-MB, Troponin I, FABP4, and FABP3) were investigated with respect to major adverse limb events (MALE). First, different features were determined and ranked as decisive to the classification model or not. A subset of the top-ranked features was then chosen depending on the rating, after which a random forest classifier was used for the chosen features. A 10-fold cross validation was applied to evaluate model performance. Receiver operating characteristic (ROC) curves were constructed for all model coefficients to calculate area under the curve (AUC) for each curve. Model performance was evaluated based on their sensitivity and specificity. Data from 917 patients was used for this study, comprised of 569 patients with PAD and 348 without PAD. Mean age for the overall cohort was 68 years, of which 65% were male. The random forest model showed a higher predictive power for FABP3 and FABP4 in all models. Considering our primary outcome of MALE, plasma proteins FABP3, FABP4, and NT-pro BNP were noted to be the most important predictors of MALE – highlighting their potential to physicians as facilitators of risk-stratification. The sensitivity for the final random forest in classifying patients with MALE was 91%, which increased to 93% upon adjusting the threshold from 50% to 40%. The same final model had a precision of 84% and an AUC of 88%. Plasma protein FABP3 holds immense potential to serve as a prognostic biomarker for PAD as it relates to MALE.