Machine learning-based all-cause mortality risk assessment integrating novel biomarkers in patients with suspected or known coronary artery disease: the AI-ANOX-EXCEED-J Study

Background We have previously demonstrated that serum levels of vascular endothelial growth factor (VEGF)-C are inversely, and those of VEGF-D are positively associated with all-cause mortality in patients with suspected or known coronary artery disease (CAD). However, whether a machine learning (ML) prediction model incorporating VEGF-C and VEGF-D outperforms a conventional model using clinical confounders and established cardiovascular (CV) biomarkers remains unknown. Methods Data on 58 clinical parameters, 3 established CV biomarkers (N-terminal pro-brain natriuretic peptide, high-sensitivity cardiac troponin I, and high-sensitivity C-reactive protein) and 4 potential novel biomarkers (VEGF, VEGF-C, VEGF-D, growth differentiation factor 15 [GDF-15]) from two multicenter prospective cohort studies involving 5,673 patients with suspected or known CAD were analyzed. The primary outcome was all-cause death over a 6-year follow-up period. Four ML algorithms were employed: random forest (RF), linear support vector machine (SVM), neural network (NN), and naïve Bayes (NB). The dataset was randomly split into derivation (70%) and validation (30%) cohorts. Model performance was evaluated using the concordance index (C-index) with 95% confidence intervals (CI). Results During follow-up, 1,054 deaths occurred. All ML models outperformed the conventional Cox proportional hazard model (C-index [95% CI]: RF 0.836 [0.832–0.840], SVM 0.833 [0.829–0.838], NN 0.816 [0.813–0.820], and NB 0.810 [0.806–0.814] vs. conventional model 0.778 [0.754–0.802]). The top 10 important variables in each ML model are shown in Table 1. Notably, VEGF-C was ranked among the top 10 most important variables in all ML models, while GDF-15 and VEGF-D were in the top 10 for three and one model(s), respectively. When re-evaluated using 20 selected variables listed in Table 1 (excluding hematocrit due to its redundancy with hemoglobin), the ML models maintained performance comparable to those using all 65 variables (C-index [95% CI]: RF 0.839 [0.835–0.844], SVM 0.836 [0.830–0.842], NN 0.819 [0.814–0.825], and NB 0.820 [0.815–0.825]). Conclusions Machine learning models incorporating VEGF-C, VEGF-D, and GDF-15 provide superior prediction of all-cause mortality compared to conventional models in patients with suspected or known CAD.