Machine learning-based model stacking and multi-key biomarker association for rapid differentiation of patients with acute chest pain：a multicenter study with subgroup bias evaluation (Preprint)

Abstract

Despite significant advancements in cardiovascular disease research, the rapid diagnosis of acute high-risk chest pain diseases such as acute myocardial infarction (AMI), pulmonary embolism (PE), and aortic dissection (AD) remains a challenge in the emergency setting. We developed a multicenter risk prediction model (Stacking-Optuna) that rapidly and accurately distinguishes between AMI, PE, and AD. This model underscores the importance of biomarkers such as cardiac troponin, brain natriuretic peptide, and D-dimer while addressing the limitations of current diagnostic methods, especially in terms of considering patient age, gender, and the combined use of different indicators. This model integrates three large-scale databases for training and validation. The results demonstrate that Stacking-Optuna exhibits exceptional discriminative ability for all three acute chest pain diseases (AMI group: AUC=0.9380 [95%CI 0.9160-0.9480], PE group: AUC=0.9480 [95%CI 0.9220-0.9560], AD group: AUC=0.9540 [95%CI 0.9260-0.9640]). Additionally, the interpretability analysis and bias evaluation further reveal the model's consistency and generalizability in a clinical context (the median AUC of bias evaluations covering fifteen subgroups were above 0.83). This clinical knowledge-based fusion and data-driven approach supports rapid and accurate risk assessment of patients with emergency chest pain in the ICU, providing a more effective diagnostic tool for patients with cardiovascular emergencies.