Machine learning-derived serum carbon dioxide trajectories in predicting prognosis of patients with acute aortic...

JOURNAL/mgres/04.03/01612956-990000000-00114/figure1/v/2026-08-19T154102Z/r/image-tiff Venous partial pressure of carbon dioxide (PvCO2) indicates postoperative adverse outcomes in patients with acute aortic dissection, and its trajectories over time affecting postoperative adverse outcomes remain unclear. This study aimed to investigate the relationship between PvCO2 trajectories and postoperative adverse outcomes, as well as clinical application through group-based trajectory modeling and machine-learning algorithms. This retrospective cohort study analyzed data from 642 patients with acute aortic dissection treated at the Fourth Hospital of Hebei Medical University from 2019 to 2024, to determine postoperative adverse outcomes, including 6-month mortality and complications. Three PvCO2 tests within the first 72 hours of admission were used to analyze PvCO2 trajectories through group-based trajectory modeling. To diagnose high-risk PvCO2 trajectories in emergency settings, least absolute shrinkage and selection operator regression was conducted to select key features, which were consequently integrated into predictive models using multiple machine-learning methods. Finally, a web-based interpretable website was developed for predicting PvCO2 trajectories in patients with acute aortic dissection. Overall, 642 patients were classed into four PvCO2 trajectories (T1, persistent decrease from a medium to a low level; T2, stable low level; T3, persistent decrease from a high to a low level; T4, stable midlevel), showing distinct patterns. After adjustment for other covariates, postoperative adverse outcomes among these groups showed significant differences, with a higher proportion of postoperative adverse outcomes in the T3 group. Subsequently, upon Shapley additive explanations analysis, 17 features with importance sequence were identified (hemoglobin, direct bilirubin, monocyte, adenosine deaminase, drinking, anion gap, chloride, platelet, lymphocyte, albumin/globulin ratio, calcium, alanine aminotransferase, α-L-fucosidase, D-dimer, antihypertensive drugs, and smoking), which were integrated into the model for predicting T3. The optimal machine-learning model was established for predicting postoperative adverse outcomes. Finally, a web-based interpretable model was developed to predict patients with acute aortic dissection and high risk of postoperative adverse outcomes. This study uncovers heterogeneity in PvCO2 trajectory, highlighting that decreased PvCO2 is associated with worse prognosis. A web-based tool for the prediction of acute aortic dissection patients at high risk of postoperative adverse outcomes was constructed to facilitate clinical application.