Interpretable Machine Learning Model Forecasts Postoperative Mortality in Acute Aortic Dissection

Researchers have developed an interpretable machine learning model and interactive website to forecast postoperative in-hospital mortality in patients with acute aortic dissection using routinely collected clinical data. The investigation, published in BMC Medical Informatics and Decision Making, combined extreme gradient boosting with Shapley additive explanations to create a prognostic tool designed to support risk stratification and clinical patient management.

Developing an accurate and practical prognostic model for acute aortic dissection has historically presented challenges when relying on routine demographic and admission records. To address this clinical need, the study authors designed a predictive workflow that utilizes readily available admission metrics to estimate mortality risk following surgical intervention.

The retrospective analysis evaluated admission records from 782 patients diagnosed with acute aortic dissection at the Fourth Hospital of Hebei Medical University. For model training and initial testing, the researchers partitioned this primary study population into a training cohort of 470 patients and an internal test cohort of 312 patients.

To assess the generalizability of the predictive system across different institutional settings, an independent external validation cohort of 318 patients with acute aortic dissection was enrolled from the Second Hospital of Hebei Medical University. Testing the model on this independent group allowed the investigators to evaluate how well the algorithms performed on unfamiliar patient records.

To identify the most informative risk factors associated with postoperative in-hospital mortality, the authors applied least absolute shrinkage and selection operator regression alongside multivariable logistic regression. These selection procedures identified nine critical clinical features: acute kidney injury, α-hydroxybutyrate dehydrogenase, Stanford A classification, carbon dioxide, blood chloride, monocyte count, hematocrit, upper gastrointestinal bleeding, and heart failure.

The extreme gradient boosting algorithm was integrated with Shapley additive explanations to ensure that individual risk forecasts remained interpretable to clinicians. Model performance was evaluated across the training, testing, and external validation datasets using receiver operating characteristic curve analysis, calibration assessments, discrimination metrics, and clinical utility evaluations.

Across the training set, internal test set, and independent external validation set, the machine learning model demonstrated sensitivities of 97.9 percent, 96.8 percent, and 95.3 percent, alongside specificities of 71.3 percent, 67.8 percent, and 70.9 percent. Overall accuracy values reached 84.5 percent, 82.4 percent, and 83.3 percent, while area under the receiver operating characteristic curve values were 0.845, 0.737, and 0.788, with corresponding F1 scores of 0.911, 0.899, and 0.903.

The investigators concluded by deploying the predictive architecture into an accessible website designed to calculate individualized risk estimates for postoperative in-hospital mortality. The authors noted that this interpretable platform provides potential utility for surgical risk stratification, patient management, and clinical decision-making in hospital care settings.

Subject of Research: Machine learning prediction of postoperative mortality in acute aortic dissection

Article Title: XGBoost–SHAP-based interpretable predictive website for postoperative in-hospital mortality in patients with acute aortic dissection

Article References: Jin, Y.-C., Zhang, X., Fang, M., Wu, Y.-Y., Wang, K.-W., Zhao, X.-S., Liu, X.-K., Dang, J.-Y., Ren, H.-Q., Geng, J.-Z., Zhao, Y.-Q., Xie, Y.-X., Zhao, Y.-P., Zhao, Y.-B., & Ma, D. (2026). XGBoost–SHAP-based interpretable predictive website for postoperative in-hospital mortality in patients with acute aortic dissection. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03889-0

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03889-0

Keywords: acute aortic dissection, postoperative mortality, XGBoost, SHAP, machine learning, risk stratification, clinical decision support, prognostic modeling, SHAP-based, interpretable, predictive, website