# Salp swarm-optimized machine learning models for predicting preoperative aortic rupture risk in acute type a aortic dissection patients

## TL;DR

This study uses machine learning optimized by the Salp Swarm Algorithm to predict aortic rupture risk in patients with acute type A aortic dissection, improving clinical decision-making.

## Contribution

The novel use of the Salp Swarm Optimization Algorithm to enhance machine learning models for predicting aortic rupture risk in ATAAD patients.

## Key findings

- The SSA-optimized Random Forest model achieved 97.41% accuracy and 0.980 ROC-AUC in predicting aortic rupture.
- Key predictors of rupture risk included eGFR, hypotension at admission, and white blood cell count.
- SMOTE was used to address class imbalance in the dataset, improving model performance.

## Abstract

Acute Type A aortic dissection (ATAAD) is characterized by acute onset and rapid progression, with aortic rupture due to dissection extension being the primary lethal mechanism. Timely identification of high-risk patients is critical for prioritizing surgical intervention to reduce rupture incidence. This study aimed to develop and validate an interpretable machine learning model to predict aortic rupture in ATAAD patients, thereby improving risk classification and supporting clinical decisions. Medical records of ATAAD patients from Xiamen Cardiovascular Hospital (January 2019–October 2024) were retrospectively analyzed. Predictors were screened via statistical significance (p
<
0.05) using seven machine learning algorithms, with the Salp Swarm Optimization Algorithm (SSA) optimizing hyperparameters for Random Forest and XGBoost models. To address class imbalance (47 rupture cases, 6.1%), SMOTE was implemented for data augmentation. Model performance was evaluated by accuracy, F1-score, precision, ROC-AUC, sensitivity, and specificity, supplemented by interpretability analyses through feature importance ranking and SHAP. Among 774 included ATAAD patients, the SSA-optimized Random Forest model achieved optimal performance (test dataset: 97.41% accuracy, 0.980 ROC-AUC, 81.82% F1-score). Key predictors included estimated glomerular filtration rate (eGFR), hypotension at admission, and white blood cell count. This work provides a quantitative tool for emergency care prioritization, with SSA enhancing model precision for high-risk patient identification, though multicenter studies are needed to validate generalizability.