# Research on a machine learning-based predictive model for postoperative neurological dysfunction in acute Stanford type A aortic dissection

## TL;DR

This study developed a machine learning model to predict neurological dysfunction after surgery for aortic dissection, showing high accuracy in identifying at-risk patients.

## Contribution

A novel XGBoost-based predictive model for postoperative neurological dysfunction in acute aortic dissection patients, outperforming traditional methods.

## Key findings

- XGBoost achieved an AUC of 0.966 in internal validation and 0.951 in external validation.
- The model outperformed logistic regression and other ML models like SVC-LK, Nu-SVC, and AdaBoost.
- SHAP identified 15 robust features from 49 initial variables for prediction.

## Abstract

This study aimed to construct and validate a machine learning (ML) model integrating preoperative, intraoperative, and postoperative multimodal clinical data for individualized prediction of postoperative neurological dysfunction (ND) in patients with acute Stanford type A aortic dissection (ATAAD).

A retrospective analysis was conducted on 1,228 ATAAD patients (Aortic Disease Center of Beijing Anzhen Hospital, January 2020–December 2023): 853 patients (January 2020–December 2022) for model training/internal validation (via 10-fold cross-validation) and 375 patients (January–December 2023) for external validation. The 853 patients were grouped into control (n = 616) and ND (n = 237), including 203 transient ND (TND) and 34 permanent ND (PND) groups. Data were analyzed using Mann–Whitney U, chi-square (χ2), and Fisher’s exact tests (p < 0.05). Four ML models (SVC-LK, Nu-SVC, AdaBoost, XGBoost) were built with perioperative data; SHapley Additive exPlanations (SHAP) selected 15 robust features from 49 initial ones. Model performance was assessed via ROC-AUC (10-fold cross-validation for training/internal validation, external validation for effectiveness), and the optimal model was identified using DeLong test (two-tailed p-values). A multidimensional analysis compared the optimal model with traditional logistic regression (LR).

The XGBoost model exhibited the best performance: AUC = 0.966 (internal validation) and AUC = 0.951 (external validation), outperforming LR and the other three ML models.

The XGBoost algorithm demonstrates superior efficacy in predicting postoperative ND in acute ATAAD patients, providing postoperative early warning, identifying high-risk patients, offering clinical guidance, and enabling timely intervention.