Predictive Analytics and Clinical Decision Support for Acute Type A Aortic Dissection: A Machine Learning Approach to 30-Day Mortality Prediction
Presented During:
Thursday, April 25, 2024: 5:38PM - 7:00PM
Sheraton Times Square
Posted Room Name:
Central Park
Abstract No:
P0259
Submission Type:
Abstract Submission
Authors:
Bo Yang (1), Carol Ling (2), Alexander Shieh (3), Chi-Ching Huang (4)
Institutions:
(1) University of Michigan, Ann Arbor, MI, (2) University of Michigan, Department of Cardiac Surgery, Ann Arbor, MI, (3) UT MD Anderson Cancer Center, Houston, TX, (4) N/A, Taipei, Taiwan
Submitting Author:
*Bo Yang
-
Contact Me
University of Michigan
University of Michigan
Co-Author(s):
Carol Ling
-
Contact Me
University of Michigan, Department of Cardiac Surgery
University of Michigan, Department of Cardiac Surgery
Alexander Shieh
-
Contact Me
UT MD Anderson Cancer Center
UT MD Anderson Cancer Center
Chi-Ching Huang
-
Contact Me
N/A
N/A
Presenting Author:
Chi-Ching Huang
-
Contact Me
N/A
N/A
Abstract:
Objective:
To develop a machine learning algorithm for the precise prediction of 30-day mortality in patients presenting with acute type A aortic dissection (ATAAD).
Methods:
A retrospective analysis was conducted using data from the Michigan Medicine Aortic Dissection database, covering the period from January 1996 to February 2023. Data were retrieved from chart reviews, the Society of Thoracic Surgeons warehouse, the national death index, and the Michigan death index database.
A random forest (RF) machine learning model was used to predict 30-day mortality. Preprocessing involved addressing missing values through multiple imputation by chained equations. Continuous variables were normalized by linearly scaling each feature to a range of 0 and 1. The dataset was separated into an 80:20 ratio for the training set and held-out testing set. We performed 5-fold cross-validation and feature selection using the training set. A total of 42 features were utilized in constructing the original RF machine learning model.
Subsequently, the top 10 features, selected using the mean impurity decrease method from the original RF, were isolated. These features were then used to develop the new RF model, which was tested on the held-out testing set to mitigate overfitting. Evaluation metrics included the Area Under the Receiver Operating Characteristic curve (AUROC) and Brier Score.
Results:
Within the cohort of 1,067 patients with ATAAD, the original RF model, using 42 features, achieved a testing AUROC of 0.825 and a Brier score of 0.099. Key features, including DeBakey type 1 or 2, right upper extremity malperfusion, age, creatinine levels, glomerular filtration rate, inotrope usage, anticoagulant usage, innominate artery dissection presentation, left common carotid dissection presentation, and Body Mass Index, contributed to the RF model's predictive capacity. The new RF model, using the 10 selected features, achieved a testing AUROC of 0.841 and a Brier score of 0.101.
Conclusions:
Our study highlights the strong predictive capability of the RF model in predicting 30-day mortality among patients with ATAAD. Considering the high mortality rate in this patient group, integrating pre-operative mortality predictions could aid surgeons in their decision-making process; however, the prediction should not be the sole determinant for proceeding with surgery. Future external validation using larger cohorts ensures the generalization of our findings.
To develop a machine learning algorithm for the precise prediction of 30-day mortality in patients presenting with acute type A aortic dissection (ATAAD).
Methods:
A retrospective analysis was conducted using data from the Michigan Medicine Aortic Dissection database, covering the period from January 1996 to February 2023. Data were retrieved from chart reviews, the Society of Thoracic Surgeons warehouse, the national death index, and the Michigan death index database.
A random forest (RF) machine learning model was used to predict 30-day mortality. Preprocessing involved addressing missing values through multiple imputation by chained equations. Continuous variables were normalized by linearly scaling each feature to a range of 0 and 1. The dataset was separated into an 80:20 ratio for the training set and held-out testing set. We performed 5-fold cross-validation and feature selection using the training set. A total of 42 features were utilized in constructing the original RF machine learning model.
Subsequently, the top 10 features, selected using the mean impurity decrease method from the original RF, were isolated. These features were then used to develop the new RF model, which was tested on the held-out testing set to mitigate overfitting. Evaluation metrics included the Area Under the Receiver Operating Characteristic curve (AUROC) and Brier Score.
Results:
Within the cohort of 1,067 patients with ATAAD, the original RF model, using 42 features, achieved a testing AUROC of 0.825 and a Brier score of 0.099. Key features, including DeBakey type 1 or 2, right upper extremity malperfusion, age, creatinine levels, glomerular filtration rate, inotrope usage, anticoagulant usage, innominate artery dissection presentation, left common carotid dissection presentation, and Body Mass Index, contributed to the RF model's predictive capacity. The new RF model, using the 10 selected features, achieved a testing AUROC of 0.841 and a Brier score of 0.101.
Conclusions:
Our study highlights the strong predictive capability of the RF model in predicting 30-day mortality among patients with ATAAD. Considering the high mortality rate in this patient group, integrating pre-operative mortality predictions could aid surgeons in their decision-making process; however, the prediction should not be the sole determinant for proceeding with surgery. Future external validation using larger cohorts ensures the generalization of our findings.
Aortic Symposium:
Dissection
Image or Table
