Sex-specific machine learning classification models improve outcome prediction for abdominal aortic aneurysms.
Authors
Affiliations (16)
Affiliations (16)
- Department of Bioengineering, University of Pittsburgh, Pittsburgh, PA, USA.
- Division of Cardiovascular Surgery, Mayo Clinic Health Systems, Eau Claire, Wisconsin, USA.
- Division of Interventional Radiology, Mayo Clinic Health Systems, Eau Claire, Wisconsin, USA.
- Department of Surgery, University of Pittsburgh, Pittsburgh, PA, USA.
- Division of Vascular Surgery, University of Pittsburgh, Pittsburgh, PA, USA.
- School of Medicine, University of Pittsburgh Medical Center, Pittsburgh, PA, USA.
- McGowan Institute for Regenerative Medicine, University of Pittsburgh, Pittsburgh,, PA, USA.
- Department of Chemical and Petroleum Engineering, University of Pittsburgh, PA, USA.
- Department of Cardiothoracic Surgery, University of Pittsburgh, Pittsburgh, PA, USA.
- Clinical & Translational Sciences Institute, University of Pittsburgh, Pittsburgh, PA, USA.
- Department of Mechanical Engineering and Materials Science, University of Pittsburgh, Pittsburgh, PA, USA.
- Center for Vascular Remodeling and Regeneration, University of Pittsburgh, Pittsburgh, PA, USA.
- Magee Womens Research Institute, University of Pittsburgh, Pittsburgh, PA, USA.
- Department of Bioengineering, University of Pittsburgh, Pittsburgh, PA, USA. [email protected].
- Clinical & Translational Sciences Institute, University of Pittsburgh, Pittsburgh, PA, USA. [email protected].
- Department of Mechanical Engineering, University of Texas San Antonio, Texas, San Antonio, United States of America. [email protected].
Abstract
Abdominal aortic aneurysm (AAA) is an abnormal dilation of the abdominal aorta that carries up to a 90% mortality rate when ruptured. Although male patients experience AAA at a higher rate than females, female patients experience AAA rupture at a rate three- to four-fold higher that of their male counterparts. The current standard clinicians use for determining when to surgically intervene is maximum transverse diameter of the AAA perpendicular to the axis of flow. However, some aneurysms below these diameter thresholds rupture. Machine learning (ML) classification models have been previously shown to predict patient outcomes with more discriminability than the diameter criterion. However, these models do not consider sex-based differences. In this proof-of-concept study, we investigate how creating sex-specific ML models impacts patient outcome prediction as compared to a general model encompassing all patients (sex agnostic). Computed tomography image sets were acquired from 537 patients (n = 159 female, n = 378 male) at the University of Pittsburgh Medical Center (UPMC) and Mayo Clinic Health Systems. Features used as input to the ML models were categorized as clinical, biomechanical, and morphological data. Prior to ML model training, patient data were randomly split for 20% holdout testing. ML models encompassing all patients (general model), only male patients (male-specific model), and only female patients (female-specific model) were trained and tested. A female-specific model and male-specific model both had a higher maximum area under the receiver-operating characteristic curve values than a general model for female patients and male patients, respectively. Equalizing the sample sizes of female and male patients in the model led to improved outcomes for female patients without decreasing performance for male patients. ML classification models show promise in improving predictions of patient outcomes for AAA. The higher AAA prevalence rate for males leads to female patients being underrepresented in AAA datasets. In this proof-of-concept study, we demonstrated that sex-specific models outperformed a general model in predicting patient outcomes. Additionally, equalizing sample sizes within the dataset improved predictions for female patients without compromising overall performance of the model. As ML applications in medicine continue to grow, it is important to consider population representation within datasets to reduce model bias.