Back to 2025 Display Posters
How Do We Get Them Home: Identification Of Multi-disciplinary Factors Predicting Non-home Discharge After Open Aortic Repair Using Machine Learning
Joshua J. Myszewski, MD, Richard D. Gutierrez, MD, Colleen P. Flanagan, MD, Zachary A. Matthay, MD, Jade S. Hiramoto, MD, Warren J. Gasper, MD, Michael S. Conte, MD, James C. Iannuzzi, MD, MPH.
University of California San Francisco, San Francisco, CA, USA.
OBJECTIVE: This study evaluates the role of clinical factors, initial physical and occupational therapy recommendations in a machine learning algorithm to identify patients likely to have a non-home discharge.
METHODS: This was a retrospective analysis of open abdominal aortic aneurysm repair (OAR) at a single quaternary center for years 2012-2020. The primary endpoint was non-home discharge following OAR. Preoperative, intraoperative, and post-operative day 1 characteristics were evaluated including initial physical and occupational therapy evaluations. A classification tree algorithm was trained with feed-forward feature selection. Model performance was evaluated by classification accuracy and weighted F1 score on a randomly selected test set representing 20% of the study sample.
RESULTS: Overall 197 OAR cases were included, with 62 undergoing non-home discharge (31%). The mean age was 70.9±10.2 years. Case urgency was 46.7% elective and 53.3% emergent. Aneurysm anatomy was 50.2% infrarenal, 27.4% juxtarenal, 13.2% pararenal, and 9.2% suprarenal.The best performing algorithm and feature set included a total of 36 features for a classification accuracy of 92.5% (F1: 0.92) using only variables from post-operative day 1 or earlier. When limited to 10 total features the classification accuracy was 85.0% (F1: 0.84). The ten most predictive features are shown below in Table 1 in order of impact on the classification accuracy.
| | |
| Table 1: Top 10 most predictive factors for non-home discharge via machine learning |
| Rank | Variable | Relative Importance |
| 1. | Hyperkalemia | 0.39 |
| 2. | Impaired Mobility on Physical Therapy Evaluation | 0.15 |
| 3. | Tylenol use Post-Operative Day 1 | 0.13 |
| 4. | Non-English Primary language | 0.09 |
| 5. | Arousal/Alertness on Physical Therapy Evaluation | 0.09 |
| 6. | Dizziness on Physical Therapy Evaluation | 0.06 |
| 7. | History of PCI | 0.05 |
| 8. | History of Generalized Anxiety | 0.02 |
| 9. | Ketamine use on Post-Operative Day 1 | 0.02 |
| 10. | History of Coronary Artery Disease | 0 |
CONCLUSIONS: This study identified novel risk factors for non-home discharge after OAR. Machine learning can be used to help identify patients in need of increased resources early in the post operative stay. Future studies on the use of task-specific models for clinical decision support in vascular surgery should explore a wide range of multidisciplinary factors and continue to explore the use of machine learning as a tool to support pre-operative risk stratification for a variety of outcomes.
Back to 2025 Display Posters
