Southern Association for Vascular Surgery
SAVS Home SAVS Home Past & Future Meetings Past & Future Meetings

Back to 2024 Abstracts


A novel risk calculator for nonhome discharge after lower extremity bypass
Brian Fry1, Jeremy Albright2, Shukri Dualeh1, Nicolas J Mouawad3, Andrew Kimball4, Jordan Knepper5, Eugene W Laveroni6, Chandu Vemuri1, Peter K Henke1
1University of Michigan, Ann Arbor, MI;2BMC2, Ann Arbor, MI;3McLaren Healthcare, Bay City, MI;4Spectrum Health, Grand Rapids, MI;5Michigan State University, East Lansing, MI;6Corewell Health, Farmington Hills, MI

INTRODUCTION: Post-discharge care is an underexamined yet critically important quality improvement target. Nonhome discharge (NHD [e.g., to a skilled nursing or rehabilitation facility]) is associated with an increase in discharge delays, post-discharge adverse events, and decreased patient quality of life. An accurate identification of patients at highest risk for NHD would greatly inform patient/surgeon shared decision making and identify opportunities for preoperative optimization. Moreover, it would prompt earlier discussion of the need for NHD, jumpstart facility identification, and guide optimal staff resource allocation for efficient discharge planning.
METHODS: All available adult patients in a statewide vascular surgery registry undergoing lower extremity bypass from January 1, 2016 - March 31, 2023 were analyzed. Our primary outcome was NHD (rehabilitation, acute care facility, nursing home/extended care, hospice, assisted living, left against medical advice). Machine learning (XGBoost) modeling was used to assess predictive accuracy based on factors associated with NHD including patient age, gender, race, comorbidities, smoking and ambulatory status, tissue loss, procedure class (elective, urgent, emergent), and length of stay. The model was developed using a training set and validated on a separate testing set to assess predictive accuracy. We then performed traditional logistic regression modeling and compared the modelsí area under the receiver operating characteristic (AUC) curves and calibration plots.
RESULTS: A total of 10,294 patients were included in the sample. At baseline, patients who had a NHD were more likely to be older, female, black, non-smokers, have tissue loss, require a vein graft, be urgent or emergent in presentation, have a longer median length of stay, have an ambulatory deficit, and more likely to have comorbid conditions. The top 5 most important factors predictive of NHD were longer length of stay, older age, presence of an ambulatory deficit, absence of claudication, and requiring an emergent procedure (Figure 1). While both models had excellent performance, machine learning outperformed logistic regression (AUC 0.859 [95% CI 0.845-0.873] vs. 0.863 [95% CI 0.849-0.876], p=0.016). Calibration plots also showed closer alignment between predicted probabilities and true event rates for the XGBoost model when compared to logistic regression (Figure 2).
CONCLUSIONS: Multiple factors contributed to a patientís risk of NHD in lower extremity bypass patients. Machine learning modeling can accurately predict NHD for greater than 86% of patients. These results will be incorporated into a novel preoperative risk calculator to inform surgeon/patient shared decision making, identify patients who would benefit from preoperative optimization, and to improve discharge planning efficiency.


Back to 2024 Abstracts