Back to 2024 Display Posters
Comparison Of Machine Learning Model With Multivariate Logistic Regression In Prediction Of 30-day Major Amputation In Patients With Peripheral Arterial Disease Following Lower Extremity Bypass
Sasank Kalipatnapu, MBBS, MS, Douglas W. Jones, MD.
UMass Chan Medical School, Worcester, MA, USA.
Introduction: Prior studies have shown that machine learning models in vascular surgery have better predictive ability than traditional logistic regression techniques. We studied the performance of an interpretable machine learning model and logistic regression to predict 30-day major amputation in patients with peripheral arterial disease (PAD) undergoing lower extremity bypass.
Methods: We queried the American College of Surgeons National Surgical Quality Improvement Program database for 23,536 patients who underwent lower extremity open procedures for PAD from 2011 to 2020. Using data from 2011 to 2018 for training(18908 patients) and data from 2019-2020 for testing (4628 patients), random forest machine learning model and multivariate logistic regression models (using both forward stepwise and backward stepwise selection techniques) were developed to predict the occurrence of 30-day major amputation.
Results: A random forest machine learning model was successful in predicting major amputation, achieving an area under the receiver-operator curve (AU-ROC) of 0.81 using all the variables from the NSQIP dataset. In comparison, a logistic regression model scored 0.83. In their choice of variables, both models used mostly similar predictors – elective vs emergency surgery, nature of the procedure performed, pre-operative disease severity, ASA class. In the model with post-operative variables as well, major reintervention on the bypass, untreated loss of patency and discharge destination were also incorporated.
| | | |
| | AU-ROC Scores (testing set) |
| Algorithm | Number of predictors | Only Pre-operative Variables | All Variables |
| Random Forest | Top 10 | 0.69 | 0.80 |
| All variables | 0.72 | 0.81 |
| Logistic Regression – Forward Stepwise selection | Top 10 | 0.73 | 0.83 |
| All variables | 0.73 | 0.83 |
| Logistic Regression – Backward stepwise selection | All variables | 0.73 | 0.83 |
Conclusion: In a dataset of patients who underwent open lower extremity procedures, a logistic regression model performed as well as a machine learning model in predicting 30-day major amputation. While machine learning techniques are promising, they are less interpretable and may not offer predictive advantages over logistic regression models.
Back to 2024 Display Posters
