Article Text
Abstract
Introduction Non-home discharge (NHD) is an important metric of quality care and patient outcomes, as well as an important reimbursement criterion. We assess the feasibility of machine learning to preoperatively predict NHD after unruptured intracranial aneurysm (UIA) treatment. Identifying patients at risk for NHD after UIA treatment could help guide providers that plan on counseling patients and help plan discharges for NHD patients.
Methods From a prospectively maintained database, all patients (n=547) treated for UIA between 2017 and 2022 was retrospectively reviewed. 21 baseline characteristics were collected, including age, sex, and underlying medical conditions. 7 UIA and treatment characteristics were also collected, including aneurysm morphology, location, modality of treatment (open surgery vs endovascular), and endovascular access route (radial, femoral). The data was randomly divided into training and testing sets with an 80:20 ratio. Given the unbalanced classes, Synthetic Minority Over-sampling TEchnique (SMOTE) was applied to the training set. Logistic regression and five machine learning algorithms were trained: Random Forest, Extremely Randomized Trees, Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and k-nearest neighbors (KNN).
Results 520 (95%) of eligible patients had no missing data and were selected for analysis. The rate of NHD was 3.8% (n=20). Random Forest was the best discriminant of NHD and had the highest mean AUROC of 0.86 (s.d. ±0.03) and accuracy of 0.93 (±0.01). Random Forest narrowly but significantly (Mann-Whitney U-test; p = 0.002) outperformed logistic regression which had AUROC = 0.84 (±0.03) and accuracy = 0.94 (±0.01). XGBoost (AUROC: 0.80±0.04, accuracy: 0.91±0.02) and Extremely Randomized Trees (AUROC: 0.80± 0.03; accuracy: 0.93±0.01) were sufficient as well. SVM (AUROC: 0.55±0.05); accuracy: 0.91±0.01) and KNN (AUROC: 0.49±0.05; Accuracy: 0.74±0.03) performed no better than random chance, however.
Discussion We demonstrate that well-designed models trained on a small dataset can successfully predict non-home discharges in unruptured intracranial aneurysm treatment. Tree-based models (Random Forest, Extremely Randomized Trees, and XGBoost) outperformed SVM and KNN, indicating that future studies may want to consider tree-based models for similar tasks. We prove that our models, and more generally, machine learning, can be used to provide precise and personalized neurosurgical care.
Disclosures S. Patel: None. K. El Naamani: None. A. Hunt: None. P. Jain: None. C. Lawall: None. C. Yudkoff: None. O. El Fadel: None. M. Ghanem: None. P. Mastorakos: None. A. Momin: None. A. Alhussein: None. R. Alhussein: None. E. Atallah: None. R. Abbas: None. R. Zakar: None. S. Tjoumakaris: 2; C; MicroVention. M. Gooch: 2; C; Stryker. N. Herial: None. H. Zarzour: None. R. Schmidt: None. R. Rosenwasser: None. P. Jabbour: 2; C; Medtronic, MicroVention, Balt, Cerus Endovascular.