1254 Artificial Intelligence Accurately Predicts Transfusion Requirements in Posterior Spinal Fusion for Adult Spinal Deformity Based on Preoperative Patient Characteristics

INTRODUCTION: Surgical management of adult spinal deformity (ASD) is a complex procedure in spinal surgery, requiring detailed planning and attentive post-operative care. The use of machine learning tools, including artificial neural networks (ANNs), holds promise for enhancing clinical decision-making in this area. METHODS: Our institutional complex spine database was screened. Inclusion criteria were patients undergoing thoracolumbosacral PSF, six or more levels of vertebrae being fused, and ASD. Patients with a cervical component, less than six levels fused, and non-ASD etiology, such as tumor and trauma, were excluded. After data pre-processing 598 patients were included in our analysis. Post feature engineering, 9 pre-operative variables, including age, gender, levels PSF, pre-operative Hemoglobin, 3-column osteotomy use, posterior column osteotomy use, BMI, sacral involvement, and ASA score were incorporated. The data was split into a training (70%), testing (15%), and validation set (15%). Multivariable logistic regression, feedforward ANN, ensemble boosted ANN, extreme gradient boosting (XGBoost), random forest, explainable boosting machine (EBM), and support vector machine (SVM) models were utilized and evaluated on accuracy, sensitivity, AUC-ROC, F1-score, and precision. Bootstrapping was performed to calculate 95% CI. RESULTS: The ensemble boosted ANN was the most balanced with an accuracy of 0.82 (CI 95% 0.72-0.9), sensitivity of 0.97 (CI 95% 0.93-1), AUC-ROC of 0.86 (CI 95% 0.76-0.94), F1-Score of 0.90, (0.83-0.94) and precision of 0.83 (0.73-0.9). Comparative models yielded satisfactory outcomes but were demonstrably inferior. CONCLUSIONS: The findings of our research suggest that machine learning techniques, especially the ensemble boosted ANN, deliver strong and consistent predictive capabilities for ASD management. Such algorithms offer significant support to surgeons in tailoring treatment strategies and forecasting patient outcomes.