EP68 Exploring Predictive Relationships Between Preoperative Patient Hip-Related Symptoms with Radiographic and Demographic Measures via Machine Learning

Abstract Purpose Hip surgeons seek to understand the complexity and interconnectedness of relationships between preoperative patient-reported symptoms, demographics, and radiographic features. Machine learning (ML) approaches might help to elucidate these potentially non-linear relationships. This study used ML regression modeling to identify radiographic and demographic features that most influence preoperative symptom severity in patients with hip-related symptoms. Methods A retrospective chart review was performed to identify patients with hip-related symptoms of femoroacetabular impingement and hip instability at a single center. Demographics and radiographic measures (e.g., femoral and acetabular measurements) from 3D CT scans were used as ML model inputs. Separate ML predictive model outcomes were preoperative iHOT-12 (n = 618 hips) and PROMIS Pain Interference (PI) scores (n = 566 hips). The random forest ML model framework was selected for its strength in modeling complex, non-linear relationships. To assess generalizability, models were developed using a train-test split (i.e., 70/30), cross-validation, hyperparameter tuning, and repeated random sampling. Model performance was evaluated using R-squared. SHapley Additive exPlanations (SHAP) analysis was used to identify the most influential model inputs (radiographic and demographic variables). Results ML models consistently overfit the training data and generalized poorly when predicting iHOT-12 and PROMIS PI scores. Although overfitting occurred, SHAP analysis identified consistent demographic and radiographic features across models that were important for symptom severity prediction: acetabular surface area, femoral torsion, alpha angle at 12 o’clock, and weight for iHOT-12 scores and acetabular surface area, femoral neck shaft angle, acetabular version at 3 o’clock, and combined version for PROMIS PI. Conclusion ML models had poor predictive ability to generalize but consistently identified radiographic variables (e.g., acetabular surface area) associated with preoperative symptom severity. Poor model generalizability suggests this study’s specific clinical variables may not adequately represent the complexity of patient symptom variability. Ongoing work aims to clarify whether these relationships exist between demographic and radiographic features and patient-reported preoperative symptoms.