RF10.5 Machine Learning Methods Identify Subgroups of Patients with Hip Pain from Radiographic and Demographic Data

Abstract

Abstract Purpose Hip pain is caused by concomitant, overlapping pathologies, challenging diagnosis and treatment decisions. Machine learning (ML) offers an unbiased method of integrating patient-specific factors and identifying groupings to help surgeons better diagnose hip conditions. This study aimed to identify symptomatic patient groupings based on preoperative imaging and demographic data, and to reveal key clinical variables driving grouping using a two-stage ML approach of clustering and classification. Methods A single-center retrospective cohort of 618 hips with pain associated with femoracetabular impingement (FAI) and instability that underwent 3D CT analysis before treatment was identified. Patient-specific 3D-CT derived femoroacetabular measurements, demographics (sex, age, height, weight, and BMI), and preoperative iHOT-12 scores were collected from medical records. 3D-CT derived measurements and demographics were used for ML model development. Separate ML models of clustering (K-Means) and classification (logistic regression) were developed to group patients and identify key clinical factors separating them. Standard ML practice was followed to define clusters (UMAP data reduction, elbow plot, and silhouette score) and evaluate the accuracy of the classification model (AUC). SHapley Additive exPlanations (SHAP) analysis was applied to the classification model to identify key clinical features separating groups for clinical interpretation. Symptom severity was measured with baseline iHOT-12 scores was compared between groups post-hoc using standard parametric statistics. Results ML models identified 5 patient groupings separated primarily by sex, followed by patterns in femoroacetabular morphology. These lead to the following group-specific clinical interpretations: (1) males with mixed-type FAI; (2) females with microinstability driven by acetabular anteversion; (3) females with instability and cam-type FAI; (4) females with instability driven by acetabular undercoverage and small cam lesions; and (5) females with pincer-type FAI. The mean iHOT-12 scores in group 1 were significantly greater than those in groups 2-5 (p<0.05), with no difference between groups 2-5 (p>0.05). Conclusion ML models identified patient groups primarily driven by sex, with subgroup separation based on radiographic presentation of FAI and/or instability. Significant differences in symptom severity were observed in (1) males with mixed-type FAI compared to all other groups. Further investigation into these patterns may guide diagnosis, treatment planning, and prognosis.

Ähnliche Arbeiten