Resident Baylor College of Medicine Neurosurgery Houston, TX, US
Introduction: Understanding risk for metastatic vertebral compression fractures (VCF) is critical for all physicians and specialties involved in cancer patient care. With risk stratification, both prophylactic medical and minimally-invasive interventions can be employed to reduce future fracture/progression of fracture. While imaging-based VCF prediction models are gaining popularity, no studies have evaluated purely clinical, machine-learning (ML) prediction models on a large series of vertebra. Herein, we present the largest study of a purely clinical, ML prediction model.
Methods: A multi-institutional, adult cohort of patients with tumor-infiltrated vertebral bodies, spanning T1 through L5, were included. Binary outcome of fracture/progression or no fracture/progression was recorded during 12-month follow-up. Twelve clinical features (vertebral location, bone lesion (lytic, blastic, both), age, sex, BMI, histology, steroid use, radiation history, chemotherapy, osteoporosis, bisphosphonate use, and mechanical back pain) were collected. Feature selection and predictive modeling were both performed using unique random forest models. Models were evaluated using 10-fold cross-validation.
Results: One-hundred-sixty-three patients with 973 vertebrae (120 fractures/progression) were included. The average time to outcome was 4.5 months. The top 9 biomarkers most predictive of outcome occurrence in order of importance were older age, higher BMI, prior radiation, uncommon histology, osteoporosis, mechanical back pain, bisphosphonate use, steroid use, and lytic lesion. For predicting future fracture/progression, the classifier achieved AUC-ROC 0.77 (95% CI: 0.72-0.82), sensitivity 0.64 (95% CI: 0.55-0.72), specificity 0.83 (95% CI 0.80-0.85).
Conclusion : Clinical profiles remain paramount in VCF prediction. Future imaging-based modeling should optimize use of clinical attributes in final models to achieve state-of-the-art performances.
.jpg)
