SURG-70. Utilizing Machine Learning Models to Guide and Optimize Clinical Management of Cancer-Associated Compression Fractures

Spinal metastases (SM) occur in an astounding 60-70% of patients with systemic cancer and can cause pain and instability. Treatment of SM involves an increasing role for minimally-invasive procedures such as percutaneous kyphoplasty to allow for faster initiation of chemoradiation. Current clinical tools, like the Spinal Instability Neoplastic Score (SINS), are limited in combining the complex patient factors that determine the optimal treatment approach. We retrospectively analyzed a prospective database of 768 SM patients treated with kyphoplasty at a single Comprehensive Cancer Center between 2009 and 2020. Collected variables included patient demographics, tumor type, pain scores, Karnofsky Performance Status (KPS), and radiographic parameters of spinal involvement and alignment. Failure of kyphoplasty was defined as non-improved pain scores, declining KPS, multiple kyphoplasties or open surgery at sites of prior kyphoplasty. Using machine-learning (ML) models and lasso logistic regression, we evaluated patient predictors of treatment outcome. A 80:20 distribution was used for training and testing, respectively. Predictive models with the highest performance are indicated by the area under the receiver operating characteristic (AUROC). Multiple myeloma (n=254) and breast cancer (n=103) were the most common primary tumor types in the cohort. The deep learning model achieved 66% average predictive accuracy with demographic variables alone. The most important demographic predictive variables were: number of vertebral levels treated, pre-procedural KPS, age, and history of osteoporosis. Among cancer types, multiple myeloma (MM) and lung cancer (LC) were the most predictive of kyphoplasty failure. Adding radiographic parameters to the model achieved 72.3% average predictive accuracy. Demographic and radiographic parameters have a significant influence on predicting success from kyphoplasty for SM. This project leverages ML techniques to detect non-linear relationships between complex clinical and radiographic variables. This model will be optimized to aid in designing clinical tools that may help guide patient management.