A machine learning tool for prediction of vertebral compression fracture following stereotactic body radiation therapy for spinal metastases.

Abstract

The most common adverse event following spine stereotactic body radiotherapy (SBRT) is vertebral compression fracture (VCF). There is interest in the development of patient-specific tools that can predict those at high risk of developing VCF. This study aimed to develop a machine learning tool able to predict the development of VCF following spine SBRT using clinical, dosimetric and tumor risk factors. A retrospective review of a prospectively maintained database of spinal segments treated with SBRT for spinal metastases was utilized. Machine learning models were applied to this dataset and their ability to predict for VCF was evaluated. Data was split into training and validation sets. Spinal Instability Neoplastic Score (SINS) is the current clinical standard for predicting spine instability in the setting of metastatic disease and served as the baseline model for comparison. Between 2008 and 2021, 1406 spinal segments were contained within the database. Logistic regression, neural network/multi-layer perceptron, support vector machine and random forest were the machine learning models applied to the dataset. Their accuracy, precision, F1-score, sensitivity and specificity was determined, together with that of SINS. Across performance metrics, the machine learning models outperformed SINS with random forest model having the best performance. Important factors that increase the risk of VCF were identified and include age, pain, extent of pre-existing VCF, location and spinal alignment. A machine learning model predicting for VCF following spine SBRT has been developed. This model outperformed the current clinical standard of SINS in the prediction of VCF.