44. Development of a novel machine learning model for prediction of adjacent fracture after cementoplasty in treating osteoporotic vertebral compression fracture

Abstract

BACKGROUND CONTEXT

There are approximately 30% of patients with osteoporotic vertebral compression fracture (OVCF) who need cementoplasty for treatment. However, the occurrence of adjacent vertebral fracture (AVF) postoperatively can lead to increased pain, delayed recovery, and poorer prognosis. Current literature identifies over 30 risk factors for AVF, including patient-specific factors, preoperative and postoperative radiographical features, and surgical-related factors. There is no effective predictive model in understanding the probability of AVF occurrence preoperatively.

PURPOSE

This study aims to develop a robust AVF predictive model using machine learning method.

STUDY DESIGN/SETTING

Retrospective cohort study.

PATIENT SAMPLE

A total of 238 patients with OVCF who underwent single level cementoplasty were included for analysis.

OUTCOME MEASURES

Adjacent fracture.

METHODS

This is a retrospective cohort analysis. Patients with OVCF who underwent single level cementoplasty between January 2016 and December 2021 were included. Exclusion criteria were pathological fractures, patients with prior cementoplasty or spinal surgeries, and follow-up less than 12 months. Total 32 preoperative clinical and radiographic features were recorded, include patient demographics, DXA, chronic diseases, vertebral height (VH), wedge angle (WA) of fracture vertebra, local kyphotic angle (LKA), presence of posterior wall fracture (PostWall), and presence of diffuse idiopathic skeletal hyperostosis (DISH), CT vertebral Hounsfield units (HU), CT psoas lumbar vertebral index (PLVI). Ten different machine learning algorithms were used to find the best model. Confusion matrix and related indicators include Accuracy, sensitivity (Se), specificity (Sp) and ROC-AUC were used to evaluate the model performance.

RESULTS

A total of 238 patients were included for analysis, with an average age of 77 years and 69% were female. Most fractures located at the TL junction (64%). The AVF rate was 27.3% during the follow-up and it occurred at postoperative 3.2 months. We found the random forest model had the best performance with 83% accuracy, AUC 0.92, Se: 82%, and Sp: 85%. Among the total 32 features, we found that the 11 most important features by orders were PostWall, HU_L2, DISH, L4_PLVI, WA, MVH, BMI, LKA, Age, and fracture level. Even using those 11 features alone, the model performance could reach 78% accuracy, AUC 0.88, Se: 80%, and Sp 76%.

CONCLUSIONS

The novel machine learning model for predicting AVF using preoperative features demonstrated excellent performance, achieving an AUC of 0.92. This model can assist clinicians and patients with OVCF in understanding the probability of AVF occurrence after cementoplasty. For patients identified as high-risk, prophylactic cementoplasty at adjacent levels or other medical interventions may provide some benefits.

FDA Device/Drug Status

This abstract does not discuss or include any applicable devices or drugs.