# Comparison of Artificial Intelligence Models Using CT Radiomics for Predicting Post-Vertebral Augmentation Residual Back Pain in Osteoporotic Vertebral Compression Fractures

## TL;DR

This study compares AI models using CT scans and clinical data to predict residual back pain after spinal surgery for osteoporotic fractures.

## Contribution

The study introduces a deep learning framework combining CT radiomics and clinical parameters to predict postoperative residual back pain.

## Key findings

- TabNet achieved the highest performance with an AUROC of 0.927 and recall of 0.833.
- Intravertebral vacuum cleft and bone mineral density were key clinical predictors.
- Radiomics features improved specificity in identifying non-RBP cases.

## Abstract

Background: Residual back pain (RBP) following vertebral augmentation (VA) represents a significant challenge in managing osteoporotic vertebral compression fractures (OVCFs). While conventional predictive models have shown moderate accuracy, their preoperative risk stratification capabilities remain suboptimal. CT-based radiomics has demonstrated success in vertebral fracture assessment, yet its integration with artificial intelligence (AI) for predicting RBP remains unexplored.

Objective: This study aims to identify the optimal AI model for predicting RBP by systematically comparing multiple algorithms that integrate CT radiomics features with clinical parameters, with the goal of enabling preoperative risk stratification for improved surgical decision-making.

Methods: This prospective study enrolled patients who underwent VA for OVCFs. Potential predictors were identified through clinical variable analysis. Radiomics features were extracted from preoperative CT images using standardized vertebral segmentation protocols. The study population was divided into training and testing cohorts at a ratio of 7:3. Five AI models were constructed through integration of clinical predictors and radiomics features. Model performance evaluation was conducted in the independent testing cohort through discrimination, calibration, and clinical utility analyses. The predictive mechanisms of the optimal model were interpreted through feature importance analysis.

Results: Among 856 enrolled patients, RBP developed in 102 cases (11.9%). TabNet exhibited optimal performance metrics (AUROC: 0.927, Recall: 0.833) among all evaluated algorithms. Feature importance analysis revealed intravertebral vacuum cleft and bone mineral density as principal clinical predictors, complemented by wavelet-based texture parameters and quantitative intensity metrics. Ablation experiments demonstrated that clinical parameters were critical for false-positive reduction, while radiomics features enhanced specificity in non-RBP identification. The model maintained consistent clinical utility across varying threshold probabilities.

Conclusion: The integration of clinical parameters and CT-based radiomics through a deep learning framework enabled accurate preoperative prediction of RBP.