Radiomics-based machine learning model integrating preoperative vertebral computed tomography and clinical features to predict cage subsidence after single-level anterior cervical discectomy and fusion with a zero-profile anchored spacer

Comprehensive Summary

This study aimed to develop a machine learning model that can predict cage subsidence after single-level anterior cervical discectomy and fusion (ACDF) with a Zero-P implant system via clinical data and radiomics. 253 patients with cage subsidence ( ≥3 mm loss of fused-segment height) were retrospectively reviewed and split into both a training set and an independent test set. Each patient’s clinical variables were collected, such as age, bone density (HU), and T1 slope angle. Radiographs were then used to measure overall cervical alignment and balance. Three predictive models were built: clinical data only, radiomics only, and a combined model. Eight algorithms were tested to find the best performing models. After testing, the combined model scored the best, with an AUC around 0.813, higher than the radiomics model (0.775) and the clinical only model (0.595). This combined model is concluded to be an accurate predictor of cage subsidence, combining radiomic information and clinical data.

Outcomes and Implications

Patients undergoing ACDF are at a higher risk for cage subsidence, which can compromise the outcome of the surgery and reduce quality of life in the patient. For a physician, understanding the risk for an individual patient of cage subsidence before the surgery is performed is a critical advantage than can help plan the surgery and the post-operative care steps. The model built in this study provides just that, detecting cage subsidence pre-operatively with high accuracy.