European Spine Journal - 2026-08-06 - Journal Article
Multimodal model predicts screw loosening at user-specified postoperative time points after single-level lumbar interbody fusion.
Tanioka S, Aydin OU, Hilbert A, Miyazaki T, Yamamoto A, Ikezawa M, Nishikawa H, Fujimoto M, Ishida F, Kamei Y, Yoshida K, Shoda M, Suzuki H, Mizuno M, Frey D
Topics
Key Takeaway
A multimodal deep learning model using preoperative vertebral body CT images predicted pedicle screw loosening after single-level lumbar interbody fusion with AUC 0.823 and sensitivity 0.914 on external testing.
Summary Depth
Choose how much analysis to show on this article page.
Summary
This study developed a multimodal model to predict pedicle screw loosening at user-specified postoperative time points following single-level lumbar interbody fusion using preoperative CT images and postoperative day as inputs. Three CT input types were compared across training (412 scans, 3 hospitals) and external testing (140 scans, 2 hospitals) cohorts; the vertebral body CT model achieved AUC 0.823 and sensitivity 0.914. Adding clinical characteristics did not improve performance, with AUC dropping to 0.721.
Key Limitation
The study does not report bone mineral density or Hounsfield unit thresholds as covariates, omitting the most established quantitative predictor of screw loosening and limiting mechanistic interpretation of the model's CT-based signal.
Original Abstract
PURPOSE
To develop a model that predicts screw loosening at user-specified postoperative time points after lumbar fusion.
METHODS
We retrospectively enrolled patients who underwent single-level lumbar interbody fusion at five hospitals: three hospitals for training and two for external testing. Preoperative CT images, clinical characteristics, and postoperative follow-up CT images were collected. Many patients had multiple follow-up CT scans, with screw loosening evaluated for each. Multimodal models were developed using preoperative CT images and the day of the follow-up CT scans as inputs and the occurrence of screw loosening on that day as the label. Three types of preoperative CT images were prepared and used for training: original, whole vertebra, and vertebral body. The results were evaluated in testing, and the best model with the highest AUC among the three was selected. Additionally, clinical characteristics were incorporated as an additional input, and the result was also evaluated.
RESULTS
A total of 212 patients (mean age, 66.5 ± 11.3, 123 male) with 552 follow-up CT scans were included: 412 follow-ups for training and validation, and 140 follow-ups for testing. The model using vertebral body showed the best performance, with an AUC of 0.823 (95% CI, 0.737-0.895) and a sensitivity of 0.914 (95% CI, 0.769-0.999). When clinical characteristics were added to the input, no improvement was observed, with an AUC of 0.721 (95% CI, 0.621-0.817).
CONCLUSION
The model predicted screw loosening at any postoperative time point with high AUC and sensitivity and has the potential to enable more thorough patient management.