68/100

Moderate evidence

Cite
More
EmailLinkedInX

Scan-based AI model predicted recovery after spinal surgery in 333-patient study

Study title: Prediction of Postoperative Neurological Function Recovery in Cervical Spondylotic Myelopathy Using MRI-based Delta-Radiomics Combined with Interpretable Machine Learning Model

Evidence type
Controlled or validated comparison
Sample
n = 333
Sources
1 sourcePeer-reviewed
Published
2 min read

Study title: Prediction of Postoperative Neurological Function Recovery in Cervical Spondylotic Myelopathy Using MRI-based Delta-Radiomics Combined with Interpretable Machine Learning Model

Researchers built a machine learning model to predict how well patients recover neurologically 12 months after decompression surgery for cervical spondylotic myelopathy, a condition that causes progressive neurological decline. The model uses changes between pre- and postoperative T2-weighted MRI scans, known as delta-radiomics. Its best version, built with XGBoost, reached an AUC of 0.899 in internal testing and 0.802 in external validation.

Read the study abstract

Cervical Spondylotic Myelopathy (CSM) presents progressive neurological deterioration and highly variable postoperative outcomes, while existing prognostic assessment methods have inherent limitations. This dual-center study developed an interpretable MRI-based delta-radiomics model for individualized CSM postoperative prognosis evaluation. A total of 237 CSM patients from Center 1 and 96 patients from Center 2 who underwent decompression surgery were retrospectively enrolled. The Center 1 cohort was stratified…

The gain

In a two-hospital study of 333 patients who had decompression surgery for cervical spondylotic myelopathy, an XGBoost model using changes between pre- and postoperative MRI scored an AUC of 0.899 on internal testing and 0.802 on external validation for predicting 12-month neurological recovery.

In brief

The study retrospectively enrolled 237 patients from one centre and 96 from a second. The first group was split into 165 training and 72 internal test patients, and the second group served as an independent external validation set. The delta-radiomic model was statistically better than models using only preoperative radiomic features or routine imaging features.

The model relies on routine MRI taken within 7 days of surgery, so the authors describe it as a possible objective tool for early prediction of recovery. Its accuracy fell from internal to external testing, and the work was retrospective.

Main points

  1. The study included 237 patients from Center 1 and 96 from Center 2, all of whom had decompression surgery for cervical spondylotic myelopathy.
  2. Center 1 patients were split 7:3 into training (165) and internal testing (72) groups; Center 2's 96 patients formed the external validation set.
  3. Features came from pre- and postoperative T2-weighted MRI; 16 preoperative radiomic, 15 delta-radiomic and 5 clinical-routine imaging features were selected.
  4. Three models were built using five machine learning algorithms and compared with DeLong and McNemar tests.
  5. The XGBoost delta-radiomic model had AUCs of 0.899 internally and 0.802 externally, significantly better than the other two models.
  6. Decision curve analysis showed higher clinical net benefit for the delta-radiomic model across wide threshold ranges.
  7. SHAP analysis identified wavelet-domain texture features as risk factors for poor prognosis, while gradient-domain kurtosis features had bidirectional effects.

The full story

The team extracted radiomic features from T2-weighted MRI before and after surgery and calculated delta-radiomic features capturing the change. LASSO regression and Spearman correlation analysis narrowed the features to a core set, and five machine learning algorithms produced three prediction models.

Performance was judged on AUC, accuracy, sensitivity, specificity and F1-score, with calibration curves and radar charts used to pick the best algorithm for each model. The delta-radiomic model showed a more balanced performance profile across cohorts and better generalizability than the alternatives.

SHAP analysis was used to explain the model's predictions. Wavelet-domain texture features were linked to poor prognosis, and gradient-domain kurtosis features showed bidirectional effects. The authors say that, because it uses routine MRI within 7 days of surgery, the model holds promise for early prediction of 12-month neurological recovery.

Sources

1 on file
  1. Journal of Imaging Informatics in Medicinedoi:10.1007/s10278-026-02521-x
    Peer-reviewed2026-11-17

Talk it over

Three questions to start on, two more if you have time
The model scored an AUC of 0.899 in internal testing on 72 patients and 0.802 in external validation on 96 patients from a second centre. How would you read that drop when judging whether it is ready for use?

Start with a question

Scan-based AI model predicted recovery after spinal surgery in 333-patient studyhttps://truvace.com/item/prediction-of-postoperative-neurological-function-recovery-in-cervical-spondylot

  1. The model scored an AUC of 0.899 in internal testing on 72 patients and 0.802 in external validation on 96 patients from a second centre. How would you read that drop when judging whether it is ready for use?

    Answer in the thread
  2. If you were a patient with cervical spondylotic myelopathy deciding on decompression surgery, how would a prediction of 12-month neurological recovery, made from MRI taken within 7 days after the operation, change what you want to know?

    Answer in the thread
  3. Both centres enrolled patients retrospectively. Who might be missing from those 333 patients, and how could that change how well the model predicts recovery elsewhere?

    Answer in the thread
If you have more time
  • The SHAP analysis linked wavelet-domain texture features to poor prognosis. What would you want a surgeon to be able to explain to a patient before relying on that kind of output?

    Answer in the thread
  • What kind of follow-up study, in which hospitals and with which patients, would settle whether the delta-radiomic model beats the preoperative-only model in real care?

    Answer in the thread

What readers say