Predicting Attainment of Clinically Important Difference in the Japanese Orthopaedic Association Back Pain Evaluation Questionnaire: A Comparison of Deep Learning and Machine Learning.

STUDY

DESIGN

Retrospective study of the prospectively collected data.

OBJECTIVE

To assess the ability of three artificial intelligence (AI) models to predict attainment of clinically important differences (CIDs) in the Japanese Orthopaedic Association Back Pain Evaluation Questionnaire (JOABPEQ).

SUMMARY OF BACKGROUND DATA: Accurate prediction of postoperative functional improvement is essential for surgical planning, yet patient-reported outcome-based predictive tools have not been established.

METHODS

We retrospectively analyzed 1149 patients from three spine centers. Outcomes were five JOABPEQ domain scores and three visual analog scale (VAS) scores. Three AI models-TabNet, a deep neural network (DNN), and elastic-net penalized logistic regression (ENLR)-were trained and validated on 981 patients from two centers using stratified five-fold cross-validation. External validation was performed on an independent cohort of 168 patients from the third center. Input features included age, sex, preoperative JOABPEQ item responses, domain scores, and VAS scores. Model performance was evaluated by the area under the receiver operating characteristic curve (AUC) and accuracy.

RESULTS

In external validation for JOABPEQ domains, TabNet achieved a mean AUC of 0.79 and accuracy of 0.74; DNN, AUC 0.77 and accuracy 0.73; and ENLR, AUC 0.78 and accuracy 0.74. For VAS outcomes, TabNet yielded a mean AUC of 0.80 and accuracy of 0.74; DNN, AUC 0.77 and accuracy 0.72; and ENLR, AUC 0.78 and accuracy 0.72.

CONCLUSIONS

All three AI models reliably predicted postoperative improvements. Such AI-based prediction models may enhance clinical decision-making and patient counseling in lumbar spine surgery.