Decision-Making in Lumbar Spinal Canal Stenosis: An Artificial Neural Networks Analysis

Parisa Azimi, Edward C Benzel, Sohrab Shahzadi, Shirzad Azhari, Hassan Rezad Mohammadi


Background: To develop an artificial neural network (ANN) model to determine a sound method for selecting patients for surgery or non-surgical options and to compare it with the traditional approach in lumbar spinal canal stenosis (LSCS) patients.

Materials and Methods: An ANN model and a logistic regression (LR) model were used as predicting models. Fifteen factors were recorded as the input variables for developed ANNs and LR: Age, gender, duration of symptoms, and measures of visual analog scale (VAS) of leg pain/numbness, the Japanese Orthopaedic Association (JOA) Score, the Neurogenic Claudication Outcome Score (NCOS), the Oswestry disability index (ODI), the Swiss Spinal Stenosis Score (SSS), the stenosis bothersomeness index (SBI), the dural sac cross-sectional surface area (DSCA), the Stenosis Ratio (SR), the Self-Paced Walking Test (SPWT), morphology grade presented by Schizas et al. and grading system introduced by Lee et al. Successful outcome was recorded based on the criteria presented by Stucki et al. ANNs were fed with patients’ data in order to choose surgical versus non-surgical treatment options. Sensitivity analysis was applied for the developed ANN model to identify the important variables. Receiver operating characteristic (ROC) analysis, Hosmer-Lemeshow (H-L) statistics and accuracy rate were calculated for evaluating the two models.

Results: The data for a total of 346 of 379 patients (143 male, 203 female, mean age 59.5±11.5 years) were available for the analysis. Patients’ information was divided into training (n = 174), testing (n = 86), and validation (n = 86) data sets. Successful outcome were 93.4% (surgery) and 89.4% (non-surgery) at 1-year follow-up. The SR, morphology grade and grading system were important variables identified by the ANN. The ANN model displayed better accuracy rate (97.8 %), a better H-L statistic (41.1 %) which represented a good-fit calibration, and a better AUC (89.0%), compared to the LR model.

Conclusion: The findings showed that an ANN model can predict the optimal treatment choice for LSCS patients in clinical setting and is superior to LR model. Our results will need to be confirmed with external validation studies.


Lumbar Spinal Canal Stenosis, Predicting, Surgical Satisfaction, Artificial Neural Networks (ANNs), Logistic Regression