Page 44 - WNS 2024 -e-Program Edition_Neat

A Machine Learning Model To Predict 5 Year Post-Operative
Back Pain in Patients with Grade 1 Lumbar Spondylolisthesis: A
Quality Outcomes Database Study
Anthony DiGiorgio, DO, MHA
UCSF School of Medicine
Praveen Mummaneni, MD, UCSF School of Medicine; Satvir Saggi, BS,
UCSF School of Medicine; Christopher I Shaffrey, MD, Duke University;
Andrew K Chan, MD, Columbia University; Mohamad Bydon,MD, Mayo
Clinic; Regis W Haid, MD, Atlanta Brain and Spine; Erica F Bisson, MD,
University of Utah; Steven D Glassman, MD, Norton Healthcare; Leah
Carreon, MD, Norton Healthcare; Kevin T Foley, MD, University of
Tennessee; Eric A Potts, MD, Goodman Campbell; Mark E Shaffrey, MD,
University of Virginia; Domagoj Coric, MD, Carolinas Spine and
Neurosurgical Associates; John J Knightly, MD, Maxim Brain and Spine;
Scott Meyer, MD, Maxim Brain and Spine; Paul Park, MD, University of
Tennessee; Michael Y Wang, MD, University of Miami; Kai-Ming G Fu, MD,
Weill Cornell; Jonathan R Slotkin, MD, Geisinger Health; Anthony L Asher,
MD, Carolinas Spine and Neurosurgical Associates; Michael S Virk, MD,
Weill Cornell; Dean Chou, MD, Columbia University
Introduction: Back pain is a common symptom in patients with lumbar
spondylolisthesis. Machine learning (ML) can predict improvement in back
pain following surgery in patients with grade 1 lumbar spondylolisthesis. We
evaluated predictors of achievement of the minimum clinically important
difference (MCID) in back pain and related disability after surgery in
patients with grade 1 spondylolisthesis using ML models.
Methods: This was a prospective analysis using the Quality Outcomes
Database consisting of patients with grade 1 lumbar spondylolisthesis. 608
patients were split into an 80% training/20% testing cohort.
Hyperparameter tuning was performed with 5 fold cross-validation.
Recursive feature selection was used to select key variables for predicting
MCID achievement in Numerical Rating Scale Back Pain(NRS-BP) and
Oswestry Disability Index(ODI). The final model was tested for accuracy on
the testing cohort.
Results: In total, 70% of patients achieved MCID for NRS-BP at the 5 year
post-operative period while 66% achieved MCID for ODI at the 5 year post-
operative period. Of the algorithms tested, logistic regression demonstrated
the best accuracy (0.77±0.03), followed by AUROC (0.75±0.04) at predicting
MCID achievement for NRS-BP at 5 years post-operatively. Similarly, logistic
regression demonstrated the best accuracy (0.71±0.04), followed by AUROC
(0.73±0.04) at predicting MCID achievement for ODI at 5 years post-
operatively. Top variables for predicting MCID for NRS-BP include baseline
NRS-BP, baseline NRS-Leg Pain (NRS-LP), baseline ODI, ASA grade, and
age at time of surgery. Top variables for predicting MCID for ODI included
baseline ODI, NRS-LP, educational level, baseline NRS-BP, and smoking
status.

Conclusion: Logistic regression performed the best of all models tested.
The top 5 variables for predicting MCID for NRS-BP included baseline NRS-
BP, baseline NRS-LP, baseline ODI, ASA grade, and age at time of surgery.
The top 5 variables for predicting MCID for ODI included baseline ODI,
baseline NRS-Leg Pain, educational level, baseline NRS-BP, and smoking
status.