Secondary Logo

Journal Logo

Presentation #66: Predictive Models for Patient‐Centered Efficacy and Discharge Destination after Elective Cervical Spine Surgery

Sivaganesan, Ahilan MD; Chotai, Silky MD; Kim, Elliott J. MD; Stonko, David BS, MS; Wick, Joseph Bradley BA; McGirt, Matthew MD; Devin, Clinton J. MD

Author Information
Spine Journal Meeting Abstracts: 2016 - Volume 2016 - Issue - p 226–228
  • Free

Introduction: Surgery is a valuable therapeutic option for degenerative cervical spine disease, however there is uncertainty as to which patients benefit. Here we introduce predictive models for clinically meaningful improvement in disability, as well as discharge destination, after cervical spine surgery (CSS).

Methods: 430 patients undergoing CSS were enrolled into a prospective registry. LOESS regression was performed to verify that a linear relationship between 12‐month Neck Disability Index (NDI) and various explanatory variables was reasonable. The following variables were used to power a multiple linear regression model for NDI: demographics, diagnosis, number / location of diseased levels, baseline symptoms and PROs, employment / insurance status, comorbidities, a history of prior surgeries, and surgical approach. Possible interactions among variables such as diagnosis, age, baseline NDI, and employment status were also accounted for in the analysis. We then used Repeated Random Sub‐Sampling (related to Monte Carlo cross‐validation) to validate the predictive performance of our model. A separate model, based on logistic regression, was constructed to predict a clinically important improvement in NDI (at least 17.3) at one year. A third model was also developed and validated, using similar methods, with the aim of predicting post‐surgery discharge destination (home versus facility).

Results: The mean NDI one year after surgery was 25.82, and the mean improvement was 16.33 points. 48% (205) of patients achieved the minimum clinically important difference (MCID) in NDI. Our predictive model for 12‐month NDI has an R‐squared of 0.69 (observed versus predicted NDI scores are plotted in Figure 1), and in validation, it achieved an R‐squared of 0.43. The predictors, in descending order of influence, are: employment, baseline NDI, diagnosis, smoking, ethnicity, claudication, narcotic use, and symptom duration. Our model for achieving a MCID in NDI has an area under the curve greater than 0.80 for the development phase and an AUC of 0.65 for the validation phase. The predictors, in descending order of influence, are: baseline NDI, motor deficit, depression, ambulation, revision surgery, employment, diagnosis, smoking, and symptom duration. Finally, our predictive model for discharge destination has an area under the curve greater than 0.80 for the development phase and an AUC of 0.75 for the validation phase (ROC curve shown in Figure 2). The predictors, in descending order of influence, are: baseline EQ‐5D, number of levels, myelopathy, depression, baseline NDI, and motor deficit.

Figure 1
Figure 1
Figure 2
Figure 2

Conclusion: We present internally validated models that can help predict disability at one year, clinically meaningful improvement in disability, and discharge destination after elective CSS. Our NDI model explains roughly 70% of the variation in 12‐month neck‐related disability. The predictive accuracy of our associated model for achieving a MCID in NDI is a good starting point, but leaves room for improvement. Our model for discharge destination has strong predictive accuracy, and with external validation at other institutions, it can become a useful tool as spine care providers seek to better understand the postoperative trajectories of their patients.

© 2016 Lippincott Williams & Wilkins, Inc.