General Session: Cervical Degenerative - Hall F
Presented by: J. Badhiwala
J. Badhiwala(1), C. Witiw(1), F. Nassiri(1), M. Akbar(1), A. Mansouri(1), G. Ibrahim(1), J. Wilson(1), M. Fehlings(1)
(1) University of Toronto, Division of Neurosurgery, Department of Surgery, Toronto, ON, Canada
Introduction: Predictors of outcome after surgery for degenerative cervical myelopathy (DCM) have been determined previously through hypothesis-driven multivariate statistical models that rely on a priori knowledge of potential confounders, exclude potentially important variables due to restrictions in model building, cannot include highly collinear variables in the same model, and ignore intrinsic correlations among variables. We sought to apply a data-driven approach, principal component analysis (PCA), to identify patient phenotypes that may predict outcomes after surgery for mild DCM using data from two related prospective, multi-center cohort studies.
Methods: We captured patients with mild DCM, defined by a modified Japanese Orthopaedic Association score of 15, 16, or 17, enrolled in the AOSpine CSM-NA or CSM-I trials. All patients underwent spinal decompression by an anterior, posterior, or combined approach. Patient outcomes were evaluated at baseline and 6 months, 1 year, and 2 years after surgery. Quality of life was evaluated by the Neck Disability Index (NDI) and SF-36v2. A heterogeneous correlation matrix was created using a combination of Pearson, polyserial, and polychoric regressions among 67 baseline variables, which then underwent eigen decomposition. Scores of significant principal components (PCs) (those with eigenvalues > 1) were included in multivariate logistic regression analyses for three outcomes of interest: achievement of the minimum clinically important difference (MCID) in 1) NDI (≤ -7.5); 2) SF-36v2 Physical Component Summary (PCS) score (³ 5); and 3) SF-36v2 Mental Component Summary (MCS) score (³ 5).
Results: A total of 154 patients met eligibility criteria and had complete data. Twenty-four significant principal components, accounting for 75% of the variance in the data, were identified. Two principal components were associated with achievement of the MCID in NDI. The first (PC 1) was dominated by variables related to surgical approach and number of operated levels; the second (PC 21) consisted of variables related to patient demographics, severity and etiology of DCM, comorbid status, and surgical approach. Both PC 1 and PC 21 also correlated with SF-36v2 PCS score, in addition to PC 4, which described patients' physical profile, including gender, height, and weight; PC 6, which received large loadings from variables related to cardiac disease, impaired mobility, and length of surgery and recovery; and PC 9, which harbored large contributions from features of upper limb dysfunction, cardiorespiratory disease, surgical approach, and region. In addition to PC 21, a component profiling patients' socioeconomic status and support systems and degree of physical disability (PC 24), was associated with achievement of the MCID in SF-36 MCS score. Biplots for PCs associated with outcomes are shown in Figure 1.
Conclusions: Through a data-driven approach, we identified several phenotypes associated with disability and physical and mental health-related QOL. Such data reduction methods may separate patient-, disease-, and treatment-related variables more accurately into clinically meaningful phenotypes that may inform patient care and recruitment into clinical trials.