A DATA-DRIVEN CLASSIFICATION OF DEGENERATIVE CERVICAL MYELOPATHY LEADS TO CLINICALLY RELEVANT SUBGROUPS WITH DISTINCT PREOPERATIVE FEATURES AND POST-SURGICAL OUTCOMES: A CSORN STUDY

Outcome prognostication of patients with Degenerative Cervical Myelopathy (DCM) is difficult due to the significant heterogeneity that exists within this population. We aim to identify subgroups of DCM patients with clinical similarities and distinct outcome patterns using a data driven approach. K-prototype clustering was applied to a national prospective DCM database to identify patient subgroups based on relevant features influencing outcomes documented in the literature (Age, time with condition, mJOA, SF12, EQ5D, MRI cord findings, motor and sensory deficits, number of levels involved, number of comorbidities and current work status). Outcome measures using mJOA, SF12 (PCS and MCS), EQ5D, NDI, work status, residual pain and satisfaction with surgery were computed for each cluster and compared using Kruskal-Wallis and Mann-Withney-U test. Data on 774 patients from was analyzed. A classification with 4 clusters was developed with each cluster having statistically significant different means for all demographics, mJOA and PROM baseline scores as well as one year post-op total and change in mJOA score, SF12, EQ5D and NDI. The 4 clusters with distinct features and prognosis were defined as follows: -cluster 1: mild disease with radiographic findings with least mJOA score and PROM improvement post-operatively; -cluster 2: youngest patients with moderate disease and high functional recovery; -cluster 3: oldest patients with most co-morbidities with moderate disease and good baseline mental state but gaining the least functionally from surgery; -cluster 4: patients with most severe disease and longest duration with most significant improvement offunctional and mental health scores. Despite the distinct features between the groups, all clusters were similarly satisfied with the surgery and would repeat it. Using an unsupervised machine learning algorithm, we identified four clinically relevant subgroups of DCM patients. Each clusters had distinguishable features that influenced outcome measures in distinct directions. Clustering of DCM patients could help physicians segment that population and help provide more nuanced counselling on outcomes using this new type of data-driven stratification of the DCM population.