(I-521) Application of Cooperative Game Theory Principles to Interpret Machine Learning Models of Proximal Junctional Failure Following C2-T2 Posterior Fusions
Resident Physician Mayo Clinic Rochester Rochester, Minnesota, United States
Introduction: Proximal junctional failure (PJF) is a known complication of cervicothoracic spine deformity surgery and may result in reduced surgical efficacy, increased postoperative pain and disability, and eventual revision surgery. Explainable machine learning models that accurately predict PJF using diverse demographic and radiographic parameters would provide a valuable tool for surgeons looking to better anticipate this potential complication and would lend new insights into the most important features contributing to PJF risk. The purpose of this study was to apply solution concepts from cooperative game theory to machine learning models to understand the most important demographic and radiographic features that underlie PJF risk in a large series of spine patients who underwent a C2-T2 instrumented fusion.
Methods: A total of 175 consecutive cases involving a C2-T2 instrumented fusion were reviewed for pre- and post-operative demographic and imaging-based biomechanical measurements to understand the predictors of proximal junctional failure in this patient population. Gradient boosting machine learning models were used to construct predictive patient-level models of proximal junctional failure. Shapley values, which apply an optimal cost-sharing rule to assign a unique distribution of the total risk of PJR to each factor in the model, were then calculated to quantify feature importance across the study population.
Results: The mean follow up for this study was 275.05 17.20 days. Gradient boosting models constructed from features identified by optimized Shapely values performed well in predicting patient-level PJF risk (mean C-statistic=0.80). The most important factors for predicting proximal junctional failure following C2-T2 posterior fusions were evidence of C2 screw loosening on 6-month follow up X-ray, presence of postoperative T1-4 kyphosis, increase in occiput to C2 Cobb angle from immediate post-op to 6-month follow up, and lower postoperative PROMIS social scores.
Conclusion : This strategy provides novel insights into global patterns of feature importance that are valuable in predicting PJF after C2-T2 fusion. Patients with certain demographic and radiographic parameters, such as postoperative C2 screw loosening, T1-T4 kyphosis, and low PROMIS Social scores, may require closer observation given an elevated risk of subsequent proximal junctional failure.
