Application Of Cooperative Game Theory Principles To Interpret Machine Learning Models Of Nonhome Discharge Following Spine Surgery

Study Design. Retrospective analysis of prospectively acquired data.Objective. The aim of this study was to identify interaction effects that modulate nonhome discharge (NHD) risk by applying coalitional game theory principles to interpret machine learning models and understand variable interaction effects underlying NHD risk.Summary of Background Data. NHD may predispose patients to adverse outcomes during their care. Previous studies identified potential factors implicated in NHD; however, it is unclear how interaction effects between these factors contribute to overall NHD risk.Methods. Of the 11,150 reviewed cases involving procedures for degenerative spine conditions, 1764 cases (15.8%) involved NHD. Gradient boosting classifiers were used to construct predictive models for NHD for each patient. Shapley values, which assign a unique distribution of the total NHD risk to each model variable using an optimal cost-sharing rule, quantified feature importance and examined interaction effects between variables.Results. Models constructed from features identified by Shapley values were highly predictive of patient-level NHD risk (mean C-statistic = 0.91). Supervised clustering identified distinct patient subgroups with variable NHD risk and their shared characteristics. Focused interaction analysis of surgical invasiveness, age, and comorbidity burden suggested age as a worse risk factor than comorbidity burden due to stronger positive interaction effects. Additionally, negative interaction effects were found between age and low blood loss, indicating that intraoperative hemostasis may be critical for reducing NHD risk in the elderly.Conclusion. This strategy provides novel insights into feature interactions that contribute to NHD risk after spine surgery. Patients with positively interacting risk factors may require special attention during their hospitalization to control NHD risk.