AI-Assisted, Literature-Informed Development and Retrospective Validation of a Point-Based Surgical Site Infection...

Study DesignRetrospective cohort study.ObjectivesPostoperative surgical site infections (SSI) remain a major cause of morbidity and cost in spine surgery. Existing risk calculators have limited applicability in this population. This study introduces a literature-informed, point-based SSI calculator designed to complement standard preoperative assessment. GPT-4 (OpenAI, San Francisco, CA) was used solely for structured literature synthesis and preliminary variable weighting.MethodsAdult patients undergoing spine surgery at a single academic center (2019-2025) were retrospectively reviewed. A 29-variable AI-assisted risk calculator integrating demographic, clinical, laboratory, and surgical factors was developed. SSI was defined using CDC/NHSN criteria with a 90-day window. Risk calculator performance was evaluated using receiver operating characteristics (ROC) curve analysis, bootstrap optimization correction, and calibration and decision curve analyses. A stratified analysis was performed comparing performance by procedure types and spinal regions.ResultsA total of 338 patients were included (SSI: 177; controls: 161). Median risk scores were significantly higher in infected vs non-infected patients (18 vs 10; P < .001). Discrimination was strong (AUC 0.7978, 95% CI 0.7521-0.8427). The optimal threshold was 17 points, yielding 61.0% sensitivity, 84.2% specificity, and 72.1% accuracy. Bootstrap validation showed minimal optimism (corrected AUC 0.7974). Calibration was excellent (calibration-in-the-large <0.001; slope 1.045; Brier score 0.1833). At the study prevalence (52.3%), PPV was 0.8106 and NPV 0.6540, with expected PPV reduction at real-world prevalence. Patients with scores ≥17 had significantly higher infection risk (RR 2.42; OR 8.31; P < .001). Stratified analyses showed consistent performance across procedure types and spinal regions.ConclusionThis internally validated, AI-assisted SSI calculator demonstrated strong discrimination and calibration. Prospective external validation is needed to determine clinical utility.