296 - AI-Enabled Risk Prediction for Sepsis and Septic Shock in Hemodialysis Patients Undergoing IR Interventions

Purpose: Patients with end-stage renal disease (ESRD) undergoing dialysis-related interventions are at heightened risk of infectious complications, including sepsis and septic shock, which are major contributors to morbidity and mortality. Traditional perioperative risk stratification tools inadequately capture these outcomes in the interventional radiology (IR) setting. Machine learning (ML) approaches may enable more accurate and individualized prediction of infection-related complications.

Materials and Methods: We analyzed 5,077 patients undergoing dialysis-related procedures in the ACS-NSQIP database (2011–2018). A Tabular Prior-Data Fitted Network (TabPFN) model was trained and validated using demographic, preoperative, and procedural variables to predict 30-day sepsis and septic shock. Data were split into training (n=3,490, 2011–2016), contemporaneous testing (n=873, 2011–2016), and non-contemporaneous testing (n=714, 2017–2018) cohorts. Missing values were imputed using k-nearest neighbor methods, and class imbalance was addressed through random oversampling. Model performance was evaluated with AU-ROC, sensitivity, and specificity. Feature contributions were assessed using SHapley Additive exPlanations (SHAP).

Results: Within 30 days of intervention, 2.34% of patients developed sepsis and 0.91% developed septic shock. For sepsis prediction, TabPFN achieved an AU-ROC of 0.870 with improved sensitivity and stable specificity across both testing cohorts. For septic shock prediction, TabPFN achieved an AU-ROC of 0.891. Across outcomes, the strongest predictors were preoperative white blood cell count, serum albumin, and INR. Subgroup analyses confirmed consistent model performance across sex, race, and age groups, with no statistically significant disparities in predictive accuracy.

Conclusion: This study demonstrates the feasibility of TabPFN-based prediction of sepsis and septic shock in dialysis-related IR procedures. TabPFN provides strong discriminative ability for sepsis risk while maintaining fairness across demographic subgroups. Such models may enable earlier recognition of high-risk patients, inform pre-procedural optimization, and support proactive surveillance strategies to reduce infection-related morbidity and mortality.