AI-Enabled Tool for 30-Day Mortality Risk Stratification in Hemodialysis Patients Undergoing Interventional Radiology Procedures

Purpose: It is essential to predict postoperative mortality risk in hemodialysis patients, enabling physicians to tailor and optimize their pre-procedural care. This study aims to develop an AI-enabled tool for stratifying 30-day mortality risk in hemodialysis patients undergoing common dialysis-related IR interventions.

Materials and Methods: Data from the American College of Surgeons’ (ACS) National Surgical Quality Improvement Program (NSQIP) database for 16,267 hemodialysis patients (mean age 59.25 ± 15.10) undergoing dialysis-related interventions between 2011 and 2018 was used for this tool (CPT codes: 35476, 36145, 36870, 36901-9, 37205-6, 37246-9, 37607, 49418-9, 49421-2, 75790-1, 75798, 75960-1). Data from 2011 to 2017 was split to be used for model training (70%), tuning (20%), and contemporaneous testing (10%). The 2018 dataset (12.22% of the whole dataset) was reserved for non-contemporaneous testing/validation. The Random Forest prediction model was evaluated on demographic factors such as age ( < 65/≥65), gender (male/female), and race (white/non-white) to enhance model interpretability. The primary outcome was 30-day mortality. Model performance was assessed using accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), and receiver operating characteristic (ROC) curves. DeLong p-values were calculated to evaluate differences in model performance across demographic subgroups.

Results: The tool demonstrated strong performance in predicting 30-day mortality across both contemporaneous and non-contemporaneous testing sets. The overall accuracy was 92.3% (47% sensitivity, 94% specificity, 15% PPV, 99% NPV) in the contemporaneous test set and 91.7% (54% sensitivity, 93% specificity, 19% PPV, 99% NPV) in the non-contemporaneous test set, with Area Under the ROC Curve (AUROC) values of 0.85 and 0.85, respectively. The model’s performance was significantly different between white and non-white patients in the non-contemporaneous validation set with an AUROC of 0.81 for white patients and 0.93 for non-white patients (P < 0.01), statistically better performance for the non-white subgroup. Stratified analyses across age and gender subgroups showed consistent predictive performance, with AUROC over 0.80 for all groups.

Conclusion: This tool provides a robust, validated approach for accurately predicting 30-day postoperative mortality risk in hemodialysis patients undergoing dialysis-related interventions.