Abstract: FR-PO0017
Competing-Risk Random Survival Forest for First Thrombectomy: A Multicenter Saudi Hemodialysis Cohort
Session Information
- Artificial Intelligence: Clinical Prediction, Decision Support, and Digital Kidney Care
October 23, 2026 | Location: Exhibit Hall A, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Ekrikpo, Udeme Ekpenyong, DaVita Health Care, Riyadh, Saudi Arabia
- Moussa, Ayman S., DaVita Health Care, Riyadh, Saudi Arabia
- Hussien, Mohamed Abdelaziz, DaVita Health Care, Riyadh, Saudi Arabia
- Abdelghaffar, Hossam Hassan, DaVita Health Care, Riyadh, Saudi Arabia
- Al kalkami, Amna, DaVita Health Care, Riyadh, Saudi Arabia
- Das, Partha, DaVita International, London, United Kingdom
- Alobaili, Saad S., King Saud University Medical City, King Saud University, Saudi Arabia
Background
Vascular access thrombosis causes dialysis interruption/hospitalization in chronic hemodialysis. Conventional models assume linearity, censor competing events, and may miss interactions among access-patency predictors. Competing-risk random survival forests (CR-RSF) accommodate nonlinear effects and competing events distribution-free. We applied CR-RSF for thrombectomy risk stratification and predictor identification
Methods
We analyzed 2477 chronic hemodialysis patients with AVF/G access from 23 DaVita Saudi clinics (Jan 2023–Mar 2026). Time zero was first cannulation; outcome was first thrombectomy with death/transplant as competing events. A 500-tree CR-RSF used 35 demographic and clinical predictors; top predictors were identified by consensus rank-sum of anti-VIMP and minimal depth. Performance was assessed by out-of-bag (OOB) time-dependent AUC, calibration, and decision-curve analysis (DCA) at 1 and 2 years
Results
Mean age was 56.0 ± 14.9 years; 57.9% were male, 94.0% had AVF, 55.3% received conventional HD, and 32.9% were diabetic. There were 518 first thrombectomies and 302 competing events. The top-10 consensus predictors were access category, anticoagulant use, Kt/V, hemoglobin, albumin, mean arterial pressure, effective blood flow rate, phosphorus, potassium, and corrected calcium. The top-10 model showed modest OOB discrimination, with AUC 0.64 at 1 year and 0.61 at 2 years, but excellent calibration: predicted vs observed risk was 7.2% vs 7.6% at 1 year and 12.6% vs 14.3% at 2 years; intercepts were near zero and slopes were 0.99. DCA showed greater net benefit than treat-all/treat-none across thresholds of 2.6%–30.6% at 1 year and 2.2%–31.2% at 2 years (Figure 1)
Conclusion
The model identified an interpretable thrombectomy-risk signature spanning access category, anticoagulation, adequacy, anemia, nutrition, hemodynamics, and mineral-bone status. Despite modest discrimination, the strong calibration and positive net benefit support clinical risk stratification. External validation is warranted
Figure 1: Decision curve analysis for the top-10 consensus CR-RSF model at 1 and 2 years