Kidney Week

Abstract: FR-PO1085

Artificial Intelligence (AI)-Assisted Urgent-Start Peritoneal Dialysis (PD) Prescription Engine: A Clinical Decision-Support Framework for Safer Dialysis Initiation

Session Information

• Home Dialysis: Modalities, Barriers, Utilization, and Clinical Outcomes
  October 23, 2026 | Location: Exhibit Hall A, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Dialysis

• 802 Dialysis: Home Dialysis and Peritoneal Dialysis

Authors

• Velusamy, Gandhimathi, University of Houston System, Houston, Texas, United States

Gandhimathi Velusamy, PhD

• Campbell, Lori, Vantive US Healthcare LLC, Deerfield, Illinois, United States

Lori Campbell

• Flores, Jackeline, Texas Tech University Health Sciences Center School of Medicine, Lubbock, Texas, United States

Jackeline Flores, MD

• Flowers, Addie, Texas Tech University Health Sciences Center School of Medicine, Lubbock, Texas, United States

Addie Flowers, MD

• Natarajan, Piruthiviraj, Texas Tech University Health Sciences Center School of Medicine, Lubbock, Texas, United States

Piruthiviraj Natarajan, MD

• George, Lekha K., Texas Tech University Health Sciences Center School of Medicine, Lubbock, Texas, United States

Lekha K. George, MD, FASN

Background

Many patients present late with ESKD without prior dialysis planning, and the default pathway is urgent HD using temporary CVC, which is associated with infection risk, repeated procedures, higher costs, hospitalisation, and morbidity. Urgent-start PD (USPD) avoids temporary HD catheters and supports transition to home dialysis; however, adoption remains limited by a lack of standardised protocols and the absence of real-time decision support during the high-risk post-catheter period.

Methods

We developed an AI-assisted, clinician-in-the-loop framework for urgent-start PD prescription planning. Clinical inputs included BSA, RKF, edema, SOB, comorbidities, and procedural factors. A hybrid prescription prediction engine combining machine learning to predict session time, fill volume, cycles per session, and sessions per week, while assigning dextrose concentration using edema/SOB-based rules. A supine positioning with low fill volumes (500–1000 mL), and short dwell times was incorporated to reduce early leaks and mechanical complications. The model was trained on a clinician-curated 74-synthetic dataset using a multi-output Random Forest classifier with cross-validation. SHAP-based interpretation was used to support explainability, and clinicians reviewed all generated prescriptions prior to implementation

Results

The multi-output Random Forest model achieved an exact match accuracy of 73% for the simultaneous prediction of all prescription parameters. Accuracy was highest for session time and cycles per session (87%), with strong performance for fill volume (87%) and sessions per week (80%). The framework standardised USPD prescription planning while preserving clinician oversight and supporting conservative, safety-focused prescribing. Observational studies demonstrating comparable or lower mortality with USPD vs catheter-based HD further support the feasibility.

Conclusion

An AI-assisted, explainable prescription prediction engine may support safer USPD initiation and reduce reliance on temporary HD catheters. By integrating predictive analytics with a recumbent, low-volume initiation strategy, this framework may improve workflow efficiency, reduce catheter-related harm, and expand access to home dialysis, addressing clinicians' pain points. Prospective validation using real-world patient data is needed.

Acknowledgment

GRACE GLOBAL

Funding

• Commercial Support – Business-GRACE Global