Abstract: FR-PO509
A Data-Driven, Evidence-Based, Patient-Centered Optimal Initiation Time for Dialysis Treatment
Session Information
- Dialysis Vascular Access
October 25, 2024 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 803 Dialysis: Vascular Access
Author
- Lee, Eva, The Data and Analytics Innovation Institute, Atlanta, Georgia, United States
Group or Team Name
- The Kaiser Nephrology Team.
Background
When to start dialysis is an important decision for end-stage CKD care. The difficulty in determining a common gold standard lies partly in the high heterogeneity among patients and available treatment conditions, and in random factors in disease progression / treatment process. This work optimizes pre-dialysis care timing.
Methods
A novel interpretable machine learning model ingests EMR / unstructured data to classify treatment effects and disease prognosis wrt each disease staging and timeline. Our ML model can classify disease-action into the same outcome groups under different conditions/features (hence multiple pathways).
A ML-knowledge treatment timing model is developed to determine an outcome-driven GFR threshold for creating access of various types that maximizes the goodness of patient’s health conditions. Patient preference is incorporated.
Results
11,913 CKD patients with 213,344 GFR s from 2011-2018 were extracted. Among those with initiated dialysis, 67% initiated HD, 32% PD, and 1% CVC. Current 'sub/optimal' initiation shows diverse variance. AI/ML uncovers critical features that affect outcome; and GFR-windows for dialysis initiation and begin dialysis that render good outcome. Table 1 shows 2 interpretable optimal timing and dialysis start time, w.r.t. patient preferences. Only 2-12% of current practice falls within these optimal initiation time. Simulating the 2 policies for 2 years on 1000 patients shows reduction of 16.4%-35% avoidable deaths and 8%-19% increase in utility reward. A new CPG for Phase I implementation was established.
Conclusion
This work has critical health practice implications. We establish the optimal initiative timing model for late-stage CKD management. The novel ML-decision system includes distinct analytic advances as it leverages hidden knowledge within EMR to enable evidence-based patient-centered policy making. The results show that current practice is far from optimal and can be improved considerably. Clinical trial has to be conducted to gauge potential outcome improvemet.
Two optimal timing policies returned from our model (simplified for brevity).
| Policy 1 | Policy 1 | Policy 2 | Policy 2 | |
| Type | Optimal Initiation time (pre-dialysis care) | Begin dialysis | Optimal Initiation time (pre-dialysis care) | Begin dialysis |
| HP | GFR < 20 | GFR < 12 | GFR < 24 | GRF < 15 |
| PD | GFR < 18 | GFR < 10 | GFR< 21 | GFR < 12 |
| Suboptimal | With uremia | GFR < 10 | With uremia | GFR < 12 |
| Suboptimal | Without uremia | GFR < 7 | Without uremia | GFR < 9 |
Funding
- Other U.S. Government Support