Abstract: FR-PO048
External Validation of a Machine Learning Model for Progression of CKD in the CREDENCE and CANVAS Trials
Session Information
- AI, Digital Health, Data Science - II
November 03, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Authors
- Tangri, Navdeep, University of Manitoba Max Rady College of Medicine, Winnipeg, Manitoba, Canada
- Ferguson, Thomas W., University of Manitoba Max Rady College of Medicine, Winnipeg, Manitoba, Canada
- Bamforth, Ryan J., University of Manitoba Max Rady College of Medicine, Winnipeg, Manitoba, Canada
- Arnott, Clare Gabrielle, The George Institute for Global Health, Newtown, New South Wales, Australia
- Mahaffey, Kenneth W., Stanford Medicine, Stanford, California, United States
- Heerspink, Hiddo Jan L., Universitair Medisch Centrum Groningen Afdeling Cardiologie, Groningen, Groningen, Netherlands
- Perkovic, Vlado, The George Institute for Global Health, Newtown, New South Wales, Australia
- Neuen, Brendon Lange, The George Institute for Global Health, Newtown, New South Wales, Australia
Background
Sodium glucose cotransporter 2 inhibitors (SGLT2i) are indicated for slowing progression of chronic kidney disease (CKD). A previously validated machine learning model (Klinrisk model) accurately predicts 40% decline in eGFR or kidney failure using routinely collected laboratory data. We sought to validate this model in the pooled CANVAS/CREDENCE trials.
Methods
The CANVAS/CREDENCE trials evaluated the effects of the SGLT2i canagliflozin on cardiorenal outcomes in patients with type 2 diabetes at high cardiovascular risk or with CKD. We validated the Klinrisk model for prediction of CKD progression, defined as greater than 40% decline in eGFR or kidney failure. The model applies results from complete blood cell counts, chemistry panels, comprehensive metabolic panels, and urinalysis. Model performance was assessed up to 3 years (median follow up 2.4 years) with the area under the receiver characteristic operating curve (AUC), Brier scores, and calibration plots of observed and predicted risks. We compared performance of the model to standard of care using eGFR (G1-G4) and urine ACR (A1-A3) KDIGO heatmap categories.
Results
Among 14,464 patients in CANVAS/CREDENCE, we found the Klinrisk model provided excellent discrimination for CKD progression (696 events at 2 years), with an AUC of 0.81 (95% confidence interval 0.78 – 0.83) for prediction of the outcome at 1 year, increasing to 0.88 (0.86 – 0.89) at 3 years. Brier scores were 0.020 (0.018 – 0.022) at 1 year, increasing to 0.056 (0.052 – 0.059) at 3 years. Calibration was satisfactory, with minor overprediction in patients randomized to canagliflozin. Compared to the KDIGO heatmap, the Klinrisk model had improved performance at every interval (Table 1).
Conclusion
The Klinrisk machine learning model using routinely collected laboratory features was highly accurate in its prediction of CKD progression in the CANVAS and CREDENCE trials.
Table 1. Results of model performance
| Klinrisk model | Klinrisk model | eGFR and ACR categories (KDIGO heatmap) | eGFR and ACR categories (KDIGO heatmap) | |
| Time frame, years | AUC (95% CI) | Brier score (95% CI) | AUC (95% CI) | Brier score (95% CI) |
| 1 | 0.81 (0.78 - 0.83) | 0.020 (0.018 - 0.022) | 0.74 (0.71 – 0.76) | 0.021 (0.018 – 0.023) |
| 2 | 0.85 (0.84 - 0.87) | 0.042 (0.039 - 0.046) | 0.79 (0.78 – 0.81) | 0.046 (0.042 – 0.050) |
| 3 | 0.88 (0.86 - 0.89) | 0.056 (0.052 - 0.059) | 0.83 (0.81 – 0.84) | 0.063 (0.059 – 0.067) |
Brier scores range from 0 to 1, with lower values representing higher accuracy.