Kidney Week

Abstract: FR-PO1098

Validation of Machine Learning-Based Screening Tools for Early Detection of CKD in Patients with Type 2 Diabetes (T2D)

Session Information

• CKD: Epidemiology, Risk Factors, and Prevention - 2
  October 25, 2024 | Location: Exhibit Hall, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: CKD (Non-Dialysis)

• 2301 CKD (Non-Dialysis): Epidemiology, Risk Factors, and Prevention

Authors

• Martinez, Julian, Arkangel AI, Bogota, Colombia

Julian Martinez

• Perez-Rondon, Alejandra, Arkangel AI, Bogota, Colombia

Alejandra Perez-Rondon

• Zea, Jose, Arkangel AI, Bogota, Colombia

Jose Zea

• Ilano, Isabella, Arkangel AI, Bogota, Colombia

Isabella Ilano

• Castano-Villegas, Natalia, Arkangel AI, Bogota, Colombia

Natalia Castano-Villegas

• Nguyen Cong, Luong, AstraZeneca, Ho Chi Minh, Viet Nam

Luong Nguyen Cong, MD

• Hadaoui, Ahmed, AstraZeneca, Algiers, Algeria

Ahmed Hadaoui, PharmD

Background

With the increasing burden of CKD among pts with T2D, early detection is critical. Annual screening is challenging, especially in developing nations. We developed 2 minimal resource models to identify pts at high risk of CKD based on estimated glomerular filtration rate (eGFR) alterations. Our study validated the models on iCaReMe global registry T2D pts.

Methods

Arkangel AI^TM software was used to develop 2 algorithms. Machine-learning architectures were used for training and ranking the algorithms. An ensemble learning model was built from the 2 algorithms. A combination of predictions from these algorithms decides whether the pt is positive or negative for CKD. These algorithms were applied to an observational iCaReMe database spanning 6 low-middle-income countries to generate an eGFR-based model. Demographic data including age, duration of T2D, sex, body mass index, and systolic and diastolic blood pressure were extracted for pts with T2D. Pts with eGFR <60 mL/min/1.73m² were considered positive cases and those with eGFR ≥60 mL/min/1.73m² were negative cases. The study determined the accuracy, sensitivity, specificity, positive predictive values, area under the curve (AUC), and F1 scores of the model.

Results

A total of 4342 pts with validated eGFR measurements were included in training the models. The ensemble learning model reported 85.19% (95% confidence interval [CI]: 82.82 to 87.55) sensitivity, 39.97% (95% CI: 38.34 to 41.59) specificity, and 48.96% (95% CI: 47.48 to 50.45) accuracy. Furthermore, the model has shown a positive predictive value of 26.06% (95% CI: 24.44 to 27.68) indicating a moderate proportion of true positive cases with CKD. The F1 score of the model was 39.91%, presenting a good predictive performance. The AUC was 62.58% showing that the model can distinguish between pts with and without CKD.

Conclusion

The model reported a good sensitivity for identifying CKD risk in the diabetic population with minimal pt information. They can help bridge the unmet need for screening for CKD among at-risk populations in resource-limited, real-world settings.

Acknowledgment: Medical writing support - Priyanka Grandhi (Fortrea Scientific Pvt. Ltd)

Funding

• Commercial Support – AstraZeneca International