Abstract: PO0525
Identifying and Clustering CKD Progression Trajectories Using Machine Learning
Session Information
- CKD Health Services Research
October 22, 2020 | Location: On-Demand
Abstract Time: 10:00 AM - 12:00 PM
Category: CKD (Non-Dialysis)
- 2101 CKD (Non-Dialysis): Epidemiology, Risk Factors, and Prevention
Authors
- Abdul Sultan, Alyshah, AstraZeneca, Cambridge, United Kingdom
- Rhodes, Kirsty, AstraZeneca, Cambridge, United Kingdom
- Doulis, Michail, AstraZeneca, Gothenburg, Sweden
- Brookes-Smith, Irena, AstraZeneca, Cambridge, United Kingdom
- Faria, Jolyon S., AstraZeneca, Cambridge, United Kingdom
- Salazar, Jose Domingo, AstraZeneca, Cambridge, United Kingdom
- James, Glen, AstraZeneca, Cambridge, United Kingdom
- MacPhee, Iain, AstraZeneca, Cambridge, United Kingdom
- Unwin, Robert J., AstraZeneca, Cambridge, United Kingdom
- Wright, David, AstraZeneca, Cambridge, United Kingdom
- Patel, Mishal, AstraZeneca, Cambridge, United Kingdom
- Metcalfe, Paul D., AstraZeneca, Cambridge, United Kingdom
- Jermutus, Lutz, AstraZeneca, Cambridge, United Kingdom
Background
There is evidence suggesting that estimated glomerular filtration rate (eGFR) slope can be used as a surrogate clinical endpoint in renal clinical trials. However, there are limited data on the characteristics of fast and slow progressors based on eGFR slope from large population-based studies.
Methods
We identified CKD patients (based on two consecutive eGFRs of <75ml/min/1.73m2 recorded more than 90 days apart) aged ≥18 years from the UK Clinical Practice Research Datalink (CPRD) between 2004 and 2019. Estimated GFR measurements over a 3-year observation period post-index date (date of 2nd eGFR measurement) were extracted. Patients were clustered based on their eGFR trajectories using statistical (linear mixed effect models (LMM)) and machine learning techniques (unsupervised machine learning and Bayesian approaches). Association between trajectory clusters and all-cause mortality was assessed using Cox regression analysis.
Results
Preliminarily, 407,108 patients with 1.8 million eGFR measurement (median 4 (IQR: 2-6) eGFR measurements per patient) were identified. Using LMM, we found 5% of patients declined rapidly with an average rate of eGFR change per year -4.78 (95%CI: -9.40 to -3.28) whereas the majority (95%) remain stable or progressed slowly. A distinct fast progressing cluster was also detected using unsupervised machine learning and Bayesian methods which showed broadly linear patterns. Overall, there was an agreement between all three clustering approaches. These findings were replicated in the validation dataset showing consistent findings. Compared to stable/slow progressors, fast progressors were 3 times more likely (Hazard Ratio (HR)=2.82: 95%CI 2.75-2.90) to die following the 3-year observation period.
Conclusion
A clear fast progressing cluster was identified with an average eGFR decline of ≥5 ml/min/1.73m2 per year with a higher risk of all-cause mortality compared to other clusters. Whilst Bayesian and unsupervised machine learning methods can detect non-linear patterns, we found broadly linear trajectories.
Funding
- Commercial Support –