Abstract: FR-PO387
Artificial Intelligence Improves CKD Progression Forecasts
Session Information
- CKD: Risk Factors for Incidence and Progression - I
November 03, 2017 | Location: Hall H, Morial Convention Center
Abstract Time: 10:00 AM - 10:00 AM
Category: Chronic Kidney Disease (Non-Dialysis)
- 301 CKD: Risk Factors for Incidence and Progression
Authors
- Neri, Luca, Fresenius Medical Care, Bad Homburg, Germany
- Bellocchio, Francesco, Fresenius Medical Care, Bad Homburg, Germany
- Barbieri, Carlo, Fresenius Medical Care, Bad Homburg, Germany
- Mari, Flavio, Fresenius Medical Care, Bad Homburg, Germany
- Tschulena, Ulrich, Fresenius Medical Care, Bad Homburg, Germany
- Stuard, Stefano, Fresenius Medical Care, Bad Homburg, Germany
Background
Accurate CKD progression forecasts are key to tailor interventions to real patients needs. Current prognostic tools rely on few dominant variables (e.g. eGFR, albuminuria) and do not incorporate potentially important patterns of association (e.g. interactions). Hence, their performance is suboptimal for high and low risk patients (i.e. early stages of CKD). We developed a risk score addressing such weaknesses with machine learning.
Methods
CKD patients (stages 3-5) registered in EuCliD® (2011-2015) entered the study cohort. Enpoints were Kidney Failure (KF) within 2 and 5 years. The algorithm was derived with random forest (RF) and tested in a partition of the original sample which was not used to develop the model. The RF algorithm exploited 82 variables. To reflect real-life clinical practice, in the validation study, the algorithm calculated a risk score for each patient based on non-missing variables abstracted from electronic clinical charts. We computed model AUC and calibration curves. We compared the predictive accuracy of RF against Tangri's Kidney Failure Risk Score (KFRS)
Results
Among 4064 patients, 2685 and 1672 patients had 2 and 5 years of follow up (fig. 1). Most influential variables for KF at 2 years were eGFR, its rate of change, proteinuria, body mass, haemoglobin, Charlson's index, phosphate. Most influential variables for KF at 5 years were blood pressure, CKD-BMD markers, eGFR, proteinuria, serum albumin, heart rate. FR outperformed KFRS both in the whole sample and among CKD3a patients (Fig 1).
Conclusion
Contrary to KFRS, the performance of RF was stable at different CKD stages and was less dependent on intial GFR and albuminuria. Differences in forecast accuracy between RF and KFRS equations may lead to very large reductions in health care cost and clinical risk due to unnecessary medical encounters.