Abstract: PO2340
A Machine Learning Algorithm to Identify Patients with Possible Non-Dialysis-Dependent CKD
Session Information
- Reassessing Race in Predicting Progression
November 04, 2021 | Location: On-Demand, Virtual Only
Abstract Time: 10:00 AM - 12:00 PM
Category: CKD (Non-Dialysis)
- 2101 CKD (Non-Dialysis): Epidemiology, Risk Factors, and Prevention
Authors
- Lobbedez, Thierry, University Hospital, Caen, France
- Dardim, Karim, Association Limousine pour l'Utilisation du Rein Artificiel à Domicile (ALURAD), Limoges, France
- Panes, Arnaud, HEVA, Lyon, France
- Poinsot, Gwendoline, HEVA, Lyon, France
- Fernandes, Jérôme, Centre Hospitalier de la Côte Basque, Bayonne, France
- Dubel, Laurence, Astellas Pharma Europe, Levallois-Perret, France
- Wolfram, Josephine, Astellas Pharma Europe B.V., Leiden, Netherlands
Background
The DAKOTAH study is a retrospective study of patients with non-dialysis-dependent chronic kidney disease (NDD CKD) in France based on data from the Echantillon Généraliste des Bénéficiaires database. A stepwise machine learning approach was used to identify patients with possible NDD CKD who could not be captured using the NDD CKD case definition (Figure).
Methods
First, the ‘potential CKD’ population was designated as patients with a diagnosis of diabetes, cardiovascular disease or hypertension, or with ≥3 prescriptions for antidiabetic and/or antihypertensive drugs, during 2012–2017. Second, an unsupervised algorithm was trained to identify patients very likely to have CKD (‘possible CKD’) in the potential CKD population. Similarity between patients was based on CKD-related variables: sex; number and duration of hospitalizations for renal failure; number of GP visits; medications; and biological exams. A distance metric between patients was defined based on these variables, and patients having similar characteristics were positioned close to one another. The algorithm learned to construct a spherical boundary around the non-CKD population, to create a decision rule for possible CKD versus non-CKD, with outliers considered possible CKD.
Results
The algorithm was validated by application to both the potential CKD population and a confirmed CKD patient pool. From the potential and confirmed CKD populations, 21% and 65% of patients were classified as possible CKD, respectively. Similarities were observed between the two groups regarding hospitalizations and selected biological exams.
Conclusion
This machine learning-derived decision rule could be a tool to identify undiagnosed patients with NDD CKD.
Funding
- Commercial Support –