OBJECTIVES: Chronic renal failure (RF) is a major health problem affecting 700 million people worldwide. Patients are often unaware of their disease, due to a lack of early diagnosis. RWD can be used to identify kidney disease at an early stage and reduce its burden.

METHODS: RWD study using THIN® France database. Patients were included from 2013 to 2023 if they had a history ≥ 3 year, a RF diagnosis, and eGFR /albuminuria value. Patients in stages G5, G4, G3b-A(2-3), G3a-A3 of the KDIGO classification were considered at VHR of RF. Several ML algorithms were trained, tuned and validated in training set, and the best model was tested in the testing set to avoid overfitting. Over 300 variables related to medical history recorded at least one year prior to the event were included as predictors.

RESULTS: 4,976 patients with an average age of 72 and 12-year history were included. 20.3 % of them were at VHR of RF, 42% were women, 78% had hypertension, 42% diabetes and 39 % cardiovascular disease. Patients at high risk of RF had mean eGFR and albuminuria of 26 and 257 and 16% of them treated with analgesics. The AUC of the first algorithms tested are all over 0.7. The beset model, random forest, had an F-score and accuracy of 0.89. and 0.80. The most important variables observed were number of prescriptions of ATC class (C08, N02, J01), age, frequency of blood pressure recording, diabetes, unrecorded smoking status.

CONCLUSIONS: These results show that the combination of AI and RWD, including medical wandering, medical decisions and patient loyalty to their doctor, offers doctors a powerful additional tool for improving the early medical management of patients with kidney disease.