Abstract: Ｏｂｊｅｃｔｉｖｅ：To develop a personalized dynamically model to predict the risk of kidney failure for patients with IgA nephropathy (IgAN) using updates of longitudinal data at each follow-up visit.
Ｍｅｔｈｏｄｏｌｏｇｙ：Three joint models were fitted to analyze the longitudinal data at each visit. We defined the baseline as the time of the kidney biopsy and fitted a clinical joint Model A, which included variables such as sex, age, eGFR, ALB, and proteinuria. We also constructed a clinical-pathological Model B, which incorporated both clinical and histological MEST features. Model C was fitted using parameters identical to those in Model A, however, the baseline was defined as the time of the clinical visit instead of the biopsy.
Ｒｅｓｕｌｔｓ：A total of 866 patients were included (650 in the development cohort and 216 in the validation cohort) and contributed 10 565 patient-years of data, and 22 533 eGFR and proteinuria measurements. Models A and B performed similarly with high predictive ability. However, the inclusion or exclusion of pathological variables did not significantly increase or decrease the accuracy of the joint models for predicting kidney failure risk. As follow-up time increased, model A's predictive performance continued to improve, reaching optimal performance around 5 years after kidney biopsy. The AUC value increased from 0.864 at kidney biopsy to 0.956 at 5 years after biopsy, and the Brier score decreased from 0.124 at the time of biopsy to 0.058 at 5 years after biopsy. The model C achieved similar results. All predictive performances were confirmed in the validation cohort. To facilitate clinical practice, we utilized the Shiny package to implement dynamic prediction. The R objects and source code have been made publicly available online.
Ｃｏｎｃｌｕｓｉｏｎ：The Joint model can be utilized for the longitudinal data generated from visits of IgAN patients, providing a high-performance dynamic prediction of kidney failure for IgAN patients. This helps achieve the objective of individualized dynamic prediction.