Machine learning prediction of 1-year thrombosis risk in membranous nephropathy: integration of renal pathology and early longitudinal trajectories

Introduction Patients with membranous nephropathy (MN) have an increased risk of thrombotic events. Existing prediction approaches primarily use clinical and laboratory measurements obtained at a single time point and rarely incorporate renal pathology or early longitudinal changes. This study integrated multiple types of information to predict the occurrence of thrombosis from 4 weeks after discharge to 1 year after being diagnosed with MN. Method This single-center retrospective cohort study included 1,070 patients with MN. Patient characteristics, laboratory findings, renal pathological features, and follow-up data were collected. Patients were randomly divided into training and validation sets at an 8:2 ratio. Latent class trajectory modeling identified longitudinal classes from measurements obtained during the index hospitalization for MN diagnosis and at 2 and 4 weeks after discharge. Predictors were selected using LASSO regression. Seven machine learning algorithms were used, including logistic boosting, adaptive boosting, generalized linear model, weighted k-nearest neighbors (kknn), naive Bayes, random forest, and support vector machine. Result Thrombotic events occurred in 150 patients (14.0%). LASSO selected 10 predictors: hypertension history, stroke or myocardial infarction history, erythrocyte sedimentation rate, D-dimer, cystatin C, estimated glomerular filtration rate, 24-hour urinary protein excretion, interstitial fibrosis and tubular atrophy score, neutrophil trajectory class, and serum albumin trajectory class. Among all models, kknn showed relatively balanced performance between the training and validation sets. The full kknn model achieved a validation AUC of 0.810 (95% CI: 0.725–0.896), sensitivity of 0.800, specificity of 0.663, and F1 score of 0.414. Conclusion Integrating clinical, laboratory, renal pathological, and early longitudinal trajectory information enabled prediction of 1-year thrombotic events in MN. The kknn model may support early risk stratification and individualized antithrombotic strategy.