Membranous nephropathy thrombosis risk model links pathology with early trajectory shifts
Patients with membranous nephropathy face more than proteinuria and edema. The disorder carries a clinically meaningful risk of arterial and venous thrombosis, and the danger extends well beyond the day of diagnosis. In this single-center retrospective cohort, investigators built a machine learning model to predict thrombotic events from 4 weeks after discharge through 1 year after diagnosis by combining clinical variables, laboratory results, renal pathology, and early longitudinal changes in blood counts and albumin.
The approach matters because thrombosis in membranous nephropathy is not driven by a single lab value. Heavy proteinuria, hypoalbuminemia, impaired kidney function, inflammation, endothelial injury, and prior vascular disease can converge on a hypercoagulable state. A model that only samples one moment in time risks missing the direction of travel: a falling albumin trajectory, a rising inflammatory profile, or persistent nephrotic-range protein loss may carry more weight than a static measurement taken on admission.
One in seven patients developed thrombosis during follow-up
The cohort included 1,070 patients, and thrombotic events occurred in 150, or 14.0%. The composite outcome included myocardial infarction, ischemic stroke, deep vein thrombosis, pulmonary embolism, and renal vein thrombosis, all objectively confirmed by imaging. Follow-up began 4 weeks after discharge, which means the analysis targeted events that emerged after the index hospitalization rather than complications already present at diagnosis.
Among the candidate predictors, LASSO regression retained 10 variables: 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. The final feature set is clinically coherent. Prior vascular disease and hypertension point to baseline arterial vulnerability, while D-dimer, proteinuria, kidney dysfunction, and albumin depletion reflect the prothrombotic biology of nephrotic syndrome. Renal pathology contributed an additional layer of risk by capturing chronic structural injury that a routine chemistry panel cannot see.
The longitudinal signal was especially important. Neutrophil and albumin trajectories summarized how inflammatory and nutritional status evolved during the index hospitalization and in the first month after discharge. In practical terms, this captures whether a patient was stabilizing or continuing to drift into a more thrombogenic state.
The kknn model balanced discrimination and calibration better than the others
Seven machine learning algorithms were compared: logistic boosting, adaptive boosting, generalized linear modeling, weighted k-nearest neighbors, naive Bayes, random forest, and support vector machine. Among them, the weighted k-nearest neighbors, or kknn, model showed the most balanced performance between the training and validation sets.
In the full model, validation AUC reached 0.810, with a 95% confidence interval of 0.725 to 0.896. Sensitivity was 0.800, specificity 0.663, and F1 score 0.414. The AUC suggests the model separated higher-risk from lower-risk patients reasonably well, while the sensitivity indicates it caught most patients who later developed thrombosis. Specificity was more modest, which is often the trade-off when the goal is not to miss high-risk patients who might need closer surveillance or individualized antithrombotic planning.
A single sentence captures the bedside consequence: a model with this profile is better suited to triage than to replace clinical judgment.
Pathology and early trajectories added information beyond single-point testing
The value of the study lies in what it layered together. Clinical scores built only from admission data can miss the biology unfolding over the first weeks after diagnosis, when proteinuria, inflammation, and albumin levels may shift quickly. Latent class trajectory modeling allowed repeated measurements to be translated into clinically readable patterns rather than raw serial values, and those patterns were then incorporated into prediction.
Renal pathology also strengthened the model. Interstitial fibrosis and tubular atrophy represent chronic tissue damage, not just active glomerular leakage. In a disease where thrombosis risk rises alongside nephrotic severity, this chronicity may signal patients whose renal injury is already advanced enough to amplify vascular risk. Combined with eGFR, cystatin C, and 24-hour urinary protein excretion, the pathology variable helped anchor the model in organ-level damage rather than laboratory noise.
Why this matters for antithrombotic decisions in membranous nephropathy
Current prophylactic anticoagulation decisions in membranous nephropathy often lean heavily on serum albumin and bleeding risk. That strategy is understandable, but it is incomplete. Albumin is important because it tracks the severity of nephrotic syndrome and the loss of natural anticoagulant proteins, yet thrombosis risk also depends on prior cardiovascular disease, inflammatory burden, kidney function, and structural renal injury. The present model attempts to bring those dimensions into one framework.
The biological rationale is straightforward. Nephrotic-range proteinuria promotes urinary loss of anticoagulant proteins, the liver responds by increasing synthesis of coagulation factors, platelets may become more reactive, and endothelial dysfunction can persist. When albumin remains low and inflammatory markers stay elevated, the coagulation system is not simply activated; it is primed. A prediction tool that captures that evolving state could help identify patients who need closer follow-up, earlier discussion of prophylactic anticoagulation, or more aggressive control of the underlying nephrotic syndrome.
Limits of a single-center retrospective model remain real
Several features constrain how far the findings can be applied. The study was retrospective and single-center, so performance may change in other hospitals, in different ethnic populations, or in centers with different biopsy practices and anticoagulation thresholds. The cohort also relied on internal validation rather than external validation, which leaves the possibility that real-world accuracy will fall outside the original dataset. Trajectory classes were built from measurements collected during the index hospitalization and at 2 and 4 weeks after discharge, so the model captures early change but not later disease drift.
The event rate also deserves attention. A composite outcome of arterial and venous thrombosis is clinically useful, but it can combine distinct mechanisms under one label. Myocardial infarction, ischemic stroke, and venous thromboembolism do not arise through identical pathways, even in nephrotic syndrome. Future validation will need to determine whether the same predictors carry equal weight across arterial and venous events, or whether separate models perform better.
Still, the central message is concrete: thrombosis risk in membranous nephropathy can be estimated more accurately when renal pathology and early longitudinal behavior are treated as part of the disease, not as optional extras. The next step is not a broader slogan about precision medicine; it is external validation in independent cohorts and testing whether this model changes prophylactic anticoagulation decisions without increasing bleeding harm.