Artificial Intelligence Model for Renal Allograft Rejection Identification, Lesion Quantification and Rejection Risk Prediction Based on Prospective Renal Transplant Pathology and Clinical Multidimensional Data

Condition(s) studied

Study summary

This prospective research collects leftover kidney biopsy tissue slides and matching routine clinical data from patients who received kidney transplants and underwent standard kidney puncture biopsy at Zhejiang University School of Medicine First Affiliated Hospital starting November 2025. A total of around 1,000 patient samples will be included, covering transplant rejection (including TCMR and ABMR subtypes, acute and chronic rejection), polyomavirus infection and recurrent original kidney disease after transplantation.

All study materials come from residual biopsy specimens generated during regular clinical examinations, with no extra invasive operations, additional medical costs or physical trauma for participants. We will scan pathological slides into digital images and combine them with patients' medical records, lab test results, medication history and follow-up information. After full anonymization and standardized labeling by senior renal pathologists following the Banff standard, we will build an artificial intelligence (AI) multi-task model.

This AI system will serve three core clinical functions: accurately distinguish different types of transplant kidney lesions, quantitatively measure tissue damage caused by rejection, and predict the risk of recurrent rejection after surgery. We will optimize and verify the model's diagnostic accuracy, stability and reliability through dataset segmentation, cross validation and algorithm adjustment.

For patients, this study brings no extra physical or economic burden. If suspicious pathological changes are found during data analysis, relevant clues will be fed back to attending doctors to support individual treatment management. For clinical providers, the finished AI tool can reduce pathologists' reading workload, lower missed diagnosis and misdiagnosis caused by individual experience differences, especially improve detection of subclinical and borderline rejection. It helps clinicians evaluate injury severity and forecast recurrence risk, so as to formulate personalized immunosuppression regimens, reduce rejection relapse and prolong graft survival.

Strict privacy protection measures are implemented throughout the whole research process: all personal identifiable information will be completely removed, and encrypted classified data management is adopted to prevent information leakage. Every participant signs a written informed consent and retains the right to withdraw from the study at any time without affecting their regular medical care. All research procedures have passed ethical review supervision, and all collected data and specimens will be properly stored or destroyed in accordance with standardized medical management rules after the study ends.

The research aims to fill the gap of prospective multi-dimensional AI auxiliary diagnosis research in kidney transplantation, promote standardized, intelligent and precise post-transplant pathological evaluation, and provide new technical support to improve long-term survival outcomes of kidney transplant recipients.

Eligibility

Primary outcome measure(s)

• ROC AUC of multi-task AI model for identification of renal allograft lesions — After model training and internal verification（through study completion, an average of 24 months）
  Area under receiver operating characteristic curve to evaluate model ability to detect rejection subtypes, polyomavirus nephropathy and recurrent primary renal disease.
• Sensitivity and specificity of multi-task AI model for renal allograft lesion diagnosis — After model training and internal verification（through study completion, an average of 24 months）
  Sensitivity and specificity of the AI model discriminating TCMR, ABMR, polyomavirus nephropathy and recurrent primary renal disease.
• Diagnostic accuracy of multi-task AI model for renal allograft lesions — After model training and internal verification（through study completion, an average of 24 months）
  Overall accuracy of the AI model in differentiating post-transplant renal pathological lesions.

Trial sites (1)

Other trials for the same condition