Kidney Week

Abstract: FR-PO1192

Cluster-Guided Models for Patient-Specific Subclinical Rejection Surveillance

Session Information

• Transplantation: Clinical - Rejection, Biomarkers, and Pharmacology
  October 23, 2026 | Location: Exhibit Hall A, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Transplantation

• 2002 Transplantation: Clinical

Authors

• Singla, Akhil, Northwestern University, Evanston, Illinois, United States

Akhil Singla, PhD

• Chen, Kenny, Northwestern University Feinberg School of Medicine, Chicago, Illinois, United States

Kenny Chen, MS

• Rebello, Christabel, Northwestern University Feinberg School of Medicine, Chicago, Illinois, United States

Christabel Rebello, MS

• Zhao, Lihui, Northwestern University Feinberg School of Medicine, Chicago, Illinois, United States

Lihui Zhao

• Park, Sookhyeon, Northwestern University Feinberg School of Medicine, Chicago, Illinois, United States

Sookhyeon Park, MD, MS

• Mehrotra, Sanjay, Northwestern University, Evanston, Illinois, United States

Sanjay Mehrotra, PhD

• Friedewald, John J., Northwestern University Feinberg School of Medicine, Chicago, Illinois, United States

John J. Friedewald, MD, FASN

Background

Subclinical acute rejection (subAR) surveillance is biopsy-dependent. Global standard-of-care (SOC)+biomarker logistic regression (LR) uses one coefficient set despite variation in donor, sensitization, and graft-function context. We tested whether patient-level SOC clustering could guide cluster-specific biomarker LR.

Methods

Discovery used 460 biopsy-paired CTOT-08 () samples from 226 recipients; validation used 128 samples from 118 Mini-Kidney recipients. Borderline biopsies were grouped with no rejection. Benchmark LR used SOC plus CXCL9/creatinine (Cr), CXCL10/Cr, donor-derived cell-free DNA (dd-cfDNA), gene expression profile (GEP), and log10 torque teno virus (TTV). Patient-level SOC clusters were learned by partitioning around medoids/Gower distance. Cluster-specific biomarker LRs were used for discovery, and Youden thresholds were applied to validation. Stability used silhouette and adjusted Rand index resampling; SOC+biomarker clusters were tested for robustness.

Results

Benchmark validation AUC was 0.69 (95% confidence interval [CI] 0.51-0.84), with sensitivity/specificity/positive/negative predictive value (PPV/NPV) 0.67/0.67/0.21/0.94. Patient SOC k=2 separated profiles by donor type, donor-specific antibody, panel reactive antibody (PRA) class I, and log10 TTV; Cluster 1 had more deceased donors (80.8% vs 26.1%) and PRA I 51-100 (30.8% vs 8.7%). Cluster 1 LR used CXCL10/Cr+dd-cfDNA+GEP+log10 TTV at 0.25; Cluster 2 used CXCL9/Cr+dd-cfDNA+GEP+log10 TTV at 0.10 probability thresholds. Validation patient SOC k=2 AUC was 0.72 (0.56-0.84), with sensitivity/specificity/PPV/NPV 0.73/0.52/0.17/0.94. Robustness checks showed validation patient SOC k=3 AUC 0.75 (0.58-0.89) and SOC+biomarker k=2 AUC 0.73 (0.58-0.86).

Conclusion

Patient-level clustering revealed interpretable recipient heterogeneity absent from a global biomarker model. Varying biomarker weights across stable profiles supports individualized surveillance with high rule-out performance.

Figure 1. Patient-level SOC clustering workflow (k=3).

Funding

• NIDDK Support