AUTHORS
Shankara Krishna Anand, MD, MS, Andrew Staron, MD2, Natasha J Burke3, Vaishali Sanchorawala, MD4, Ashish Verma, MBBS5
AFFILIATIONS
1 Stanford Health Care, Stanford, California, United States
2 Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
3 Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
4 Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
5 Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
ABSTRACT
Introduction: Renal involvement occurs in up to 70% of patients with immunoglobulin light-chain (AL) amyloidosis. Progression to end-stage kidney disease (ESKD) impacts prognosis, dialysis requirement, and treatment selection. The Palladini renal staging system (Blood 2014) stratifies risk of progression to ESKD into three stages using estimated glomerular filtration rate (eGFR < 50 mL/min/1.73 m²) and 24-hour urine total protein (UTP > 5 g/24h). However, it was derived using Kaplan-Meier estimation, which censors death and overestimates ESKD risk in a disease with substantial competing (largely cardiac) mortality. It also provides group-level rather than patient-specific probability. We developed and internally validated an individualized, competing risks model for ESKD in AL amyloidosis.
Methods: We queried the Boston University Amyloidosis Center databank for patients with newly diagnosed AL amyloidosis, renal involvement, and measurable proteinuria (24-hour UTP > 0) between 1994 and 2023 (n=1,262). ESKD (dialysis or kidney transplant) was the event of interest and death without ESKD the competing event, modeled using Fine-Gray subdistribution hazards. Missing covariates were handled by multiple imputation (MICE, m=20). Predictors were selected by LASSO (eGFR and log-transformed 24h UTP), with age retained as a prespecified clinical covariate. At 1, 3, and 5 years, we compared our model with Palladini staging on discrimination (time-dependent area under the curve [AUC]), overall accuracy (index of prediction accuracy [IPA], a scaled Brier score), calibration (integrated calibration index [ICI]), and clinical utility (decision-curve analysis); AUC and ICI were internally validated by bootstrap cross-validation (out-of-bag, 500 replications), IPA and decision-curve analysis by repeated 10-fold cross-validation.
Results: Over a median follow-up of 10.3 years, 218 patients (17%) progressed to ESKD while 685 (54%) died without ESKD, reflecting substantial competing mortality. Utilizing competing risk (Aalen-Johansen) estimation, no stage I patient developed ESKD at any time point. The 1-, 3-, and 5-year incidence was 6%, 12%, and 16% for Palladini stage II and 15%, 28%, and 33% for Palladini stage III. When death was censored by Kaplan-Meier, the 3-year stage III estimates rose from 28% to 35%, illustrating the upward bias that likely explains the higher 3-year risks originally reported in the Palladini study (0%, 7%, 60% for stages I-III).
In our individualized model, lower eGFR (subdistribution hazard ratio [sHR] 0.98 per mL/min/1.73 m²), higher UTP (sHR 1.40 per log-unit), and younger age (sHR 0.97 per year) were each independently associated with progression to ESKD (all p<0.001). The inverse association with age reflects competing risk: younger patients are less likely to die before reaching ESKD. Internally validated discrimination exceeded Palladini staging at 1 year (time-dependent AUC 0.82 vs. 0.76; difference 0.07, 95% CI 0.01-0.13, p=0.03) and remained higher at 3 years (0.79 vs. 0.76). The model demonstrated improved accuracy, reducing squared prediction error by 13% relative to a covariate-free model vs. 9% for Palladini staging (index of prediction accuracy 0.13 vs. 0.09 at 3 years, p=0.02). This model was calibrated at the patient level: predicted 3-year risks were within 1.6 percentage points of observed on average, compared to 11.5 points for Palladini staging (ICI 0.016 vs. 0.115).
In decision-curve analysis, our model yielded higher net benefit than Palladini staging across 3-year risk thresholds of ~10-30%. At a 10% threshold, applying the individualized model would avoid ~14 unnecessary high-risk evaluations/referrals per 100 patients without missing any true ESKD progressors. The model further resolved risk heterogeneity within each Palladini stage. Among stage II patients, 3-year ESKD incidence was 15% in the upper half of model-predicted risk vs. 8% in the lower (Gray p=0.01); among stage III patients, 36% vs. 19% (p<0.001).
Conclusion: In renal AL amyloidosis with measurable proteinuria, a parsimonious, competing-risks model using eGFR, proteinuria, and age provides individualized, well-calibrated absolute estimates of ESKD risk and delivers greater clinical net benefit than categorical staging. Because roughly half of these patients die before reaching ESKD, competing risks estimation gives a more realistic picture of what each patient faces.
DISCLOSURES
Shankara Krishna Anand, MD, MS: No relevant conflicts of interest to declare.