Machine‐learning based subgroups of AL amyloidosis and cumulative incidence of mortality and end stage kidney disease

Abstract

AbstractImmunoglobulin light chain (AL) amyloidosis is a multisystem disease with varied treatment options and disease‐related outcomes. Current staging systems rely on a limited number of cardiac, renal, and plasma cell dyscrasia biomarkers. To improve prognostication for all‐cause mortality and end‐stage kidney disease (ESKD), we applied unsupervised machine learning using a comprehensive set of clinical and laboratory parameters. Our study cohort comprised 2067 patients with newly diagnosed, biopsy‐proven AL amyloidosis from the Boston University Amyloidosis Center. Variables included 31 clinical symptoms and 28 baseline laboratory values. Our clustering algorithm identified three subgroups of AL amyloidosis (low‐risk, intermediate‐risk, and high‐risk) with distinct clinical phenotypes and median overall survival (OS) estimates of 6.1, 3.7, and 1.2 years, respectively. The 10‐year adjusted cumulative incidences of all‐cause mortality were 66.8% (95% CI 63.4–70.1), 75.4% (95% CI 72.1–78.6), and 90.6% (95% CI 87.4–93.3) for low, intermediate, and high‐risk subgroups. The 10‐year adjusted cumulative incidences of end‐stage kidney disease (ESKD) were 20.4% (95% CI 6.1–24.5), 37.6% (95% CI 31.8–43.8), and 6.7% (95% CI 2.8–11.3) for low‐risk, intermediate‐risk, and high‐risk subgroups. Finally, we trained a classifier for external validation with high cross‐validation accuracy (85% [95% CI 83–86]) using a subset of easily obtainable clinical parameters. This marks an initial stride toward integrating precision medicine into risk stratification of AL amyloidosis for both all‐cause mortality and ESKD.