Abstract
Early diagnosis of light chain amyloidosis (AL) is critical to prevent irreversible organ damage and improve survival. However, delayed recognition is common due to the non-specific nature of the symptoms and signs of the disease. Many patients with systemic AL have a known precursor plasma cell disorder such as monoclonal gammopathy of undetermined significance or smoldering myeloma (collectively referred to here as MGUS+). To inform early detection strategies, we conducted a large population-based study to identify specific clinical diagnoses that precede an AL diagnosis in individuals with MGUS+.
In this NIH-funded study, we used Medicare health insurance claims data to select a longitudinal cohort of MGUS+ patients identified between 2016-2017. Patients were observed for up to 5 years prior to their MGUS+ diagnosis and followed for a minimum of 2 years afterward, until the earliest progression to AL, death, or end of the study period, whichever came first. The earliest 2 years were designated as the “prevalent” and the remaining follow-up time as the “incident” periods. Twenty-three pre-specified clinically relevant red flag diagnoses associated with AL (e.g., nephrotic syndrome, heart failure) were monitored via ICD-9-CM/ICD-10-CM codes contained in Part A (inpatient) and Part B (carrier and outpatient) claims. Diagnoses recorded during the prevalent period were not re-evaluated during the incident period; the timing of new (incident) diagnosis was captured based on its first occurrence during follow-up. To identify incident diagnoses and timing thereof associated with future AL, we applied nonparametric, time-to-event modeling using Bayesian Additive Regression Trees (BART), a machine learning method. BART is especially suitable for this study as a) it accomodates time-dependent covariates, capturing both the incidence and the timing of red flag diagnoses, b) identifies complex interactions without requiring pre-specification, therefore allowing combinations or clusters of red flags within and across organ systems, and c) provides interpretable rankings of variable importance based on selection probabilities within the ensemble. In addition to red flag diagnoses, the model also included sociodemographics (age, sex, dual enrollment to Medicaid as a proxy for poverty, and Census-tract level residence in urban, sub-urban or rural area).
The cohort consisted of 8,681 MGUS+ patients (median age 73, Q1-Q3 68-77 years; 50.4% male; 79.5% non-Hispanic white; 12.6% Black; 10% dual-enrolled in Medicaid; 8% rurality). Median follow-up was 73.8 (Q1-Q3 65.6-82) months. A total of 451 (5.2%) progressed to systemic AL during the post-MGUS+ follow-up period. The strongest predictors of AL risk indentified by BART multivariate modeling were incident nephrotic syndrome (importance score = 0.132), timing of nephrotic syndrome (0.063), and timing of proteinuria (0.059). Other high-ranking red flag predictors included incident cardiomyopathy (0.053), incident pleural effusion (0.037); prevalent cardiomyopathy (0.041), incident dysphagia (0.028), timing of heart failure (0.027), incident chronic kidney disease (0.025), and incident carpal tunnel syndrome (0.023). Additional relevant signals included weight loss (prevalent: 0.022; incident: 0.011), edema (timing: 0.019; prevalent: 0.014), and several multisystemic features such as syncope, arrhythmia, dyspnea, and hepatosplenomegaly (importance scores ranging from 0.011–0.015). Among sociodemographic factors, age was the strongest predictor of AL risk (0.129), followed by race (Black vs. white: 0.023), poverty status (0.018), and rurality (0.014). The BART survival model demonstrated good out-of-sample performance, with a Harrell's C-index statistic of 0.73, indicating strong model validity.
We identified several clinical diagnoses primarily involving renal and cardiac systems that preceded an AL diagnosis in a nationally representative cohort of older MGUS+ individuals. These findings demonstrate the utility of administrative claims data in flagging high-risk individuals and the potential of machine learning approaches for detecting complex, time-dependent patterns of disease evolution. Detailed results, to be discussed during the presentation, provide the foundation for predictive algorithms that can be integrated into electronic health systems to enable earlier detection of AL among older high-risk MGUS+ patients in clinical settings.
