Abstract
Introduction: Cancer health disparities in multiple myeloma (MM) are well-established - African Americans (AAs) are at a higher risk for MM compared to Caucasian Americans (CAs). However, little is known about the risk factors driving this disparity. With the advancement in machine learning in the era of electronic health records, we aimed to assess racial differences in predictors for MM in patients diagnosed with monoclonal gammopathy of undetermined significance (MGUS), the premalignant condition of MM.
Methods: Patients diagnosed with MGUS from 1999-2021 in the Veterans Health Administration were identified. We used a natural language processing-based algorithm to determine MGUS diagnosis and progression to smoldering MM and MM. We excluded individuals of race other than CA or AA and individuals with diabetes mellitus (as diabetes, its underlying mechanism, and treatments may interact with MM progression). We considered the following potential predictors: age, sex, time-varying weight, height (for body mass index, BMI), m-spike, creatinine, hemoglobin, hemoglobin A1C, platelets, neutrophils, albumin, and total protein. For each patient who progressed, an observation window was defined as the time from the first to the last available time-varying data before progression. The length of the window was used to match cases to controls with ≥1 data point beyond the window to ensure no loss to follow-up. Each racial group was then split into training (70%) and testing (30%) sets. Ad-hoc features summarizing time-varying predictors/markers within the window included average, variance (var), maximum (max), minimum (min), max increase and decrease/month, and area under the curve for changes (AUTCC). The AUTCC integrates the long-term change in the predictor into one metric. To overcome the issue of imbalanced ratio of cases with progression to cases without (~1:9), we sampled the minority cases and create synthetic data points centered around them in the training set. Significant predictors were identified through an adapted Lasso regression method to recursively compute the importance of each predictor using the training data. Pearson correlation coefficients (r) were used to present the association between each predictor and progression. Model performance was evaluated by 10-fold cross validation random forest algorithms using the testing sets and presented as AUC scores.
Results: The analytic cohort included 16,430 (72.1% CAs and 27.9% AAs) MGUS patients. Average AUC score was 0.79 (95% confidence interval, CI, 0.76-0.82) for CAs and 0.83 (95% CI 0.81-0.86) for AAs. In AA patients with MGUS, the top risk factors for progression were max increase of m-spike (r=0.11, 95% CI 0.08-0.14), creatinine AUC (r=0.11, 95% CI 0.07-0.14), and max decrease of m-spike (r=0.06, 95% CI 0.07-0.08), while the top protective factors are max increase of creatinine/month (r=-0.32, 95% CI -0.40, -0.23), max albumin (r=-0.29, 95% CI -0.40, -0.20), and max decrease of creatinine/month (r =-0.18, 95% CI -0.25, -0.11). In CA patients with MGUS, the top risk factors for progression were platelets AUC (r=0.09, 95% CI 0.08-0.11), max decrease of platelets/month (r=0.08, 95% CI 0.05-0.10) and min platelets (r=0.05, 95% CI 0.02-0.07), while the top protective factors were var m-spike (r=-0.13, 95% CI -0.15, -0.11), var BMI (r=-0.12, 95% CI -0.15, -0.10), and min creatinine (r=-0.12, 95% CI -0.14, -0.1).
Conclusion: In both AAs and CAs, m-spike, platelets, and creatinine were the most influential predictors of progression from MGUS to MM. In addition to low m-spike for AA patients, stable creatinine and albumin levels predict reduced risk of progression. In CA patients, stable BMI, creatinine, and platelet levels predict reduced risk of progression. Our findings may contribute to improved prediction of progression for patients diagnosed with MGUS.
Disclosures
Sanfilippo:ACS-IRG: Research Funding; Covington & Burling LLP: Other: Expert Case Review; Health Services Advisory Group: Consultancy; K01 NHLBI: Research Funding; NHLBI NIH: Other: Loan Repayment program; Quinn Johnston: Other: Expert Case Review.
