Machine Learning Model Predicts Monoclonal Gammopathy Using Routine Laboratory Values

Abstract

PURPOSE

Monoclonal gammopathy (MG) is a disorder defined by the presence of monoclonal immunoglobulin. Monoclonal protein (M-protein) serves as an important biomarker for the spectrum of MG that includes plasma cell myeloma (PCM) and, its precursor, MG of undetermined significance (MGUS). Recent evidence suggest early identification of MGUS before the onset of PCM is associated with improvement in overall survival. Despite these benefits, the recognition rate of MG remains inadequate. Addressing this care gap would help identify at-risk patients who may benefit from targeted evaluation to prevent adverse outcomes. However, to date, to our knowledge, there are no widely used machine learning (ML) models that predict MG. Therefore, our aim was to leverage routine laboratory data to train and evaluate the performance of ML-based risk models for M-protein associated with MG.

METHODS

The study population was composed of deidentified laboratory data procured from a cohort of 232,813 individuals within a large US outpatient network. The 7-year longitudinal data set included 1,610 patients with the following inclusion criteria: age 50–85 years, ≥3 complete blood count and metabolic panel results, and at least one protein electrophoresis result. ML models were developed using XGBoost. The reference outcome was M-protein.

RESULTS

The seven-variable risk classifier model accurately predicted M-protein within 5 years and achieved an AUC of 0.84. The most important predictors were absolute lymphocyte trajectory, age, RBC, total protein, RBC distribution width, blood urea nitrogen, and relative eosinophils.

CONCLUSION

Our ML risk classifier accurately predicted the presence of M-protein using routine laboratory data. Although prospective studies are warranted, the results support the clinical utility of the model to improve timely recognition for patients at risk for MG.

INTRODUCTION

Monoclonal gammopathy (MG) is a condition defined by the presence of monoclonal immunoglobulin.¹ Monoclonal protein (M-protein) serves as an important biomarker for the spectrum of MG. Downstream, plasma cell myeloma (PCM) represents the most clinically impactful entity and contributes a significant cause of morbidity and mortality, worldwide.² Cancer statistics from the United States demonstrate that the disease affects 7.3 individuals per 100,000, with a 5-year survival rate of 62.4%.³ The diagnosis of PCM is supported by the following hematologic criteria: serum M-protein ≥3 g/dL or urinary M-protein ≥500 mg/24 h and/or ≥10% clonal plasma cells in the bone marrow.^1,4 Although up to 30% of patients are asymptomatic at presentation, the majority require treatment at the time of diagnosis.⁵ Moreover, complications over the course of the disease are substantial and include increased risk of end-organ damage, manifested by hypercalcemia, renal insufficiency, anemia, and bone lesions.^6,7

CONTEXT

• Key Objective

• Can readily available routine laboratory data train and test the performance of a machine learning–based risk model for monoclonal gammopathy (MG)?

• Knowledge Generated

• The seven-variable risk classifier model accurately predicted MG and achieved an AUC of 0.84. The most important predictors were absolute lymphocyte trajectory, age, RBC, total protein, RBC distribution width, blood urea nitrogen, and relative eosinophils.

• Relevance (U. Topaloglu)

• Early detection of MG has been shown to improve overall survival, yet current recognition rates remain low, leaving many at-risk patients unidentified until progression to plasma cell myeloma. This study demonstrates that a machine-learning risk classifier using only routine laboratory data can accurately predict monoclonal protein presence, highlighting a scalable tool that may enhance timely diagnosis and improve outcomes for patients with MG.*

• *Relevance section written by JCO CCI Associate Editor Umit Topaloglu, PhD.

Upstream, MG of undetermined significance (MGUS) is the principal precursor of PCM, and it is characterized by serum M-protein <3 g/dL, <10% clonal bone marrow plasma cells, and absence of end-organ damage.¹ Compared with PCM, the prevalence of MGUS is much higher and estimated to affect at least 3,000 per 100,000 persons, or 3% of the general population older than 50 years.⁸ Owing largely to its often-asymptomatic clinical presentation, evidence of MGUS may be detected retrospectively in nearly all cases of PCM, often many years in advance of the initial diagnosis.^9,10 Indeed, the annual and lifetime risk of progression from MGUS to PCM is significantly increased at 1% and 40%, respectively.^9,11,12

In addition to its role as a precursor for PCM, MGUS is also independently associated with increased all-cause mortality and significant comorbidities, including cardiovascular, peripheral vascular, and kidney disease.¹³ Moreover, evidence from large-scale observational studies suggest that the early identification of MGUS—before the onset of PCM—is associated with a 13%-15% improvement in overall survival.^14-16 Despite the benefits of early detection, the clinical recognition rate for MGUS remains inadequate. To be sure, <10% of patients diagnosed with PCM have a documented prior history of MGUS, representing a substantial care gap in hematologic oncology.¹⁴ This care gap precludes the use of validated risk-stratification tools, leaving high-risk individuals without the monitoring necessary to intercept disease progression.¹⁷

Recently, machine learning (ML) has emerged as a powerful method for harnessing large volumes of health care data to predict disease risk across a broad swathe of clinical domains.^18-22 However, to date, to our knowledge, there are no widely implemented ML tools that predict the risk of MG among general adult outpatients. Addressing this knowledge gap would aid in detecting high-risk patients within this heterogeneous population who may benefit from further diagnostic evaluation—as well as targeted counseling, monitoring, and management—at an earlier stage to prevent adverse outcomes.⁷

Against this background, the aim of our study was to leverage (1) longitudinal hematologic and biochemical data from a geographically diverse US cohort and (2) ML methods to develop a readily obtainable risk prediction model for abnormal M-protein associated with MG.

METHODS

Study Design and Population

The study was reviewed and approved by the institutional review board at Western Institutional Review Board—Copernicus Group. Informed consent was waived due to the lack of feasibility of obtaining consent from all participants, and the data reviewed were deidentified, records-based, and retrospective in nature. The initial study cohort included deidentified data from 232,813 adult participants from the Southwest, Mid-South, and Pacific regions of the United States. The overall median age was 68 years, and 38% of participants were male.

Data Collection and Measurements

Complete blood count (CBC) and metabolic panel testing were performed on the equivalent high-throughput instrument platform across all testing sites and were standardized according to best practices.²³ Where applicable, reference intervals for select variables were as follows: white blood cell count (3.8-11.2 ×10⁹/L), absolute lymphocyte count (ALC; 0.7-4.5 ×10⁹/L), platelet count (150-450 ×10⁹/L), relative distribution width (11%-15%), creatinine (0.80-1.40 mg/dL), blood urea nitrogen (BUN; 8-24 mg/dL), and calcium (8.6-10.3 mg/dL).

The training and test cohorts were derived from the laboratory information system (LIS) at outpatient facilities within Sonic Healthcare USA with dates of service between January 1, 2017, and December 31, 2024. Outcomes were determined by reviewing records contained in the LIS.

Candidates were selected based on availability of sufficient demographic (eg, age, sex) and laboratory data. Minimum criteria for inclusion in the study included at least three CBC and metabolic panel values. To better capture the longitudinal patterns and dynamic kinetics, natural cubic splines were computed using R Splines package for selected variables using time indexed as days since the first result (R Foundation for Statistical Computing, Vienna, Austria). Spline coefficients summarize early/mid/late trajectory components.²⁴ First (baseline) and last values were also recorded, and linear slopes were derived when ≥2 time points were available. The outcome of interest was M-protein detected by protein electrophoresis.

Statistical Analyses

All analyses were performed in R (packages including data.table, dplyr, xgboost, pROC, caret, splines, and doParallel). After feature engineering (first/last values, in addition to linear slopes, and spline coefficients for time point analysis), we assembled a modeling matrix joined to the binary positive monoclonal antibody test result label. To mitigate potential collinearity between related features (eg, first/last values and derived slopes), we applied model-based feature selection and regularization within XGBoost in addition to cross-validation in the model. Importantly, XGBoost’s inherent handling of correlated predictors minimizes instability, and we confirmand that the final feature set did not exhibit problematic redundancy. This approach ensures that inclusion of splines, slope, and first/last values captures complementary longitudinal information rather than introduction of bias. The data set was split with stratified sampling into training (80%) and held-out test (20%). The stratification was performed on the outcome variable (presence of the monoclonal antibody). This ensured that the proportion of the positive and negative cases was preserved across training and test sets, thereby avoiding imbalance-related bias in performance evaluation. Within the training data, we conducted a constrained feature search over combinations of k = 6-24 predictors drawn from the engineered pool of 48 initial features plus their slopes and splines from CBC and metabolic panels (eg, age at first observation; first values for RBC, total protein, eosinophils, BUN, RBC distribution width [RDW], total protein—serum albumin [globulin gap], mean corpuscular volume; spline coefficients, such as lymphocytes spline [second coefficient] and hemoglobin spline [first coefficient]; and slopes, such as calcium and albumin).

After initial model and feature importance weighing, 24 total features remained. For each candidate set, an XGBoost classifier (binary: logistic; η = 0.05, max_depth = 6, subsample = 0.9, colsample_bytree ≈ 0.8-0.85, min_child_weight = 3) was tuned using 10-fold cross-validation with up to 8,000 rounds and early stopping. Using 10-fold cross-validation, we evaluated the model's AUC and analyzed confusion matrix metrics—sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV)—by testing predicted probabilities on a validation set at thresholds from 0.30 to 0.80. This range was selected to balance the trade-off between sensitivity and specificity while avoiding extreme cutoffs (<0.3 or >0.8) that provided little clinical utility in preliminary assessments. We performed a cross-validated AUC evaluation and, on a validation split, derived confusion matrix metrics by thresholding predicted probabilities from 0.30 to 0.80. A composite score (0.30·sensitivity + 0.30·specificity + 0.20·PPV + 0.30·NPV) guided model selection. To favor rule-out utility, we additionally filtered to the top NPV quantiles before choosing the highest-scoring configuration. The selected model was then retrained with its optimal rounds on the training data and evaluated on the held-out test set with receiver operating characteristic/AUC and 95% CIs (DeLong) using pROC. Parallelization applied doParallel to accelerate cross-validation.

For descriptive comparisons between the model-predicted positive and model-predicted negative groups, continuous variables were summarized as median and compared using Wilcoxon rank-sum test. To control for multiple comparison across features, P values were adjusted using the Benjamini-Hochberg false-positive discovery rate; P < .05 was considered statistically significant.²⁵

Data and Resource Availability

The deidentified data sets generated during and/or analyzed in the current study are available from the corresponding author upon reasonable request.

RESULTS

Baseline Characteristics

The initial registry included 232,813 adult patients with an outpatient clinical laboratory result between January 1, 2017, and December 31, 2024. To account for the well-established challenge of interlaboratory variation, CBC and metabolic panel testing were performed on equivalent high-throughput instrument platforms and standardized according to best practices.²³ Test results were omitted from the study if they demonstrated any of the following: initial age of the patient <50 years or >85 years; fewer than three distinct CBC and metabolic panel time points; or less than one protein electrophoresis result.

After applying exclusion criteria, the data set contained 1,610 patients (Fig 1) with a median age of 68 years. By CBC and metabolic panel, the cohort included participants with the following baseline (median) results for WBC, absolute lymphocytes, creatinine, and calcium, respectively: 6.10 ×10⁹/L, 1.63 ×10⁹/L, 0.93 mg/dL, and 9.50 mg/dL (Table 1). The end point prevalence (positive M-protein) was 45% in training and 44% in test, with median observation times of 50 and 48.4 months, respectively (Table 1).

FIG 1.

Participant flow diagram. The flow diagram depicts the number of adult participants with protein electrophoresis results in the original data set, before removal due to exclusion criteria. Omitted participants included those with age <50 years or >85 years; less than three CBC results; and less than one protein electrophoresis result. The final data set included 1,610 participants. CBC, complete blood count.

TABLE 1.

Baseline Characteristics

Category	Training	Test
Patient demographics
Number of unique individuals	1,308	302
Age, years, median	68	67
Male, %	38	38
Laboratory values
WBC, median, ×109/L	6.10	6.20
Hemoglobin, median, g/dL	13.2	13
Lymphocytes, median, ×109/L	1.63	1.56
Relative eosinophils, median, %	2.50	2.30
Platelet, median, ×109/L	235	241
Creatinine, median, mg/dL	0.93	0.96
BUN, median, mg/dL	16	17
Calcium, median, mg/dL	9.50	9.50
Total protein, median, g/dL	6.90	6.90
End point, positive monoclonal protein
Events, %	45	44
Observation time, months	50.9	48.4

NOTE. Observation time is defined as the average number of months of laboratory data procured for the cohort.

Abbreviation: BUN, blood urea nitrogen.

Machine Learning Models

Models using multiple variables were trained and tested to predict positive M-protein associated with MG. Based on XGBoost analysis, the seven-variable cross-validated Monoclonal Gammopathy Classifier was accurate and achieved an overall AUC of 0.84 (Fig 2). The PPV and NPV for the model were 0.62 and 0.89, respectively. Overall, the most meaningful predictors of progression to abnormal M-protein were early absolute lymphocyte trajectory (spline-derived), age (initial), RBC (initial), total protein (initial), RDW (initial), BUN (initial), and relative eosinophils (EOS; initial; Fig 3). The addition of other routine laboratory variables (eg, total protein—albumin gap) did not improve the performance of the model.

FIG 2.

ROC curve of the machine learning MG risk classifier model showing prediction performance for positive monoclonal protein. The seven-variable classifier included ALC (spline), age, RBC, total protein, RDW, BUN, and eosinophils. ALC, absolute lymphocyte count; BUN, blood urea nitrogen; FPR, false positive rate; MG, monoclonal gammopathy; PCM, plasma cell myeloma; RDW, RBC distribution width; ROC, receiver operating characteristic; TPR, true positive rate.

FIG 3.

Variable importance. ALC, absolute lymphocyte count; BUN, blood urea nitrogen; EOS, relative eosinophils; RDW, RBC distribution width.

Further evaluation of the model compared the positive M-protein and negative M-protein groups against key variables. Compared with the negative group, the positive M-protein group demonstrated decreased RBC (4.10 v 4.45 ×10¹²/L, P < .01); decreased total protein (7 v 7.16 mg/dL, P < .01); and increased age (70 v 66 years, P < .01), RDW (13.40 v 13, P < .01), BUN (17 v 16 mg/dL, P < .01), and EOS (2.60% v 2.30% ×10⁹/L, P < .05; Table 2).

TABLE 2.

Machine Learning Monoclonal Gammopathy Risk Classifier: Summary

Category	Positive Monoclonal Protein Group	Negative Monoclonal Protein Group	P
Age, years, median	70	66	<.01
ALC spline, median, ×109/L	–0.03	–2.29	.01
RBC, median, ×1012/L	4.10	4.45	<.01
RDW, median, %	13.40	13	<.01
Eosinophils, median, %	2.60	2.30	<.01
BUN, median, mg/dL	17	16	<.01
Total protein, median, g/dL	7	7.16	<.01

NOTE. Positive monoclonal protein group: cohort predicted by the model to demonstrate positive monoclonal protein. Negative monoclonal protein group, cohort predicted by the model to demonstrate absence of a monoclonal protein. P > .05.

Abbreviations: ALC, absolute lymphocyte count; BUN, blood urea nitrogen; RDW, RBC distribution width.

DISCUSSION

Leveraging real-world data and ML, this retrospective cohort study trained and tested a risk prediction model for MG across a large, diverse US adult population. The performance of the seven-variable ML MG risk classifier was robust with an overall AUC value of 0.84.

A key strength of this investigation is that the risk prediction model is uniquely based on ML XGBoost analysis, which may be better suited to solve complex problems with discrete and longitudinal features.^26-28 Consistent with previous plasma cell dyscrasia studies, our findings demonstrate that the following static variables are associated with MG compared with non-MG: lower RBC,²⁹ older age,¹ higher RDW,³⁰ higher BUN,³¹ and elevated EOS.³² Somewhat unexpectedly, the MG group was associated with lower total protein. In contrast to PCM, this finding has been described in early MGUS and may be secondary to a decrease in serum albumin levels that can outpace the gain in M-protein, resulting in decreased total protein.^33,34 Notably, the median laboratory test results across key parameters (ie, RBC, RDW, total protein, BUN, and EOS) for both groups were well within established reference intervals for these tests, suggesting the target population is enriched for early-stage MG, rather than advanced, symptomatic PCM. Among longitudinal variables, ALC spline was the most significant contributor to the classifier and accounts for 25% of the variable importance to the model. Compared with the gradual ALC decline seen in the normal aging process,³⁵ patients with a positive M-protein exhibited steady ALCs. This finding adds to a growing body of evidence demonstrating the complex and dynamic sequence of immune dysfunction, proliferation/control, and evasion as a hallmark of plasma cell dyscrasias.^36-38 Collectively, the results underpin the important role of ML to (1) detect subtle and, often intricate, abnormalities in routine blood chemistry parameters; and (2) predict M-protein associated with MG.

Another strength of the MG risk classifier is that the model employs existing and readily obtainable information found within the widely used CBC and metabolic panel and could be broadly deployed in a value-based manner. This is particularly important because MG is underrecognized in the primary care setting,¹⁴ particularly in underresourced areas.² Using ML, the delay in diagnosis could be abrogated by the classifier as it detects subtle changes in CBC and metabolic panel parameters over time and alerts clinicians to initiate further evaluation. Moreover, the therapeutic landscape for early (smoldering) PCM and/or MGUS continues to be actively investigated³⁹; and, with emerging evidence, timely recognition of early-stage disease will likely become critical to enroll eligible patients in clinical trials and/or realize potential treatment benefits.

Recently, work by Cai et al⁴⁰ successfully generated a ML model using CBC and flow cytometric parameters to rapidly screen for patients with PCM. Other PCM models have focused on survival, prognosis, and/or morbidity after the initial diagnosis.^41,42 Although there are similarities, the current study uniquely uses a large and diverse US population to predict which patients are at high risk for demonstrating M-protein associated with MG within 5 years. By contrast, the Cai ML model relies predominantly on an Asian population and serves primarily to rapidly screen patients exclusively for PCM.⁴⁰ Taken together, these data highlight the success and potential of ML risk classifiers to improve the overall clinical care across the spectrum of MG and plasma cell dyscrasias.

There are several limitations to this investigation that should be considered. First, the study cohort defined MG based primarily on laboratory values (positive M-protein), rather than established clinical diagnosis. However, multiple studies have demonstrated that MG is frequently underrecognized, particularly in the primary care setting.^1,33 As a result, studies that exclusively rely on clinically confirmed MG preclude interrogation of a significant undiagnosed population. Due to the robust size and inclusion of undiagnosed disease, the findings seen here may be more applicable to the general population in the United States. If implemented broadly, it could also improve the overall identification and diagnosis of disease. Second, the outcome of interest (positive M-protein) does not capture light-chain MGUS (LC-MGUS), which is characterized by an abnormal free light chain (FLC) ratio in the absence of a heavy-chain M-protein. However, recent data suggest that the overall prevalence of LC-MGUS is remarkably small (approximately 0.3%) and that standard FLC reference intervals can lead to an elevated false-positive rate. Using positive M-protein detected by protein electrophoresis as the outcome of interest avoids diagnostic noise where standard reference ranges for FLC ratios are still being refined.^14,43 Third, the outcome of interest encompasses MG of clinical significance (MGCS), a spectrum of nonmalignant conditions characterized by protein-mediated organ damage across varied systems.⁴⁴ Although these conditions are not classified as traditional blood cancers, identifying MGCS is important for clarifying what is often a highly complex clinical diagnosis and, possibly, preventing end-organ damage. In neurology, for example, detecting MGCS is important for diagnosing treatable conditions such as anti–myelin-associated glycoprotein neuropathy or polyneuropathy, organomegaly, endocrinopathy, M-protein, and skin changes (so-called POEMS)⁴⁵ syndrome. In rheumatology, detecting MGCS is essential to identifying rare disorders such as Schnitzler syndrome or cryoglobulinemic vasculitis. Although not necessary for routine cases, detection allows for clone-directed therapies (eg, rituximab) that can stop irreversible nerve or tissue damage that standard treatments might miss.^46,47 Fourth, the study data set includes testing typically ordered in response to clinical suspicion (eg, protein electrophoresis), rather than used as screening tools. This suspicion-driven testing enriches the cohort for individuals already at elevated risk and introduces potential selection bias, a well-recognized challenge in real-world health care data sets. Despite these limitations, real-world data remain important for population-based ML research because they capture the complexity of routine clinical practice and patient heterogeneity in different ways compared with controlled trials and unselected populations. This is consistent with recent work highlighting the important role of real-world data in bridging the gap between clinical trials and actual practice.^48,49 Finally, the study design did not evaluate clinical variables for plasma cell disorders, including imaging, family history, prior exposure to environmental toxins, body weight, molecular-cytogenetic abnormalities, and other risk factors.^50-52 Although incorporating these elements may have improved the accuracy, it also would have added more complexity and, arguably, decreased the overall usability of the model. Nevertheless, these limitations raise important questions of how the risk prediction model functions in targeted populations associated with plasma cell neoplasms. These questions and others should be answered in prospective cohorts, enriched for known clinical risk factors for plasma cell dyscrasias.

Developed and validated in a study cohort of 232,813 patients, the ML MG risk classifier was accurate and achieved an AUC of 0.84. The most important predictors of MG were ALC trajectories, age, RBC count, and total protein. The model was derived from routine demographic and laboratory values in a large and geographically diverse US outpatient cohort, which strengthens its overall generalizability and usability. To our knowledge, it is the first ML model reported in the literature to predict MG using routine laboratory data. Given its underrecognition in the community setting, the ML risk classifier has the potential to significantly improve timely, cost-effective identification of a population with elevated risk for MG at an earlier stage for evaluation.

DATA SHARING STATEMENT

A data sharing statement provided by the authors is available with this article at DOI https://doi.org/10.1200/CCI-26-00013.

AUTHOR CONTRIBUTIONS

Conception and design: Mercedeh Movassagh, Cihan Kaya, Mohamed E. Salama, Joseph Aoki

Administrative support: Mohamed E. Salama

Provision of study materials or patients: Con Skordis, Mohamed E. Salama

Collection and assembly of data: Con Skordis

Data analysis and interpretation: Mercedeh Movassagh, Cihan Kaya, Ella Aoki, Tarush Kothari, Mohamed E. Salama, Joseph Aoki

Manuscript writing: All authors

Final approval of manuscript: All authors

Accountable for all aspects of the work: All authors

AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST

The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/cci/author-center.

Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments).