Post-MGUS Diagnosis Serum Monoclonal-Protein Velocity and the Progression of Monoclonal Gammopathy of Undetermined Significance to Multiple Myeloma

Natural Language Processing-Assisted Classification Models to Confirm Monoclonal Gammopathy of Undetermined Significance and Progression in Veterans' Electronic Health Records

PURPOSE

To develop and validate natural language processing (NLP)-assisted machine learning (ML)-based classification models to confirm diagnoses of monoclonal gammopathy of undetermined significance (MGUS) and multiple myeloma (MM) from electronic health records (EHRs) in the Veterans Health Administration (VHA).

MATERIALS AND METHODS

We developed precompiled lexicons and classification rules as features for the following ML classifiers: logistic regression, random forest, and support vector machines (SVMs). These features were trained on 36,044 EHR documents from a random sample of 400 patients with at least one International Classification of Disease code for MGUS diagnosis from 1999 to 2021. The best-performing feature combination was calibrated in the validation set (17,826 documents/200 patients) and evaluated in the testing set (9,250 documents/100 patients). Model performance in diagnosis confirmation was compared with manual chart review results (gold standard) using recall, precision, accuracy, and F1 score. For patients correctly labeled as disease-positive, the difference between model-identified diagnosis dates and the gold standard was also computed.

RESULTS

In the testing set, the NLP-assisted classification model using SVMs achieved best performance in both MGUS and MM confirmation with recall/precision/accuracy/F1 of 98.8%/93.3%/93.0%/96.0% for MGUS and 100.0%/92.3%/99.0%/96.0% for MM. Dates of diagnoses matched (±45 days) with those of gold standard in 73.0% of model-confirmed MGUS and 84.6% of model-confirmed MM.

CONCLUSION

An NLP-assisted classification model can reliably confirm MGUS and MM diagnoses and dates and extract laboratory results using automated interpretation of EHR data. This algorithm has the potential to be adapted to other disease areas in VHA EHR system.

Smoldering multiple myeloma: Reviewing the rationale for intervention

Smoldering multiple myeloma has been recognized for over 40 years and represents a pre-symptomatic phase of the 2nd most common hematologic malignancy. 1/3 of patients will remain asymptomatic at 10 years. There is an identifiable subset of patients that will develop CRAB within 2 years of recognition and these patients are considered for therapeutic intervention before the development of potentially irreversible complications. Obstacles to widespread implementation of therapeutic guidelines are limited by the variable definitions associated with this high-risk group as well as the poor concordance between classification schemes. Analysis of clinical trial outcomes as well as uniform eligibility helps determine whether a given patient should be considered for therapeutic intervention outside of a clinical trial.

How I approach smoldering multiple myeloma

The current standard of care in smoldering multiple myeloma (SMM) is close surveillance, outside of clinical trials. Efforts are being made to understand the pathobiologic process that leads to the progression of SMM to active MM. This review provides a critical description of available data, including risk factors and risk models of progression, as well as clinical trials investigating interventions for this patient population. We describe 2 cases in which patients were seen before the concept of a myeloma-defining event was established. Today, based on the International Myeloma Working Group criteria, both patients would have been identified as experiencing myeloma-defining events, and therapy would have been initiated. These cases show that occasionally, patients can undergo observation only, even when they exceed criteria for high-risk SMM.