Identifying Eligibility for Chimeric Antigen Receptor T-Cell Therapy Among Diffuse Large B-Cell Lymphoma Patients Using Real-World Data and Unsupervised Machine Learning

OBJECTIVES: In the United States, patients with relapsed or refractory diffuse large B-cell lymphoma (DLBCL) who have failed at least two lines of systemic therapy may be eligible for Chimeric Antigen Receptor T-cell (CAR-T) therapy. Investigating lines of therapy to evaluate eligibility for CAR-T can be difficult for researchers using real-world data, including electronic health records and claims data. This is because the heterogeneity of treatment regimens, many of which do not precisely align with treatment guidelines, make identifying switches in lines of therapy challenging. We explore whether unsupervised machine learning may be useful for identifying switches in treatment lines for DLBCL patients.

METHODS: We used 2007 - 2022 electronic health record data (TriNetX) to identify 8,850 DLBCL patients using ICD-10-CM diagnosis codes (C83.3). We identified 32 drugs used for DLBCL treatment from National Comprehensive Cancer Network (NCCN) and American Cancer Society (ACS) guidelines and the literature. For each patient, drugs delivered within a 7-day window were grouped into multi-drug encounters. Mini-Batch K-Means clustered encounters on drug domains identified with Multiple Correspondence Analysis (MCA). We developed treatment line switch indicators with clustering results and information from NCCN guidelines and expert opinion. We identified combinations of indicators that were strong enough to indicate treatment line switches.

RESULTS: An algorithm based on multiple indicators successfully identified meaningful treatment line switches, identifying 204 patients who were potentially eligible for CAR-T therapy. The indicators, based on clustering results, attempted to rule out maintenance treatment, modifications due to toxicity issues, and treatments that were usually not considered to be new lines of treatment. Some relatively weak indicators also successfully identified most line switches, but likely misclassified some cases.

CONCLUSIONS: Unsupervised machine learning is a promising approach for identifying treatment line switches. However, selection of switch indicators requires careful consideration of tradeoffs between sensitivity and specificity.