AUTHORS
Ahmed Sawas, MD, Selina Radlein, PhD2, Hannah Mandle, MS3, Patrick Ward, PhD4
AFFILIATIONS
1 Flatiron Health, New York, New York, United States
2 Flatiron Health, New York, New York, United States
3 Flatiron Health, Durham, North Carolina, United States
4 Flatiron Health, Durham, North Carolina, United States
ABSTRACT
Background: Patients with chronic lymphocytic leukemia (CLL) previously exposed to both Bruton tyrosine kinase inhibitors (BTKi) and B-cell lymphoma (BCL)2 inhibitors (BCL2i) represent an increasingly common population with limited evidence to guide subsequent treatment. Existing real-world studies describe treatment patterns but not treatment selection. We used a large language model (LLMs) to enable scalable extraction of clinical reasoning from unstructured electronic health record (EHR) notes, providing insights into real-world treatment decision-making.
Methods: This retrospective cohort study used the US-based, EHR-derived, deidentified Flatiron Health Research Database with >137,000 patients with CLL. Adult patients with CLL who received both BTKi and BCL2i within the first 2 lines of therapy (1L/2L), either sequentially or concurrently, were eligible. Patients whose first dual-class exposure occurred beyond 2L were excluded. The index event was the first qualifying CLL-directed treatment following dual-class exposure and prior to Richter transformation.
An LLM extracted structured information from clinician-patient discussions: 1) disease status, 2) treatment discontinuation reasons, 3) clinician-documented treatment-selection drivers, 4) patient- and disease-related decision factors, 5) alternative therapies discussed and 6) clinician interpretation. Variables and taxonomies were clinician-defined and iteratively refined. Discontinuation reasons were grouped into clinically meaningful transition categories (progression-related, treatment intolerance, fixed-duration completion, patient-driven, and other). Treatment strategies were grouped into therapeutic classes. Descriptive analyses characterized treatment transitions, decision drivers, and subsequent treatment strategies. Exploratory analyses evaluated real-world time to next treatment (rwTTNT) overall and by transition category.
Results: Among 1471 patients with dual BTKi/BCL2i exposure, 9% initiated a subsequent CLL-directed treatment and 2% had an observation period of ≥6 months before subsequent therapy.
At index, clinical progression was the most common disease status. Treatment transitions were progression-related in 58%; toxicity-related in 33%; and fixed-duration completion, patient-driven, or other in 9%.
The most common subsequent strategies were alternative covalent BTKi (acBTKi; 29%), venetoclax-based therapy (20%), non-covalent BTKi (13%), observation (9%), and other therapies (29%), including chemoimmunotherapy, anti-CD20 therapy, PI3K inhibitors, clinical trials, and cellular therapy.
Primary treatment-selection drivers were prior class failure, progressive disease, prior toxicity, prior response history, and expected efficacy. Patient factors included age, frailty, cardiovascular disease, infection history, and preference; disease factors included high-risk biology and transformation risk. Limited therapeutic options were documented in 46%, and sequencing uncertainty in 27%.
Progression- and toxicity-related transitions most commonly led to acBTKi. Median rwTTNT was 8 months (IQR, 4-14) overall, with median follow-up of 16 months (IQR, 8-26). Median rwTTNT was longer following toxicity-related transitions than progression-related transitions (21 [IQR, 11-27] vs 15 [IQR, 6-23] months).
Conclusions: Following dual BTKi/BCL2i exposure, treatment selection in CLL was associated with clinician-documented factors beyond prior treatment history, including treatment transition, patient fitness, disease characteristics, and prior treatment experience. LLM-enabled extraction of unstructured EHR documentation provided a scalable approach to characterizing real-world clinical decision-making, complementing traditional patient journey analyses. Frequent retreatment with covalent BTKi/BCL2i and limited use of non-covalent BTKi reflect uncertainty regarding optimal sequencing after dual BTKi/BCL2i exposure and underscores the need for prospective studies and real-world evidence to inform treatment selection. Longer rwTTNT among patients transitioning because of treatment intolerance compared with disease progression suggests that the clinical context underlying treatment selection may influence subsequent clinical outcomes.
DISCLOSURES
Ahmed Sawas, MD: Roche - Current equity holder in publicly-traded company; Pfizer - Honoraria; Eli Lilly - Honoraria; Flatiron Health, Inc. - Research Funding; Flatiron Health, Inc. - Current Employment