FIVE CLINICALLY DISTINCT CHRONIC LYMPHOCYTIC LEUKEMIA (CLL) PRESCRIBER PERSONAS IDENTIFIED VIA UNSUPERVISED MACHINE LEARNING USING NORSTELLALINQ REAL-WORLD DATA
Author(s)
Isabella Even-Chen, BA1, ilan behm, MPH2, Shefali Patel, MS3, Juan Diego Irizarry-Cole, PhD1, Rahul Das, PhD4, Atharva Manjrekar, MS5, Allison Perry, PhD1.
1Norstella, New York, NY, USA, 2Norstella, Englewood, CO, USA, 3Norstella, Ferndale, MI, USA, 4Norstella, Yardley, PA, USA, 5Norstella, West Hartford, CT, USA.
1Norstella, New York, NY, USA, 2Norstella, Englewood, CO, USA, 3Norstella, Ferndale, MI, USA, 4Norstella, Yardley, PA, USA, 5Norstella, West Hartford, CT, USA.
OBJECTIVES: To identify and characterize distinct HCP personas in CLL using HDBSCAN clustering applied to integrated structured and AI-imputed real-world data, and to quantify differences in treatment utilization, patient complexity, and clinical decision-making across segments.
METHODS: An HCP-level dataset was constructed using NorstellaLinQ’s US real-world linked open claims, structured EHR, and clinical notes for providers with >=1 CLL patient (January 2019-September 2025). Features included BTKi and BCL-2 utilization rates, chemotherapy use, treatment duration by line, patient distribution across lines 1-4, BTKi rechallenge and switching rates, biomarker positivity prevalence, and clinical trial participation. A gradient-boosted model imputed behavioral features for HCPs without unstructured EHR data (AUC=0.90). HDBSCAN clustering was applied to the combined feature matrix with iterative stability assessment.
RESULTS: Five personas were identified across 9,424 providers: Early-Line Generalists (n=2,111) with low BTKi utilization (6%) and short durations (62 days); Chemo-Forward Early-Line Traditionalists (n=1,132) with minimal biomarker burden; BTKi-First Early-Line Precision Managers (n=2,284) with near-universal BTKi adoption (98%) and long average durations (444 days); High-Complexity Late-Line Strategists (n=1,845) with moderate BTKi utilization (72%), high late-line involvement, and complex pretreated patient panels; and Late-Line Targeted Maximizers (n=2,052) with the highest BTKi (95%) and BCL-2 (87%) utilization, longest average durations (424 days), highest CLL patient volume (70 patients), and greatest late-stage and biomarker-positive burden. BTKi utilization ranged 16-fold across clusters.
CONCLUSIONS: HDBSCAN applied to integrated structured and AI-imputed RWD identifies five clinically interpretable CLL HCP personas with substantially different prescribing patterns, patient complexity profiles, and sequencing behaviors. This segmentation framework supports more targeted evaluation of treatment adoption patterns and provider-level heterogeneity in CLL care.
METHODS: An HCP-level dataset was constructed using NorstellaLinQ’s US real-world linked open claims, structured EHR, and clinical notes for providers with >=1 CLL patient (January 2019-September 2025). Features included BTKi and BCL-2 utilization rates, chemotherapy use, treatment duration by line, patient distribution across lines 1-4, BTKi rechallenge and switching rates, biomarker positivity prevalence, and clinical trial participation. A gradient-boosted model imputed behavioral features for HCPs without unstructured EHR data (AUC=0.90). HDBSCAN clustering was applied to the combined feature matrix with iterative stability assessment.
RESULTS: Five personas were identified across 9,424 providers: Early-Line Generalists (n=2,111) with low BTKi utilization (6%) and short durations (62 days); Chemo-Forward Early-Line Traditionalists (n=1,132) with minimal biomarker burden; BTKi-First Early-Line Precision Managers (n=2,284) with near-universal BTKi adoption (98%) and long average durations (444 days); High-Complexity Late-Line Strategists (n=1,845) with moderate BTKi utilization (72%), high late-line involvement, and complex pretreated patient panels; and Late-Line Targeted Maximizers (n=2,052) with the highest BTKi (95%) and BCL-2 (87%) utilization, longest average durations (424 days), highest CLL patient volume (70 patients), and greatest late-stage and biomarker-positive burden. BTKi utilization ranged 16-fold across clusters.
CONCLUSIONS: HDBSCAN applied to integrated structured and AI-imputed RWD identifies five clinically interpretable CLL HCP personas with substantially different prescribing patterns, patient complexity profiles, and sequencing behaviors. This segmentation framework supports more targeted evaluation of treatment adoption patterns and provider-level heterogeneity in CLL care.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
RWD18
Topic
Methodological & Statistical Research, Real World Data & Information Systems
Topic Subcategory
Health & Insurance Records Systems
Disease
Oncology