Registry-based machine learning in Sezary syndrome: Assessment of prognostic signal and age-masked racial inequities.

Abstract

e22615 Background: Sézary syndrome (SS) is a rare leukemic cutaneous T-cell lymphoma with limited contemporary population data on survival, disparities, and prognostication. Methods: SEER-22 identified adults with first primary SS (ICD-O-3 9701/3), 2000–2021 (N = 403). Overall survival (OS) was evaluated by Kaplan–Meier and Cox models adjusting for age, sex, calendar year, SEER Summary Stage, and county-level socioeconomic status (SES). To evaluate machine learning (ML)–based registry risk stratification, we split the cohort by diagnosis year into training and held-out test sets and trained four survival models using identical covariates: multivariable Cox, elastic-net Cox, random survival forest, and gradient-boosted Cox. Model performance was evaluated using Harrell’s C-index, time-dependent Uno AUC (12/36/60 months), calibration, and Brier score. Results: Median age was 68; 77% were White and 19% Black. Black patients were younger at diagnosis (46.8% < 60 vs 25.4% in Whites; p = 0.0009). Median OS was 48 months; 1- and 5-year OS were 83.2% and 42.7%. Median OS was 37.0 months in Black patients versus 49.0 months in White patients (Other/Unknown: 36.0 months). Survival improved in 2011–2021 compared with 2000–2010 (median OS 56.0 vs 39.5 months; 12-month OS 87.5% vs 75.7%; 60-month OS 48.4% vs 35.0%). Racial OS was similar at 12 months (83.9% Black vs 82.9% White) but diverged thereafter. In Cox models, the crude Black vs White hazard ratio (HR) was 1.19 (0.86–1.63) but increased after age/sex/era adjustment; in the fully adjusted model, Black race remained independently associated with higher mortality (HR 1.61, 95% CI 1.14–2.27; p = 0.007), along with age (per decade HR 1.38, 95% CI 1.23–1.53; p < 0.001), while stage/SES were not. Time-windowed analyses suggested excess hazard beyond the first year. In the SS test set, all ML approaches demonstrated near-chance discrimination (C-indices ~0.5; low Uno AUCs) with minimal improvement in calibration. By contrast, applying the same modeling pipeline to a substantially larger SEER mycosis fungoides cohort yielded strong discrimination (C-index ~0.84). Conclusions: SS survival has improved modestly in the modern era but remains poor. Younger age at diagnosis masks inequity in unadjusted analyses; adjusted models reveal persistent longer-term mortality disadvantage for Black patients. Routine registry variables provide limited prognostic signal for SS, underscoring the need for biomarker- and treatment-sequence–enriched datasets to enable actionable risk stratification and address disparities.