Interpretable machine learning for prognostic survival prediction in mycosis fungoides: Model development and external validation.

Abstract

e22612 Background: Mycosis fungoides (MF) has heterogeneous outcomes and persistent disparities, yet lacks widely adopted, validated prognostic tools. We developed and externally validated machine learning (ML) survival models using registry variables to identify independent predictive indicators and generate individualized survival predictions. Methods: MF (ICD-O-3 9700/3) diagnosed 2000–2021 was identified in SEER-22 database. Overall survival (OS) was from diagnosis to death (any cause) or last follow-up. Independent predictors were evaluated with a fully adjusted multivariable Cox model including race, age, sex, Summary Stage, and county-level household income quartiles (2021 USD). To develop a transportable prediction tool using routinely captured registry variables, ML models were trained in a complete-case SEER-22 cohort using age, sex, race group, stage, income group, rural–urban residence, and diagnosis period (≤2010 vs 2011+). We trained LASSO-penalized Cox (LASSO-Cox), random survival forest (RSF), and gradient-boosted Cox (XGBoost-Cox) models using a temporal split by year of diagnosis (training ≤2016; test > 2016) with cross-validated tuning. Performance was assessed by Harrell’s C-index, Uno time-dependent AUC (12/24/36/48 months), and Brier scores. External validation applied the locked SEER-22 models to an independent SEER-8 MF cohort with harmonized covariates at 12–60 months. Results: In the fully adjusted Cox model, Black vs White race was associated with worse OS (HR 1.76, 95% CI 1.60–1.93) and Asian/Pacific Islander vs White with improved OS (HR 0.81, 95% CI 0.66–0.98). Age and advanced stage were strong predictors of death, and male sex was modestly but significantly associated with poorer survival. A socioeconomic gradient was observed: lower-income counties ( < $50k) had higher hazards than the highest-income quartile (≥$95k), with stepwise decreases across higher income levels. ML modeling included 13,447 patients (train 9,849; test 3,598). Internal test discrimination was high (C-index: LASSO-Cox 0.835; RSF 0.818; XGBoost-Cox 0.824), with LASSO-Cox AUC 83.9–84.9% across 12–48 months. External validation (SEER-8, n = 4,753) showed good transportability for LASSO-Cox (C-index 0.757; AUC 76.6–78.2% across 12–60 months) and RSF (C-index 0.765), while XGBoost-Cox performed poorly (C-index 0.562; AUC 54.9–57.6%). Conclusions: Parsimonious ML survival models using routinely available demographic, clinical stage, and socioeconomic surrogates provide robust MF prognostication, with LASSO-Cox and RSF demonstrating strong external generalizability.