AUTHORS
Bo Yang, BS, Jingkun Li2, Yan Zhuang3, Haibo Cheng4, Jing Luo5, Yushan Li6, Yuxiang Wang7, Samantha M Thomas8, Jingke Yang, MD, PhD9, Jennifer K Plitcha10, Baixin Ye, MD11
AFFILIATIONS
1 Duke University School of Medicine, Durham, North Carolina, United States
2 School of Public Health, Harbin Medical University, Harbin, Heilongjiang, China
3 Duke University School of Medicine, Durham, North Carolina, United States
4 Virginia Polytechnic Institute and State University, Blacksburg, Virginia, United States
5 School of Mathematical and Statistical Sciences, Arizona State University, Tempe, Arizona, United States
6 Bouvé College of Health Sciences, Northeastern University, Boston, Massachusetts, United States
7 Wuhan Britain-China School, Wuhan, China
8 Duke Cancer Institute, Duke University School of Medicine, Durham, North Carolina, United States
9 Academy of Clinical Research and Study, Washington, District of Columbia, United States
10 Duke Cancer Institute, Duke University School of Medicine, Durham, North Carolina, United States
11 Department of Hematology, Renmin Hospital of Wuhan University, Wuhan University, Wuhan, Hubei, China
ABSTRACT
Background
The International Prognostic Index (IPI) remains a widely used clinical tool for risk stratification in diffuse large B-cell lymphoma (DLBCL). However, the composite IPI score summarizes a limited set of clinical features and may omit prognostically relevant information, including comorbidity burden, demographic factors, and socioeconomic context. Although machine learning (ML) survival models have shown improved prognostic performance, it remains unclear whether these gains reflect algorithmic complexity or the incorporation of a richer predictor set. We sought to compare a composite IPI-score Cox reference model with conventional and ML-based survival models to assess the relative contribution of variable completeness versus algorithmic complexity to overall survival prediction in DLBCL.
Methods
We identified 32,312 patients with newly diagnosed DLBCL from the National Cancer Database (NCDB) between 2004 and 2018. The primary outcome was overall survival, defined as time from diagnosis to death from any cause or last follow-up. Missing predictor data were handled using multiple imputation. The cohort was randomly divided into a 70% training set and a 30% held-out test set. Six survival models were developed: composite IPI-score Cox, multivariable Cox, Elastic Net Cox, Random Survival Forest (RSF), XGBoost survival, and Bayesian Cox. All non-IPI models used the same 18-variable predictor set, including clinical, comorbidity, demographic, and socioeconomic variables. Model performance was evaluated in the held-out test set using Harrell’s C-index, time-dependent AUC, calibration, decision curve analysis, and continuous net reclassification improvement (NRI) and integrated discrimination improvement (IDI). Exploratory age-stratified analyses were performed in patients aged <60, 60–69, and ≥70 years.
Results
The composite IPI-score Cox model achieved a C-index of 0.644 (95% CI, 0.626–0.661). Expanding the predictor set within a conventional multivariable Cox model improved discrimination to 0.708 (95% CI, 0.692–0.723), representing a gain of +0.064 over the IPI reference model. Elastic Net Cox achieved identical discrimination (C-index, 0.708), supporting the robustness of the linear benchmark. RSF achieved the highest point estimate for discrimination (C-index, 0.715; 95% CI, 0.700–0.730), but the incremental gain beyond multivariable Cox was small (+0.007). XGBoost survival and Bayesian Cox achieved C-indices of 0.685 and 0.704, respectively. Continuous NRI relative to the IPI reference model was highest for RSF (0.607; 95% CI, 0.568–0.646), while multivariable Cox also showed substantial improvement (0.508; 95% CI, 0.471–0.545). Exploratory age-stratified analyses showed smaller differences among models in the 60–69-year stratum and larger separation in younger and older patients, suggesting that age representation may contribute to predictive information loss. Across models, continuous age, composite IPI score, comorbidity burden, and socioeconomic variables consistently ranked among important predictors.
Conclusions
In this large NCDB cohort of patients with DLBCL, most of the observed improvement in overall survival prediction over a composite IPI-score Cox reference model was captured by expanding the predictor set within a conventional Cox framework. More complex ML algorithms provided only modest additional discrimination beyond the full-variable Cox model. These findings suggest that modernization of DLBCL prognostic tools should prioritize complete, clinically meaningful, and interpretable predictor domains before increasing algorithmic complexity. Because the outcome was all-cause overall survival, these models should be interpreted as overall survival prediction tools rather than disease-specific biology or treatment-intensification tools. External validation is required before clinical implementation.
DISCLOSURES
Bo Yang, BS: No relevant conflicts of interest to declare.