Development and multicenter validation of a machine learning framework for predicting severe myelosuppression in nasopharyngeal carcinoma: Evidence from large-scale real-world data and prospective clinical trials.

Abstract

6023 Background: Severe treatment-related myelosuppression (Tr-MS) in nasopharyngeal carcinoma (NPC) necessitates dose reductions, compromising survival. Current static models rely on baseline "snapshots," failing to capture dynamic bone marrow fluctuations. We developed a dynamic deep learning system using longitudinal real-world data to predict cycle-specific Tr-MS risk. Methods: We analyzed 103,919 longitudinal records from 12,065 NPC patients (Nanfang/SYSUCC). A rolling-window strategy incorporated dynamic pre-dose labs, prior-cycle nadirs, and cumulative exposure to predict Grade ≥3 Tr-MS (CTCAE v6.0). Models (Logistic/XGBoost/LightGBM/TabPFN) were trained (6:3:1 split) and validated in two independent external cohorts (n=197) and three prospective trials (NCT03919552, NCT06767488, NCT06017895; n=173). The model was deployed as an EHR-integrated tool for automated real-time risk stratification. Results: Dynamic models demonstrated robust discrimination across all validation hierarchies (Table). In internal validation, AUCs ranged from 0.84 to 0.96. Performance remained stable in external cohorts (AUC 0.75-0.87) and prospective trials (AUC 0.79-0.86). SHAP analysis identified cumulative chemotherapy dosage and prior-cycle nadirs as dominant risk factors. Notably, the HIS-integrated tool successfully automated data retrieval, eliminating manual entry burden. Conclusions: This first EHR-integrated, cycle-specific dynamic system for Tr-MS in NPC captures longitudinal marrow exhaustion, enabling a shift from reactive rescue to precise, proactive prevention. Performance of dynamic prediction models across validation cohorts. Endpoint Cohort AUC (95% CI) Sens Spec Anemia Internal (n=1,231) 0.96 (0.96-0.98) 0.91 0.91 Anemia External (n=197) 0.87 (0.85-0.89) 0.89 0.86 Anemia Prospective (n=173) 0.86 (0.82-0.89) 0.88 0.86 PLT Internal (n=1,231) 0.90 (0.87-0.92) 0.79 0.84 PLT External (n=197) 0.80 (0.77-0.81) 0.81 0.82 PLT Prospective (n=173) 0.81 (0.80-0.83) 0.80 0.83 WBC/Neut Internal (n=1,231) 0.84 (0.82-0.85) 0.71 0.79 WBC/Neut External (n=197) 0.75 (0.73-0.77) 0.69 0.77 WBC/Neut Prospective (n=173) 0.79 (0.77-0.81) 0.69 0.78 Abbreviations: AUC, area under the ROC curve; PLT, thrombocytopenia; WBC/Neut, leukopenia/neutropenia; Sens, sensitivity; Spec, specificity.