Development and validation of a predictive model for extranodal natural killer/T-cell lymphoma

Current survival prediction model for extra-nodal natural killer/T-cell lymphoma (ENKTL) have poor accuracy. We developed and validated a machine-learning (ML) algorithm using clinical and pathological co-variates. Data from 977 subjects with ENKTL from 4 cohorts were analyzed. Model performance was evaluated using Harrell's c-index, receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCAs). The Gradient Boosting Machine (GBM) outperformed the other 15 ML algorithms tested, leading to the development of the ENKTL-ML score. C-indexes of the evaluation and external validation cohorts were 0.82 (95% Confidence Interval [CI], 0.76, 0.87), 0.84 (0.81, 0.88) and 0.83 (0.72, 0.94). The ENKTL-ML score effectively stratified subjects into 3 groups with distinct survival outcomes. Our model was more accurate than the IPI, KPI, PINK-E, and NRI (all P < 0.001) models. An online calculator is available at https://highcloud.shinyapps.io/ENKTL_ML_Scores/ . The ENKTL-ML score should help physicians predict survival of people with ENKTL.