The prognostic predictive SER model for NK/T-cell lymphoma in the era of modern immunotherapy

As immune checkpoint inhibitors (ICI)-based combination therapies are increasingly explored for treating NK/T-cell lymphoma (NKTCL), there is a critical clinical need to identify patients who will benefit from ICI without relying on costly genomic testing.

A machine learning model was developed using routine blood tests and clinical characteristics from 364 ICI-treated NKTCL patients. The case records of 1259 NKTCL patients discharged from Sun Yat-Sen University Cancer Center, Guangzhou, between January 2018 and December 2023 were retrospectively analyzed. After screening, 364 ICI-treated patients were included in the study. These patients were randomly assigned to training (n = 254) and validation (n = 110) cohorts in a 2:1 ratio. Lasso regression and five machine learning algorithms, including random forest (RF), were applied for feature selection and clinical benefit prediction. The RF model demonstrated optimal predictive performance using five key features. To predict overall survival, we combined the RF model with two critical clinical factors-Ann Arbor stage and Eastern Cooperative Oncology Group (ECOG) performance status-to develop the stage-ECOG-RF (SER) model. This model generates a risk score to quantify the probability of poor survival following ICI treatment. In total, the SER model including seven features is significantly associated with clinical outcomes and long-term survival.

Five feature variables-lymphocyte count, platelet count, bone marrow involvement, cholesterol (CHO), and EBV-DNA copy number-were selected from 23 laboratory tests and clinical characteristics with complete data (0% missing rate). In the training cohort, the RF algorithm showed an area under the receiver operating characteristic curve (AUC) of 0.878, outperforming extreme gradient boosting (XGBoost), support vector machine (SVM), decision trees (DT), logistic regression and SVM algorithms. The RF model demonstrated sensitivity of 83.3% and specificity of 78.9%. In the validation cohort, the AUC of the RF model was 0.752, with sensitivity of 68.8% and specificity of 69.1%. The SER model, which integrates the RF model with Ann Arbor stage and ECOG, attained time-dependent area under the receiver operating characteristic curve (AUC(t)) values of 0.736 and 0.650 for predicting 3- and 5-year overall survival. This surpasses the prognostic index of natural killer lymphoma (PINK-E) and the net reclassification index (NRI) models, which showed AUC(t) values of 0.722 and 0.532, and 0.707 and 0.541 at 3 and 5 years, respectively.

Based on routine blood tests and clinical data, the SER model for ICI therapy of NKTCL-optimized with the RF algorithm and incorporating Ann Arbor stage and ECOG-demonstrates superior predictive performance compared to PINK-E and NRI. It provides a valuable reference for early prediction of ICI therapy failure and long-term survival.