New prognostic models for extranodal natural killer T-cell lymphoma, nasal-type using Cox regression and machine learning

Jianbin Sun; Xiaoyan Ke; Mingzhi Zhang; Yuhui Wang; Fengyang An; Yulin Zhao; Li Zhu

doi:10.21037/tcr-20-3017

New prognostic models for extranodal natural killer T-cell lymphoma, nasal-type using Cox regression and machine learning

Transl Cancer Res. 2021 Feb;10(2):613-626. doi: 10.21037/tcr-20-3017.

Authors

Jianbin Sun¹, Xiaoyan Ke², Mingzhi Zhang³, Yuhui Wang¹, Fengyang An¹, Yulin Zhao⁴, Li Zhu¹

Affiliations

¹ Department of Otorhinolaryngology, Head and Neck Surgery, Peking University Third Hospital, Beijing, China.
² Department of Hematology, Peking University Third Hospital, Beijing, China.
³ Department of Oncology, the First Affiliated Hospital of Zhengzhou University, Zhengzhou, China.
⁴ Department of Rhinology, the First Affiliated Hospital of Zhengzhou University, Zhengzhou, China.

Abstract

Background: The prognostic index of natural killer lymphoma (PINK) is recommended for use as a prognostic model for determining the best non-anthracycline-based treatment for extranodal natural killer T-cell lymphoma, nasal-type (ENKTL). However, this model does not provide an accurate individual risk estimation for patients; therefore, our retrospective study was conducted to determine this risk.

Methods: Clinical data from 250 patients with ENKTL treated with non-anthracycline-based regimens were analyzed. The statistically significant clinical characteristics were selected as the parameters for our models. The patient data from 250 patients were randomly divided into 5 groups for 5-fold cross validation before final models were established on all of the patients' data. A statistical model nomogram based on a Cox proportional hazards model, and a machine learning model based on the lightGBM algorithm, were constructed. Concordance index (C-index) and calibration curve, areas under the curve (AUC) values, and binary error were used to evaluate two models.

Results: Five variables [age, the Chinese Southwest Oncology Group and Asia Lymphoma Study Group ENKTL (CA) staging system, Eastern Cooperative Oncology Group (ECOG) score, B symptoms, and lactate dehydrogenase (LDH)] were significant and were selected as parameters for creating the statistical model nomogram, while lesion sites (anatomical regions, lymph nodes and primary lesion site) took place of CA staging system in machine learning model. During cross validation, the mean C-indices of training cohort and validation cohort for statistical model nomogram and PINK were 0.851±0.008, 0.843±0.029, 0.758±0.019 and 0.757±0.080, respectively, while the mean 3-year AUC for machine learning model were 0.920±0.010 and 0.865±0.035, respectively. The calibration curves and binary errors showed a good correlation between the predicted result and the reality.

Conclusions: These two models could provide ENKTL patients with an accurate individual risk estimation in the era of non-anthracycline-based treatment.

Keywords: CA staging system; Extranodal natural killer T-cell lymphoma, nasal-type (ENKTL); machine learning; nomogram; non-anthracycline-based treatment.