A diagnostic scoring model of ENKTCL in the nose-Waldeyer's ring based on logistic regression: Differential diagnosis
Jun-Yi Xiang1, Xiao-Shan Huang1, Na Feng2
1Department of Radiology, The Second Affiliated Hospital of Zhejiang Chinese Medical University, Hangzhou, Zhejiang, China.
Objective:
To establish a logistic regression model based on CT and MRI imaging features and Epstein-Barr (EB) virus nucleic acid to develop a diagnostic score model to differentiate extranodal NK/T nasal type (ENKTCL) from diffuse large B cell lymphoma (DLBCL).
Methods:
This study population was obtained from two independent hospitals. A total of 89 patients with ENKTCL (n = 36) or DLBCL (n = 53) from January 2013 to May 2021 were analyzed retrospectively as the training cohort, and 61 patients (ENKTCL=27; DLBCL=34) from Jun 2021 to Dec 2022 were enrolled as the validation cohort. All patients underwent CT/MR enhanced examination and EB virus nucleic acid test within 2 weeks before surgery. Clinical features, imaging features and EB virus nucleic acid results were analyzed. Univariate analyses and multivariate logistic regression analyses were performed to identify independent predictors of ENKTCL and establish a predictive model. Independent predictors were weighted with scores based on regression coefficients. A receiver operating characteristic (ROC) curve was created to determine the diagnostic ability of the predictive model and score model.
Results:
We searched for significant clinical characteristics, imaging characteristics and EB virus nucleic acid and constructed the scoring system via multivariate logistic regression and converted regression coefficients to weighted scores. The independent predictors for ENKTCL diagnosis in multivariate logistic regression analysis, including site of disease (nose), edge of lesion (blurred), T2WI (high signal), gyrus like changes, EB virus nucleic acid (positive), and the weighted score of regression coefficient was 2, 3, 4, 3, 4 points. The ROC curves, AUCs and calibration tests were carried out to evaluate the scoring models in both the training cohort and the validation cohort. The AUC of the scoring model in the training cohort were 0.925 (95% CI, 0.906-0.990) and the cutoff point was 5 points. In the validation cohort, the AUC was 0.959 (95% CI, 0.915-1.000) and the cutoff value was 6 points. Four score ranges were as follows: 0-6 points for very low probability of ENKTCL, 7-9 points for low probability; 10-11 points for middle probability; 12-16 points for very high probability.
Conclusion:
The diagnostic score model of ENKTCL based on Logistic regression model which combined with imaging features and EB virus nucleic acid. The scoring system was convenient, practical and could significantly improve the diagnostic accuracy of ENKTCL and the differential diagnosis of ENKTCL from DLBCL.
