Differential diagnosis of pediatric cervical lymph node lesions based on simple clinical features

This study aims to to establish a diagnosis model based on simple clinical features for children with cervical histiocytic necrotizing lymphadenitis or malignant lymphoma. Simple clinical features of pediatric patients were analyzed to develop a diagnosis model based on a comparison of classical machine-learning algorithms. This was a single-center retrospective study in a tertiary pediatrics hospital. Pediatric patients treated for cervical histiocytic necrotizing lymphadenitis or malignant lymphoma treated at our institution in recent 5 years were included. Demographic data and laboratory values were recorded and binary logistics regression analysis was applied to select possible predictors to develop diagnostic models with different algorithms. The diagnostic efficiency and stability of each algorithm were evaluated to select the best one to help establish the final model. Eighty-three children were included with 45 cases of histiocytic necrotizing lymphadenitis and 38 cases of malignant lymphoma. Peak temperature, white blood cell count, monocyte percentage, and urea value were selected as possible predictors based on the binary logistics regression analysis, together with imaging features already reported (size, boundary, and distribution of mass). In the ten-round random testing sets, the discriminant analysis algorithm achieved the best performance with an average accuracy of 89.0% (95% CI 86.2-93.6%) and an average AUC value of 0.971 (95% CI 0.957-0.995). Conclusion: A discriminant analysis model based on simple clinical features can be effective in differential diagnosis of cervical histiocytic necrotizing lymphadenitis and malignant lymphoma in children. Peak body temperature, white blood cell count, and short diameter of the largest mass are significant predictors.