Abstract
Background. Neuroblastoma (NB) is one of the most common malignant solid tumors in childhood, in some cases presenting challenges for differential diagnosis at initial presentation. Serum tumor markers – neuron-specific enolase (NSE) and lactate dehydrogenase (LDH) – have limited diagnostic accuracy. Circulating C18 lipoform of GD2 may represent a promising liquid biopsy marker for NB, however, multiparametric diagnostic models combining GD2 with serum markers remain largely unexplored. Objective – to develop and internally validate a logistic regression model for noninvasive diagnosis of NB in children based on the combination of serum C18 lipoform GD2, NSE, and LDH levels, and to assess whether a neural network model (multilayer perceptron) improves diagnostic performance compared to logistic regression. Materials and methods. Serum concentrations of C18 lipoform GD2 were measured using a validated high-performance liquid chromatography-tandem mass spectrometry method, NSE and LDH were determined by standard techniques in primary patients with suspected neurogenic tumor between July 2021 and March 2024 (32 months) at the Dmitry Rogachev National Medical Research Center of Pediatric Hematology, Oncology and Immunology. Missing NSE and LDH data were handled by multiple imputation (m = 20). A logistic model with three continuous predictors (GD2, NSE, LDH) was fitted in each imputed dataset, and coefficients were pooled using Rubin's rules. As an alternative, a neural network model (3 inputs, one hidden layer with 2 neurons, logistic output) was trained on the 6th imputation and applied to all 20 datasets. Model performance was assessed by AUC (95% confidence interval (CI)), as well as sensitivity, specificity, positive predictive value and negative predictive value for the logistic and neural network decision rules on the pooled multiply imputed dataset at the probability threshold determined by the Youden index. Internal validation was performed using strict 5-fold cross-validation on each imputed dataset. Differences in AUC between the two decision rules were assessed using DeLong's test. Results. This prospective single-center diagnostic cohort study included 155 patients aged 0–18 years with clinical and instrumental suspicion of a neurogenic tumor prior to specific therapy (89 with NB, 66 controls). The logistic model demonstrated an AUC of 0.842 (95% CI 0.829–0.856); only C18 lipoform GD2 remained a significant predictor (OR 1.07 per 1 nM; 95% CI 1.04–1.11; p 0.001). The mean AUC under strict 5-fold cross-validation was 0.838 (range – 0.812–0.867). The neural network decision rule showed a similar AUC on pooled data (0.841; 95% CI 0.828–0.855) but a lower and more variable mean AUC under cross-validation (0.759; range – 0.684–0.841). DeLong's test revealed no statistically significant difference between the AUCs of the logistic and neural network models (p = 0.904). Conclusion. Circulating C18 lipoform GD2 may serve as a key component of the diagnostic model in patients with suspected neurogenic tumor, while NSE and LDH contribute minimal additional value. The neural network model does not provide a significant AUC advantage and is less stable than logistic regression. Given the comparable discriminatory ability, greater robustness, and interpretability, the logistic regression model was selected as the final decision rule for noninvasive NB diagnosis and implemented as an online risk calculator.

This publication has 16 references indexed in Scilit: