# Differences in immune indicators among normal, high-risk, and esophageal cancer populations and development of a predictive model

## TL;DR

This study identifies immune differences between normal, high-risk, and esophageal cancer groups and builds a predictive model for early cancer risk assessment.

## Contribution

A novel immune-based machine learning model is developed for non-invasive early detection of esophageal cancer risk.

## Key findings

- EC grouping significantly impacts 54 immune metrics (Pillai = 0.585, P < 2.2×10^−16).
- 12 immune metrics, including Tcm CD4+ T cells and FB CD8+ T cells, show significant inter-group differences.
- The neural network model achieved 81.2% accuracy and AUC values above 0.90 for group differentiation.

## Abstract

To systematically compare the differences in immune characteristics among three populations (normal group, high-risk group, and esophageal cancer [EC] patients group) and construct a predictive model based on immune metrics, thereby exploring its value in EC risk assessment and early diagnosis.

A total of 440 participants were enrolled, including 173 normal individuals, 162 high-risk individuals, and 105 EC patients. Peripheral blood samples were collected, and 75 immune metrics were detected using flow cytometry. First, 21 highly correlated indicators were eliminated through the Pearson correlation matrix (|r|≥0.7), leaving 54 indicators for further analysis. Multivariate analysis of variance (MANOVA) was used to assess the effect of EC grouping on immune metrics, adjusting for gender and age. Significant indicators were analyzed using analysis of covariance (ANCOVA) and the false discovery rate (FDR). Stratified sensitivity analysis was conducted to verify the robustness of the results. Finally, a predictive model was built using 12 key immune metrics, and 10 covariates. The model performance was evaluated by 10-fold cross-validation, neural network algorithm, accuracy, Kappa value, and area under the receiver operating characteristic curve (AUC).

MANOVA showed that EC grouping had a significant overall impact on the 54-immune metrics (Pillai = 0.585, P < 2.2×10^−16). After ANCOVA and FDR correction, 12 immune metrics with significant inter-group differences were identified (P′ < 0.05), including Tcm CD4+ T cells, Naive CD8+ T cells, Tem CD8+ T cells, and FB CD8+ T cells. Stratified sensitivity analysis confirmed stable differences in key indicators (FB CD8+ T cells and Act NK cells) remained stable among subgroups. The neural network predictive model exhibited excellent discriminative efficacy: the overall cross-validation accuracy was 81.2%, the Kappa value was 0.709, and the AUC values for pairwise comparisons among the three populations were all higher than 0.90 (0.905 to 0.927).

This study demonstrates distinct and stable immune profiles across EC risk groups. The immune-based machine learning model effectively differentiates normal, high-risk, and cancer populations, offering a promising non-invasive tool for early risk assessment and screening of esophageal cancer.