Computational Approach Improves Diagnostic Performance Based on Tumor-Associated Autoantibodies in Esophageal Squamous Cell Carcinoma.

Although tumor-associated autoantibodies (TAAbs) are promising biomarkers for early cancer detection, the diagnostic performance of individual TAAbs for esophageal squamous cell carcinoma (ESCC) remains limited. In this study, we constructed machine learning (ML) models for ESCC diagnosis using the serum levels of 6 TAAbs (c-myc, NY-ESO-1, p53, p62, RalA, and Sui1) and compared their performance with that of a conventional approach. When the training and test datasets were combined, serum samples from 339 patients with ESCC and 152 healthy controls were analyzed. To address class imbalance, the training data were oversampled using the synthetic minority oversampling technique. Classification models were developed using 5 ML algorithms, among which the svmLinear (support vector machine with a linear kernel), svmRadial (support vector machines with a radial basis function kernel), and neural network models demonstrated robust performance without overfitting. At specificities exceeding 90%, these models achieved substantially higher sensitivities than those obtained using any single TAAb. Notably, the svmLinear and neural network models achieved sensitivities greater than 40% for early-stage ESCC (stage 0/I) in the test dataset. Overall, the integration of multiple TAAbs with ML substantially improved the diagnostic performance for ESCC, highlighting the clinical potential of ML-based approaches.