Early vs late fusion of qMIDS with machine learning based classifiers for oral cancer diagnosis from gene expression data

Different fusion architectures of the quantitative Malignancy Index Diagnostic System (qMIDS) method with machine learning classifiers for oral cancer diagnosis from gene expression data are investigated. The proposed fusion architectures are tested on geographically and ethnically diverse datasets and the evaluation results showed performance improvements with late fusion, which significantly outperformed early fusion in more matched model configurations than the reverse in every cohort, although at the best-performing configuration the two architectures were statistically indistinguishable in most cohorts. Inter-dataset analysis showed robustness of the proposed architectures when the training/test samples are from different populations.