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

Late fusion of qMIDS with machine learning classifiers for oral cancer diagnosis improves performance over early fusion across geographically and ethnically diverse gene-expression datasets.

• Why it matters: Fusion strategy can affect diagnostic accuracy, and prior uncertainty remains about which architecture generalizes across populations and dataset shifts.
• What they did: The study evaluates multiple qMIDS fusion architectures (early vs late) combined with machine learning classifiers on geographically and ethnically diverse datasets, including inter-dataset tests with different populations.
• The result: Late fusion significantly outperformed early fusion in more matched model configurations in every cohort, while at the best-performing configuration the architectures were statistically indistinguishable in most cohorts; robustness was shown when training/test samples came from different populations.