Does Fusion Stage Matter? Comparing Feature-Level and Decision-Level Multimodal Learning for Lung Cancer Subtyping

Lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC) have distinct molecular and histopathological characteristics, making accurate subtyping important for understanding disease biology and guiding treatment. In this study, we examine how the stage of multimodal integration affects LUAD-LUSC classification using RNA-Seq, DNA methylation, histopathology whole-slide images (WSI), and clinical data from The Cancer Genome Atlas (TCGA). We used a matched cohort of 823 patients and compared feature-level early fusion with decision-level late fusion across several combinations of these modalities. To make the comparison fair, we also evaluated an equal-feature setting in which 40 features were retained from each modality. Six classical machine-learning classifiers were assessed using accuracy, macro-F1, AUC-ROC, sensitivity, and specificity. RNA-Seq alone provided very strong discrimination, achieving 97.55% accuracy and an AUC-ROC of 0.9992, suggesting that transcriptomic information carries much of the signal needed to distinguish LUAD from LUSC. Multimodal integration did not consistently outperform this RNA-Seq baseline. However, the fusion strategy became more important as additional heterogeneous data were introduced. In the full four-modality setting, late fusion achieved 97.96% accuracy compared with 96.33% for early fusion under the equal-feature setting. Ablation analysis further showed that removing RNA-Seq produced the largest decline in late-fusion accuracy. Overall, these results suggest that the value of multimodal learning depends not only on which data modalities are included, but also on how they are combined, particularly when one modality already provides highly discriminative information.