Benchmarking Early, Latent and Decision Fusion of RNA Sequencing with Other Omics Layers for Lymph Node Metastasis Prediction in Head and Neck Cancer

Abstract

Lymph node spread is one of the strongest signs of poor outcome in head and neck squamous cell carcinoma, and small nodal deposits are often missed before surgery. We asked whether adding DNA methylation, copy number and protein data to RNA sequencing improves prediction of pathological node status from the primary tumour, and whether the way the layers are joined matters. We used 416 tumours from The Cancer Genome Atlas, of which 242 were node positive. Starting from RNA sequencing alone, we built six data combinations and tested each with three fusion strategies: early fusion of all features, latent late fusion of autoencoder codes, and decision fusion of single layer models. Six model families were compared: logistic regression, support vector machine, random forest, XGBoost, a multilayer perceptron, and MO-LSFN, our deep model that standardises the autoencoder codes, places them on a 32 by 32 map ordered by correlation with a mask for missing data, and classifies the map with a wide residual network. All 76 models used the same fivefold cross validation, and every step that learns from data was fitted on training patients only. Prediction was difficult for every model, with areas under the receiver operating characteristic curve between 0.555 and 0.649. Logistic regression on RNA sequencing alone reached 0.609. The best model, decision fusion with logistic regression on all four layers, reached 0.649 (95% confidence interval 0.591 to 0.705), but this gain was not statistically significant. Adding layers changed the mean score by less than 0.01, and latent fusion was slightly better than early fusion for conventional models (0.619 against 0.606). MO-LSFN gained the most from extra layers, rising from 0.556 on RNA alone to 0.628 on all four, and each of its parts helped in the ablation study, but it did not beat the simpler models. These results suggest that the molecular profile of the primary tumour carries only a weak signal for nodal spread, and that fusion design matters less than is often assumed.