Abstract 1094: Harnessing neural networks for cancer prediction: clinical staging of head and neck cancer

Abstract Introduction: Predictive oncology through the use of deep learning offers potential to both classify and assist in the treatment of head and neck cancer, a disease that has a relatively static rate of cure. More specifically, neural networks may allow for the classification of genomic expression data in a more complex manner. The present study applies neural networks to the larger NCI Cancer Genome Atlas (TCGA) and our previously published smaller U133A Affymetrix gene expression dataset. Methods: The models were constructed using Pytorch, built from modular blocks of neural networks (Chunks) to facilitate extensive combinations of layers. Every Chunk consisted of a linear layer and optional predefined layers: rectifier (ReLU), normalization, and random neuron dropout (to prevent overfitting). The DeepChunk model (consisting of multiple Chunks) had additional tunable hyperparameters (optimization function, learning rate, etc.) The neural networks were trained to classify T and N staging data separately for both TCGA (split by HPV status) and Affymetrix chip datasets between early (T1/2) versus advanced (T3+) T stages, N0 versus N1+, respectively. Models were trained on larger datasets (HPV- T and HPV N, n >= 350) using randomly split data with 80% reserved for training and 20% reserved for testing, with proportional staging data across the split. The remaining datasets (n <= 50) had a split of 70% training and 30% testing. Hyperparameters (tuned for both the model and specific chunks) were assessed (by test accuracy and area under receiver operator characteristic curve, AUC-ROC) on the whole gene expression data. Settings selected during the hyperparameter screen were assessed with 20 trials using randomized initialization and data splits. Results: The models (test accuracy [95% CI], AUC-ROC [95% CI]) trained on the TCGA HPV- N (0.619 [0.601, 0.637], 0.63 [0.602, 0.659]) and TCGA HPV+ T (0.72 [0.672, 0.768], 0.733 [0.671, 0.794]) datasets performed significantly better than the other models trained. The remaining models failed to significantly predict above random chance (with 95% CIs overlapping with 0.5). Unstable performance from these models could be attributed to overfitting or quickly converging during training with lack of improvement afterwards. Another potential cause involves high variance from the smaller size of training data. Conclusions: The process used to tune and train the neural networks harbors potential in the analysis of large gene datasets. At the same time, it may benefit from feature reduction or functional annotations as the one dimensional feature lists are not inherently ordered or clustered, as is the case with two dimensional image neural networks. Citation Format: Amrit Menon, Nathan DeMichaelis, Dalton Schutte, Frank G. Ondrey. Harnessing neural networks for cancer prediction: clinical staging of head and neck cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 1094.