Abstract LB-251: Investigation of few-shot deep learning techniques on a small, high dimensional head and neck carcinoma dataset for prediction of cancer recurrence

Abstract

Abstract Introduction: Artificial Intelligence (AI) techniques may be useful in predictive oncology. We examined the feasibility of applying open source Few-shot learning to genomics data for the purpose of predicting head and neck cancer recurrences. We compared basic feed forward neural networks (vanilla networks) to a special architecture designed for few-shot learning called a Siamese network. Methods: Affymetrix U133A gene chip data (D=22,215, 41 tumor/13 normal) of patients with non-recurrent (46.34%) and recurrent (53.66%) head and neck carcinomas (nRHNC and RHNC, respectively) was analyzed. Shallow (one-layer) and deep (three-layer) variants were examined for both vanilla and Siamese networks to predict RHNC using the PyTorch framework. Four models were built using PyTorch: one shallow vanilla (SV) network with a single layer of 100 neurons, based on the top performing model from the authors' previous work, one deep vanilla (DV) network with layers of 256, 64, and 8 neurons, one shallow Siamese (SS) network with a single layer of 100 neurons, and one deep Siamese (DS) network with layers of 256, 64, and 8 neurons. Each model was initialized, trained, and evaluated 11 times. Kaiming initialization and Stochastic Gradient Descent with Nesterov Momentum (0.5) was employed with typical assumptions. Results: The model with the highest mean F1 score, AUC, and sensitivity was the SV network (0.7001, 0.6667, and 0.6800 respectively) the model with the highest sensitivity was DV (0.7217). When comparing the mean metrics for the following pairs: SV/DV, SS/DS, SV/SS, and DV/DS, the only difference of means that were found to be statistically significant was the SV/SS pairing for F1, AUC, and specificity. Conclusion: The simplest model, with the least representational capacity, performed best but not in a statistically significant manner in nearly all cases. Interestingly, the DS model performed better than the SS model AUC and specificity while the SV model performed better than its deep counterpart in those metrics. In this initial analysis, we conclude that it is feasible to apply PyTorch AI to Affymetrix gene chips to obtain meaningful data about head and neck cancer recurrences. Citation Format: Frank G. Ondrey, Dalton Schutte. Investigation of few-shot deep learning techniques on a small, high dimensional head and neck carcinoma dataset for prediction of cancer recurrence [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr LB-251.