Training with Small Medical Data: Robust Bayesian Neural Networks for Colon Cancer Overall Survival Prediction

Fast and accurate cancer prognosis stratification models are essential for treatment designs. Large labeled patient data can power advanced deep learning models to obtain precise predictions. However, since fully labeled patient data are hard to acquire in practical scenarios, deep models are prone to make non-robust predictions biased toward data partition and model hyper-parameter selection. Given a small training set, we applied the systems biology feature selector in our previous study to avoid over-fitting and select 18 prognostic biomarkers. Combined with three other clinical features, we trained Bayesian binary classifiers to predict the 5-year overall survival (OS) of colon cancer patients in this study. Results showed that Bayesian models could provide better and more robust predictions compared to their non-Bayesian counterparts. Specifically, in terms of the area under the receiver operating characteristic curve (AUC), macro F1-score (maF₁), and concordance index (CI), we found that the Bayesian bimodal neural network (late fusion) classifier (B-Bimodal) achieved the best results (AUC: 0.8083 ± 0.0736; maF₁: 0.7300 ± 0.0659; CI: 0.7238 ± 0.0440). The single modal Bayesian neural network classifier (B-Concat) fed with concatenated patient data (early fusion) achieved slightly worse but more robust performance in terms of AUC and CI (AUC: 0.7105 ± 0.0692; maF₁: 0.7156 ± 0.0690; CI: 0.6627 ± 0.0558). Such robustness is essential to training learning models with small medical data.