Speaker
Description
Machine learning has proven highly effective in cancer classification using RNA expression data. These models are versatile and can be applied to various cancer types with minimal prior knowledge, provided sufficient training samples are available. However, machine learning models often function as black boxes, making result interpretation challenging. Traditional neural networks also lack reliable confidence estimates for their predictions. To overcome this limitation, we propose a probabilistic approach using Bayesian neural networks (BNNs). Unlike conventional models, BNNs quantify uncertainty, providing confidence estimates for individual classifications. Our results show that BNNs maintain high specificity even in noisy datasets or when key genes from the training set are missing, though with a trade-off in sensitivity. This robustness makes BNNs particularly valuable for clinicians using diverse gene expression methods, ensuring reliable predictions across varying conditions.