A novel neural network-based survival analysis model
Neural Networks - 2003 Special issue: Advances in neural networks research — IJCNN'03
Double-blind evaluation and benchmarking of survival models in a multi-centre study
Computers in Biology and Medicine
| Hi-index | 0.00 |
Approximately 50% of all patients with intraocular melanoma die of metastatic disease, despite successful treatment of the primary tumour. The main factors associated with mortality include: tumour diameter; ciliary body involvement; extraocular tumour spread; epithelioid cell type; high mitotic rate; and chromosome 3 loss. We report the development of a web-based prognostic tool, which integrates all these factors. The cohort comprised 2655 patients (1369 male, 1286 female, mean age: 60.86 years) with histopathological data on 1282 patients and cytogenetic information on 405 patients. There were 871 deaths, 517 of which were from metastatic disease. A Conditional Hazard Estimating Neural Network (CHENN) model has been developed, and used to model the survival probability conditioned on observed clinical data. Such model is trained in the Bayesian framework, which allows model training, model regularization, model comparison and feature selection. The CHENN model is nonlinear, embeds the Cox proportional hazards model, and can correct Cox estimates in the cases with a nonlinear relationship between covariates and survival probabilities, and/or the proportional hazards assumption is not verified. As a result of these studies, we can confidently reassure many patients with intraocular melanoma that their survival probability is not significantly worse than that of the general population. This improves their well-being and avoids unnecessary and expensive screening tests. Conversely, we can reliably identify those with a high risk of metastatic death. These patients are referred to an oncologist for systemic screening and several have undergone partial hepatectomy with significant prolongation of life. Our prognostication enhances the feasibility of future randomized, prospective studies evaluating protocols for systemic screening and adjuvant therapy.