MSc. Stats Project Presentation: Comparison of cox proportional hazards model and artificial neural networks on clinical data

CANDIDATE:   CHRIS WERNER-WESTOBY

ABSTRACT:

Artificial neural networks (ANNs) are nonlinear models that use networks of nodes. In survival analysis, traditional methods such as the Cox proportional hazards (CPH) model have been used much more than ANNs. The ANN models have potential advantages over the CPH which include accommodating nonlinearity and requiring fewer assumptions. This project compares the Cox proportional hazards ANN (CPHNN) and neural-multitask logistic regression (N-MTLR) to the CPH model on clinical datasets. The datasets are the bipolar dataset with the age of development for bipolar disorder, major depressive disorder or schizoaffective disorder as the response and the free light chain dataset with days until death as the response. On the smaller bipolar dataset, the N-MTLR did not fit the data well, while the CPHNN performed slightly better than the CPH using the C-index. The selection of ANN model was important. For the free light chain dataset, the models had similar C-index performance.