Modeling the covariates effects on the hazard function by piecewise exponential artificial neural networks: An application to a controlled clinical trial on renal carcinoma

Background: In exploring the time course of a disease to support or generate biological hypotheses, the shape of the hazard function provides relevant information. For long follow-ups the shape of hazard function may be complex, with the presence of multiple peaks. In this paper we present the use of a neural network extension of the piecewise exponential model to study the shape of the hazard function in time in dependence of covariates. The technique is applied to a dataset of 247 renal cell carcinoma patients from a randomized clinical trial. Results: An interaction effect of treatment with number of metastatic lymph nodes but not with pathologic T-stage is highlighted. Conclusions: Piecewise Exponential Artificial Neural Networks demonstrate a clinically useful and flexible tool in assessing interaction or time-dependent effects of the prognostic factors on the hazard function.