Background.  Linear and nonlinear decision models have been explored for discovery of protein patterns as biomarkers.  Often these models have similar performance but low correlation. The objective of this study was to investigate the feasibility of combining such models (i.e., fusion model) to potentially improve decision performance.  To investigate the feasibility of a fusion decision strategy for proteomic biomarker discovery, we profiled the proteins in 43 serum samples collected from 21 patients with heart failure and 22 cardiac patients with adequate left ventricular ejection fraction as controls.  Samples were processed without fractionation using hydrophobic H4 protein chips and analyzed by surface enhanced laser desorption ionization mass spectrometry (SELDI) (Ciphergen Biosystems, Fremont, CA).  Using peak detection and processing algorithms provided in the Biomarker Wizard Software, 19 protein peaks were identified all having signal to noise ratios of >3.0.  Next, stepwise linear peak selection was performed to identify the most discriminatory peaks.  The process resulted in 4 peaks subsequently merged using two different decision models. Specifically, a linear discriminant model was developed to predict the presence of CHF based on the 4 peaks.  The same peaks were also merged using an artificial neural network to potentially capitalize on diagnostically important data nonlinearities.  The diagnostic performance of both decision models was assessed using Receiver Operating Characteristics (ROC) analysis.  Results.  Based on the leave-one out data sampling scheme, both decision models had similar ROC performance with area indices of 0.66 and 0.67 respectively.  On the other hand, the fusion decision strategy resulted in a decision model with statistically significantly better ROC area index of 0.88 which supports our hypothesis.   Although the performances of both decision models were comparable, correlation analysis (r = 0.49) showed that they had low case agreement and each model had an advantage over the other in different parts of the case spectrum.     Conclusion.  A fusion neural network that combined the linear and nonlinear models boosted the decision-making performance. Therefore, combining decision models that do not think alike is a promising decision strategy for proteomic data analysis.