The diagnosis of loss of coolant accidents (LOCA) in nuclear reactors has attracted a great deal of attention in condition monitoring of nuclear power plants (NPPs) because the health of cooling system is crucial to the stability of the nuclear reactor. Multi-layer perceptron (MLP) neural networks have commonly been applied to LOCA diagnosis. The data used for training these models consists of a number of time-series data sets, each for a different break size, with the transient behavior of different measurable variables in the coolant system of the reactor following a LOCA. It is important to select a suitable architecture for the neural network that delivers robust results, in that the predicted break size is deemed to be accurate even for a break size that is not included in the training data sets. The objective of this paper is to present a simple method for measuring the robustness of diagnostic models for predicting the break size during the loss of coolant accidents. A robustness metric is proposed based on the leave-one-out approach and the mean squared error resulting from a diagnostics model. Using this metric it becomes possible to compare the robustness of different diagnostic models. Given data obtained from a high fidelity simulation of the coolant system of a nuclear reactor, four different diagnostic models are obtained and their properties compared and discussed. These models include a fully connected multi-layer perceptron with one hidden layer, a fully connected multi-layer perceptron with two hidden layers, a multi-layer perceptron with one hidden layer that is pruned using the optimal brain surgeon algorithm, a group method of data handling (GMDH) neural network, and an adaptive network based fuzzy inference system (ANFIS).