In this study, we investigate a Hindmarsh-Rose-type model with the structure of recurrent neural feedback. The number of equilibria and their stability for the model with zero delay are reviewed first. We derive conditions for the existence of a Hopf bifurcation in the model and derive equations for the direction and stability of the bifurcation with delay as the bifurcation parameter. The ranges of parameter values for the existence of a Hopf bifurcation and the system responses with various levels of delay are obtained. When a Hopf bifurcation due to delay occurs, canard-like mixed-mode oscillations (MMOs) are produced at the parameter value for which either the fold bifurcation of cycles or homoclinic bifurcation occurs in the system without delay. This behavior can be found in a planar system with delay but not in a planar system without delay. Therefore, the results of this study will be helpful for determining suitable parameters to represent MMOs with a simple system with delay.2010 Mathematics Subject Classification. Primary: 34K18, 34K17; Secondary: 34K60. Key words and phrases. Hindmarsh-Rose-type model, Hopf bifurcation, differential-difference equation, mixed-mode oscillations, recurrent neural feedback. 37 38 SHYAN-SHIOU CHEN AND CHANG-YUAN CHENGTo understand the behavior of neurons with time delay, it is important to clarify the behavior of reduced neuronal models. We begin by reviewing the Hindmarsh-Rose (HR) model, which was first introduced in the 1970s. Connor et al. [11,12] were the first to establish a model for the alternative generation of action potentials. This model, which is similar to the classical four-dimensional Hodgkin-Huxley model [25], contains fast sodium, delayed rectifier potassium, leakage, and additional potassium conductance (i.e., the transient A-current). Rose and Hindmarsh [38] simplified the six-dimensional Connor-Stevens model to the two-dimensional Hindmarsh-Rose (2DHR) model by a transformation of variables. Furthermore, they extended the 2DHR model to a three-dimensional HR model with the addition of a slow variable to describe the subthreshold of the inward and outward currents. With suitable parameters, the models can simulate repetitive firing [23], bursting [24] and thalamic neurons [38] with detailed ionic currents, where repetitive firing is primarily induced through quadratic recovery. On a physiological level, the HR model can simulate the bursting neurons of the pond snail Lymnaea [7,6,9,8,40]. Zemanova [45] used the HR model to investigate the structural and functional clusters of a corticocortical network by modeling the cortical area of the network with a sub-network of interacting excitable neurons. Tsuji [42] proposed a 2DHR-type model that preserves both the time-scale parameter and the first component of the vector fields in the FitzHugh-Nagumo model [20,35] and showed that the model has properties of both Class 1 and Class 2 neurons. Chen [10] studied the number and stability of equilibria and codimension-two bifurcations in the 2DHR-type model and t...