In the broad framework of condition-based maintenance, the final objective of bearing condition monitoring is to evaluate the health state and to estimate the remaining useful life of the bearings. The latter is a particularly challenging task, considering that remaining useful life is inextricably linked to a projection of what will happen in the future. Often, health indices, whose reliability relies on their effectiveness and consistency, are used for bearing condition monitoring. Most of the existing health indices pursue the property of monotonicity, but generally there are no obvious boundaries between the different health states of the bearings. Hence, it is quite difficult to give an objective and independent estimation of the health states, especially in real applications under different operating conditions and in the presence of noise. Furthermore, it is also difficult to set the failure threshold for a given health index when it is employed in different applications. In this paper, a new health index called 'MAC 2 PSD' is proposed, based on the moving-average cross-correlation (MAC 2) of the power spectral density (PSD) of the vibration signals. An interesting property of MAC 2 PSD is its capability to track the health condition and to discriminate clearly between the different health states. As shown in the paper, the MAC 2 PSD can also be used to estimate the remaining useful life by using its values during the defect-propagation phase. The effectiveness of MAC 2 PSD is shown by means of two different cases of bearing run-to-failure experimental data, from two different test rigs. Additionally, the capability to avoid false positives is evaluated by means of bearing vibration data measured on a locomotive in commercial service.