The objective of the present study is to synthesize and investigate using the same set of sixty-one real-world accidents the human head injury prediction capability of the head injury criterion (HIC) and the head impact power (HIP) based criterion as well as the injury mechanisms related criteria provided by the simulated injury monitor (SIMon) and the Louis Pasteur University (ULP) finite element head models. Each accident has been classified according to whether neurological injuries, subdural haematoma and skull fractures were reported. Furthermore, the accidents were reconstructed experimentally or numerically in order to provide loading conditions such as acceleration fields of the head or initial head impact conditions. Finally, thanks to this large statistical population of head trauma cases, injury risk curves were computed and the corresponding regression quality estimators permitted to check the correlation of the injury criteria with the injury occurrences. As different kinds of accidents were used, i.e. footballer, motorcyclist and pedestrian cases, the case-independency could also be checked. As a result, FE head modeling provides essential information on the intracranial mechanical behavior and, therefore, better injury criteria can be computed. It is clearly shown that moderate and severe neurological injuries can only be distinguished with a criterion that is computed using intracranial variables and not with the sole global head acceleration.