Thank you again for your reply. We have finally solved this problem by writing 32bits sequence before each operation.
Now we have some accuracy problem. For example, we measure with our calibrated multimeter 45,80mV but the code obtained from AD7793 (indeed the mean of 50 codes) is 49.90mV. We have this problem no matter the gain used.
We are trying full-scale calibration, zero calibration... all the possible ways but we always obtain this apparent offset between the "real" voltage and the converted voltage... any hint?