This paper addresses the problem of melt index (MI) control in the polymerisation of propylene in slurry reactors. Good MI control performance is usually hampered by measurement delays and a low sampling frequency. The traditional solution to the control problem is to control the hydrogen concentration in the reactor (H-2 serves as the chain transfer agent in the reaction and therefore has a good correlation with the MI). The H-2 concentration in the gas cap of the reactor is relatively easy to measure and therefore chosen to close a SISO control loop. However, this concentration is not always a good measure for the H-2 concentration in the slurry phase where the reactions occur. Consequently, MI variations can still be large at good control of H-2 in the gas cap. In the present paper, the use of a soft sensor for the H-2 concentration in the slurry phase is investigated by means of simulation. When used in combination with a model predictive control method, it is shown that plant performance significantly improves both at disturbances and at grade changes.