Segregation is an important phenomenon which heavily affects the final mechanical properties of steel products. The presence of several complex physical phenomena resulting in final segregation pattern in as-cast products makes the quantitative prediction of macro-segregation for industrially relevant casting processes extremely difficult. In the present work, a reliable prediction of important rolled product quality (in terms of mechanical and Charpy impact properties) which are linked to segregation is achieved for plate rolled pipeline steels by exploiting data related to the as-cast structure and caster operational data (including casting machine condition) through the application of neural networks. In particular, a hierarchical approach is proposed for the prediction of the Charpy Impact Value, in order to reflect the physical link between this quantity and the Ultimate Tensile Strength. The neural predictor has been developed by exploiting real industrial data and its performance can improve through time by enlarging the database that is used for its training.