Non metallic inclusions are present in all continuously cast steels. They can affect the physical and mechanical characteristics of products and, consequently, their service properties. The amount and type of non-metallic inclusions have a direct influence on properties such as Welding, fatigue, ductility or corrosion resistance. Internal cleanliness is a main problem in the case of a electric steel making plant. In that case cleaning effort is done in the ladle furnace where the equilibrium between slag and steel bath is the key for the final result. Anyway, the transference between phases through the interface is complex, with several intermediate processes and not completely known. Also many properties needed for diagnose the reactions are very difficult to measure continuously, so they are just approximated with any kind of estimator or using simplifications. Those inconveniences are cause of problems and, in general, the estimation of the possible result, and in consequence the need of new additions, is reserved to a human expert. That situation is very negative due to the subjective of the technician and the dependence of his criteria. In order to avoid these problems, here is presented a new approach of estimation with the creation of a model based in historical data, capable to produce in-line the expected value of inclusions in the final strip from the steel composition and some process parameters. Model is created with different techniques, including Multiadaptive Regression Splines and Self Organizing Maps, has been trained with data from more than 2,000 heats and it has required several hundreds of analysis of steel samples in the ladle. Results show how that kind of model is capable to provide important information of oxide inclusions avoiding the need of a human expert and providing a very interesting tool for the development of an automatic control system of inclusions.