The aim of our study was to investigate the data hungriness of the various modelling techniques and the aim was not to find the best modelling technique in AUC terms. To our knowledge, the data hungriness of various modelling techniques has not been assessed before for medical prediction problems. However, a few studies addressed this topic in the context of progressive sampling for the development of a power law to guide the required sample size for prediction modelling. For example, arithmetic sampling was applied with sample sizes of 100, 200, 300, 400 etc. to 11 of the UCI repository databases to obtain insight into the performance of a naive Bayes classifier [8]. This study led to required sample sizes from 300 to 2180 to be within 2% from the accuracy of a model built from the entire database. Other researchers modelled 3 of the larger databases from the UCI repository using different progressive sampling techniques [9]. Using the C4.5 modelling technique, which we consider a CART variant, sample sizes of 2000 for the LED database, 8000 for the CENSUS database and 12000 for the WAVEFORM database were required for a model being no more than 1% less accurate than a model based on all the available data.