Generating models of mental retardation from data with machine learning

The article focuses on generating simple and expressive domain models of Mental Retardation (MR) from data using knowledge discovery and data mining (KDD) methods. 2137 cases (mild or borderline MR) and 2165 controls (randomly selected) from the National Collaborative Perinatal Project (NCPP), a multicentric study involving pregnant mothers and the outcomes, constituted our sample. Twenty attributes (prenatal, perinatal and postnatal), thought to play a role in MR were utilized. The outcome variable (class) was, whether the child was retarded or not, based on the IQ score. Tree learners (C4.5, CART), rule inducers (C4.5 Rules, FOCL) and a reference classifier (Naive Bayes) were the machine learning algorithms used for model building. The predictive accuracy ranged from 68.4% (FOCL) to 70.3% (Naive Bayes). CART obtained a sensitivity of 79.0% and also generated highly stable and simple trees across fifty random two-third training), one-third (testing) partitions of the sample. The algorithms identified emotional/behavioral problems in children as a significant predictor of MR risk. Our study shows that the KDD methods hold promise in recovering useful structure from medical data.