Statistical Model To Identify Factors Leading To Autism In Children With Application On Pediatric Hospital Ain Shams University

Abstract

Autism spectrum disorders(ASDs) represent a group of neurodevelopmental disorders characterized by impairments in verbal and non-verbal communication, social withdrawal and stereotypical behaviors, which may or maynot be associated with cognitive deficits, self-injurious behaviors and other neurological comorbidities. The increase of ASDs prevalence cannot be fully related to advances in diagnostics or sudden genetic shifts whereas there is a growing agreement among clinicians and scientists that ASDs resulted from an interaction between genetic and environmental factors. One environmental factor that has received great attention is the body burden of heavy toxic metals as mercury, lead and aluminum. Heavy metals exposure is an increasing global problem and many previous studies demonstrated that heavy metals induce deleterious effects in humans. This study drew a comparison between the results obtained on a given set of data gathered on a sample of Egyptian autistic children against age and sex matched healthy controls using different statistical and data mining techniques being represented in parametric and nonparametric methods which are Logistic regression, Discriminant analysis, Classification and regression tree (CART), Artificial neural network and Random forests to determine the possible risk factors that may lead to autism and to reach the ideal model to be used as a tool to predict autism occurrence. The variables studied in this thesis are House age, Age at conception, Consanguinity, Aluminum pans, Hair Lead level, Hair Mercury level, Dental amalgam, Passive smoking, Fish consumption and Gender. A split sample cross validation method is used to assess the validity of the classification models and it splits the sample into training and testing samples representing 75.7% and 24.3% of the total sample. The traditional statistical techniques applied in this study represented in Forward Stepwise logistic regression and Stepwise discriminant analysis revealed that they have the same final model variables which are: House age, Age at conception, Aluminum pans, Consanguinity, and Hair lead level.Also, The data mining non parametric statistical techniques represented in Artificial neural network, Classification and regression tree (CART) and Random forests revealed better classification accuracy where the results sobtained according to the training data set showed that the artificial neural network has the best performance in the establishment of the prediction model and its classification accuracy is 82.08%, it is the recommended classifier for the diagnosis of autism.Moreover, the Classification and Regression trees (CART) as well as Random Forest models have the same classification accuracy which is equal to 78.30%. While the traditional models represented in Forward Stepwise logistic regression and Stepwise discriminant analysis models have accuracy rate of 77.40% and 76.40%% respectively showing less classification accuracy than those of the nonparametric data mining techniques. Furthermore, the study showed that the classification accuracy of the cross validated group cases (Testing data set) for Logistic regression and Discriminant analysis is 76.50% and 79.40% while for Artificial Neural Networks, Classification and Regression Trees (CART) and Random forest is 79.41%, 76.47%, 79.41% respectively.

Key words: Autism, House age, Age at conception, Consanguinity, Hair Lead level, Hair Mercury level, Dental amalgam, Logistic regression, Discriminant analysis, Artificial Neural Networks, Classification and Regression Trees (CART) and Random forest.