Detecting Autistic Traits using Computational Intelligence & Machine Learning Techniques

Autistic Spectrum Disorder (ASD) is a developmental disorder that describes certain challenges associated with communication (verbal and non-verbal), social skills and repetitive behaviours. Self-administered ASD assessment tools, also known as screening tools, are typically conducted by a caregiver, medical staff and require responses to a large number of items. The validity and accuracy of assessments based on these tools relies upon classification methods which have antiquated technologies and this should be of concern for users in the healthcare community. A possible way to improve the classification accuracy and efficiency of the current screening tools is to adopt intelligent methods based on machine learning (ML) and computational intelligence. The latter can be utilised to identify a concise set of items by using new technologies such as mobile platforms, thus improving screening, or be able to steer those in seek of help toward a more accurate diagnosis. To automate the classification process and enhance the predictive accuracy of the test, the processing of data, based on the outcome of the computational intelligence, can be conducted using the former method.

This thesis proposes a new ML architecture for ASD screening that consists of a rule-based classification method called Rules Machine Learning (RML) which generates high predictive rules that can be easily understood by different users. Moreover, a new feature selection method known as Variable Analysis (Va) is proposed; this significantly reduces the number of features needed for ASD screening methods while maintaining performance. The last proposal in this thesis is an easy and accessible mobile screening application called ASDTests, which enables vital autism features to be collected from three primary datasets: adults, children, and adolescents from which thorough descriptive and predictive analyses are performed. To measure the performance of the RML and Va methods, large numbers of experiments have been conducted using various feature selection and ML techniques on the considered datasets. The bases of the comparisons are: evaluation metrics including sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NVP), and harmonic mean. The results clearly demonstrated that the new ML method was able to choose fewer items from the three datasets than the other methods considered while maintaining acceptable levels of specificity, sensitivity, and predictive accuracy. The concise sets of items and classifiers generated are of high interest to the different individuals interested in ASD screening. These results can also assist in early detection of ASD traits, thus facilitating access to necessary support systems for the physical, social, and educational well-being of the patient and their family in addition to increasing the likelihood of improved outcomes for the patient.