Abstract
In recent years, classifiers have been shown to provide highly accurate results in predicting problems; however, the success relies on the classifier as well as the employed dataset. In this study, the success of various classifiers for predicting autism spectrum disorder cases is analysed in terms of different metrics obtained both after testing on the same set and after applying 10-fold cross validation. When the same dataset has been used for the training and testing steps, all the algorithms have shown excellent results. Nevertheless, when 10-fold cross-validation has been applied, sequential minimal optimization has become the most successful classifier in terms of all the performance metrics. Although the results have proven the success of classifiers in predicting problems such as the one addressed in this study, autism spectrum disorder is a complex disorder, and the tools employing classifiers should be used under the supervision of certified professionals or clinicians.
TopIntroduction
Autism spectrum disorder (ASD) is a complex developmental disability with two main characteristics: (i) difficulty with social interactions and communication and (ii) repetitive behaviours, interests, and activities (American Psychiatric Association, 2013). As well as adults and adolescents with ASD, children suffering from this disorder also exhibit restricted and repetitive patterns of behaviours, interests, or activities and experience significant difficulties in social communication and interaction. In boys, indications are typically recognised around one to two years of age (McPartland & Volkmar, 2012). Late diagnosis and over-diagnosis are common problems particularly among girls; many children with ASD are not finally diagnosed until they are older. Sometimes girls with ASD are not diagnosed until late in their development (Gould & Ashton-Smith, 2011). Early diagnosis allows early intervention and thereby provides better outcomes (Crais & Watson, 2014). On the other hand, sometimes indications similar to the ones found in ASD can occur to a moderate degree when children are stressed, depressed and anxious due to some reasons (White et al., 2009). For instance, children may exhibit restricted interests and social avoidance when they have dyslexia, speech and language disorders, and intellectual disabilities. There is diversity in the type and severity of ASD indications (American Psychiatric Association, 2013). Therefore, while some individuals in the severe range may need substantial support in their daily lives, some in the mild range may live independently. Although treatment efforts are generally individualised and can involve behavioural therapy and teaching of coping skills, psychotropic medications are widely utilised in alleviating associated behavioural and emotional indications (Accordino et al., 2016).
The Autism Diagnostic Interview-Revised (ADI-R) (Lord et al., 1994), consisting of 93 questions, is the most established and one of the most widely used autism-specific diagnostic interviews (Pasco, 2010). In this interview, an overall diagnostic algorithm collects scores from items in a number of domains relating to communication, social interaction, repetitive, restricted and stereotyped behaviours and the age of onset, and generates outcomes relating to The Diagnostic and Statistical Manual of Mental Disorders (DSM) and The International Classification of Diseases (ICD) definitions of core autism only (Lord et al., 1994). Autism Diagnostic Observation Schedule (ADOS) (Lord et al., 2000), a semi-structured interactive assessment conducted by a certified professional, is the best validated and most widely used diagnostic assessment. It consists of four modules and in accordance with the age and expressive language level of the case being assessed, the appropriate module is selected. Module 1 of ADOS is suitable for young nonverbal children. Module 4 is suitable for verbally fluent adolescents and adults. Modules 2 and 3 are suitable for the remaining groups (Pasco, 2010). Cut-off scores obtained when the appropriate module has been applied are the indicator of whether outcomes fulfil criteria for ASD (Lord et al., 2000). Nevertheless, the diagnostic algorithm of ADOS does not take scores of items relating to restricted, repetitive and stereotyped behaviours into consideration. Consequently, information from other sources is required for a full clinical diagnosis.
In this study, it is aimed to analyse the success of various classifiers for predicting ASD cases in terms of different metrics. It is known that when the same dataset is used for training and testing steps, unrealistic results may be obtained (Varoquaux & Cheplygina, 2022). However, applying cross validation generally leads to more realistic results that can be generalised (Korjus et al., 2016). Therefore, both approaches are to be used in order to make a fair comparison between the classifiers.
Key Terms in this Chapter
Virtual Reality: It is the use of computer modeling and simulation that enables an individual to interact with an artificial three-dimensional visual or other sensory environment.
Autism Spectrum Disorder: It is a neurological and developmental disorder that affects how people interact with others, communicate, learn, and behave.
SMO: It is an algorithm used to solve the quadratic programming problem that arises during the training of support vector machines.
Classification Algorithms: They are used to categorise data into a class or category.
Performance Metrics: They are used to evaluate machine learning models and measure their effectiveness. They help to determine whether the machine learning model is making accurate predictions, and how well it is generalising to new data.