Are Approaches of „Machine Learning“ and „Support Vector Machines“ Suitable to Improve the ASD Diagnostic Process in Children and Adolescents?

Background: Existing diagnostic instruments do, in fact, identify individuals with ASD accurately, but sometimes fail to differentiate individuals with ASD from those with other psychiatric disorders and complex neuro-behavioral profiles (such as ADHD, emotional and personality disorders and others). For these reasons, the identification of features which discriminate between different but overlapping phenotypes is of great importance. Recently, pattern classification methods based on machine learning algorithms have been used to predict or classify individuals of different phenotypes.

Objectives: The objective of the talk is to focus on both, machine learning algorithms and support vector machines to classify distinguishing dimensions of the ASD diagnostic process .

Methods: These innovative approaches were applied to the data of 2,568 children, adolescents and adults to identify those items of the applied diagnostic tools which show the best discriminatory quality. All patients underwent the gold standard diagnostic procedures, ASD diagnosis was confirmed in 1,359 individuals. In almost the same amount of patients (N= 1,209) ASD was excluded and a differential diagnosis was found (e.g. ADHD, language disorder). The outlined machine learning methods (Decision Tree) as well as Support Vector Machine analyses were used to develop algorithms that differentiate ASD from other disorders.

Results: For children and adolescents, a reduced number of differentiating items that exhibited good diagnostic accuracy was identified. Sensitivity ranged from 79 – 96%, specificity was found between 63 – 77%.

Conclusions: In sum, for young children accuracy was higher than for older. Social orientation in interaction behavior seems to be a core and specific symptom of ASD differentiating ASD from other disorders with multitude symptom overlap.