Predicting ASD Severity from Stereotypies Complexity Patterns through an Innovative Machine Learning System: A Proof of Concept Study.

Background: Stereotypies, despite their high frequency and strong diagnostic significance within autism, have not yet been fully elucidated due to their broad spectrum of presentation and pattern complexity. The VICTORY project (A Video Catalogue from Observational Retrospective Study on Stereotypies) a cross-sectional cohort study assessing presentation patterns, clinical Severity, and extinction modalities of stereotypies in autism, offers a new possibility to relate the complexity of stereotypies presentation to autism severity.

Objectives: The aim of this study is to assess the feasibility of predicting ASD severity in individual subjects from stereotypies patterns using innovative machine learning systems. This possibility would also enable further understanding as to which factors are significantly involved.

Methods: Twenty expert caregivers wearing a body cam recorded specific stereotypic behavior in a natural context during the everyday activities of 67 autistic subjects for 3 months of close follow-up. After a few minutes of recording, the possibility to interrupt their behavior by intervening physically to divert attention was recorded. A team consisting of one senior child neuro-psychiatrist together with a senior psychologist reviewed all the video recordings (1868) selecting 780 of them as the most meaningful to summarize the whole spectrum in each individual within the given time window. Each video was classified according to components (motor, sensorial, vocal, intellective), complexity (2 classes, simple and complex), body parts involved, sensory channels involved (hearing, sight, proprioception, taste, pain, smell), extinction modality and basic demographic features. Ninety-two variables were used to represent the input for preprocessing. The existence of a poor linear correlation among features of stereotypies patterns and ADOS score prompted us to use a machine learning system approach. An evolutionary algorithm (a TWIST system based on the KNN algorithm) was used to subdivide the dataset into training and testing sets as well as to select features yielding the maximum amount of information. After this pre-processing, 19 input variables were selected and different machine learning systems were used to develop a predictive model based on a training testing crossover procedure able to distinguish subject with an ADOS total score ranging from 8 to 20 from those with an ADOS total score ranging from 22 to 28.

Results: Acting on these inputs, the best supervised machine learning system(MLS) obtained a global accuracy of 84.96% (85.12 % - sensitivity and 84.79 % -specificity) in predicting the ADOS score class. Most of the stereotypies features selected by the algorithm were complex, with 2 or 3 different components in the same pattern among motor, sensorial, intellectual and vocal. A semantic connectivity map based on fourth generation unsupervised MLS depicted the association among high severity ADOS class with stereotypies made-up of 3 different components.

Conclusions: Machine-learning systems show a promising potential in highlighting the complex relationship between stereotypies patterns and ASD severity .