ADVANCES IN AUTISM INTERVENTIONS FOR CHILDREN WITH AUTISM SPECTRUM DISORDER
FOLINIC ACID IMPROVES VERBAL COMMUNICATION IN CHILDREN WITH AUTISM
CLUSTER RANDOMIZED TRIAL OF THE CLASSROOM SCRTS INTERVENTION FOR ELEMENTARY STUDENS WITH AUTISM SPECTRUM DISORDER
TEACHING PARENTS BEHAVIORAL STRATEGIES FOR AUTISM SPECTRUM DISORDER (ASD): EFFECTS ON STRESS, STRAIN AND COMPETENCE
EXTRA AXIAL CEREBROSPINAL FLUID IN HIGH RISK AND NORMAL RISK CHILDREN WITH AUTISM AGED 2 – 4 YEARS: A CASE CONTROL STUDY
BEYOND INFECTION: MATERNAL INMUNE ACTIVATION BY ENVIRONMENTAL FACTORS, MICROGLIAL DEVELOPMENT AND RELEVANCE FOR AUTISM SPECTRUM DISORDERS
THREE 2018 REPORTS: Prevalence of Autism Spectrum Disorder in United States:
Importance of monitoring early symptoms and severity in autism spectrum disorder
Advances in the genetics of autism
SOME PAGES WEB
There are many webs concerning the autism desorder, some examples:
Autism spectrum Australia
Wu Medical Center
MOLECULAR AUTISM Part of Sprinter Nature
Prediction of Autism at 3 Years from Behavioural and Developmental Measures in High-Risk Infants: A Longitudinal Cross-Domain Classifier Analysis
G. Bussu1 · E. J. H. Jones2 · T. Charman3 · M. H. Johnson2 · J. K. Buitelaar1 · the BASIS Team
© The Author(s) 2018. This article is an open access publication
Abstract We integrated multiple behavioural and developmental measures from multiple time-points using machine learning to improve early prediction of individual Autism Spectrum Disorder (ASD) outcome. We examined Mullen Scales of Early Learning, Vineland Adaptive Behavior Scales, and early ASD symptoms between 8 and 36 months in high-risk siblings (HR; n = 161) and low-risk controls (LR; n = 71). Longitudinally, LR and HR-Typical showed higher developmental level and functioning, and fewer ASD symptoms than HR-Atypical and HR-ASD. At 8 months, machine learning classified HRASD at chance level, and broader atypical development with 69.2% Area Under the Curve (AUC). At 14 months, ASD and broader atypical development were classified with approximately 71% AUC. Thus, prediction of ASD was only possible with moderate accuracy at 14 months. Keywords Autism · Early prediction · Machine learning · Data integration · Individual prediction · High-risk · Longitudinal study
Although symptoms of Autism Spectrum Disorders (ASD) typically emerge early in life, a reliable diagnosis is usually not achieved before age 3 or later (Steiner et al. 2012). Evidence suggests that the best prognosis for ASD currently lies in early targeted intervention aimed to improve later outcome by modifying emergent atypical developmental trajectories (Fernell et al. 2013; MacDonald et al. 2014). A recent
follow-up study on the effects of parent-mediated social communication intervention in infants at high familial risk of ASD between 9 and 14 months shows a treatment effect on symptom severity extended 24 months after intervention end (Green et al. 2017). However, the sustained delivery of behavioural intervention to all infants at risk for ASD based only on traits would be too expensive, and the risk/benefit ratio may be less favourable for infants who would have developed typically anyway. Thus, individual prediction of later development of ASD as soon as early signs emerge could help to better target early intervention strategies. Limits to detection of ASD before 24 months come from the high heterogeneity of the disorder and the relatively late emergence of the core characteristics of ASD. Heterogeneity in onset, aetiology, phenotype, neurobiology, and developmental trajectory points to multiple underlying processes acting together and leading to the disorder rather than a unitary biological process (Jones et al. 2014; Lai et al. 2013; Vorstman et al. 2017). Therefore, the investigation of different types of data is essential to capture the different aspects of the disorder. Machine learning holds the potential to provide a robust algorithm for prediction of later clinical outcome combining complementary information from different sources in an efficient way, and allowing the identification of
Electronic supplementary material The online version of this article (https ://doi.org/10.1007/s1080 3-018-3509-x) contains supplementary material, which is available to authorized users.
- G. Bussu firstname.lastname@example.org 1 Donders Institute for Brain, Cognition and Behaviour, Radboud University Medical Centre, Kapittelweg 29, 6525 EN Nijmegen, The Netherlands 2 Centre for Brain and Cognitive Development, Birkbeck, University of London, 32 Torrington Square, London WC1E 7JL, UK 3 Department of Psychology, Institute of Psychiatry, Psychology and Neuroscience, King’s College London, De Crespigny Park, Denmark Hill, London SE5 8AF, UK
Journal of Autism and Developmental Disorders
the most predictive combination of measures. The application of these methods has already shown promise for classification of children with ASD (Ingalhalikar et al. 2014; Uddin et al. 2013; Wee et al. 2014). In the present study we apply machine learning algorithms to predict clinical outcome at 36 months from different combinations of behavioural and developmental measures at 8 and 14 months. Despite a general consensus on the added value of data integration for prediction of ASD, the method has not previously been applied to behavioural measures and standard developmental assessments. Research in the early recognition and diagnosis of ASD has been focused on prospective longitudinal studies of infants at high-risk for autism because they have an older sibling with ASD. High-risk infants (HR) have about a 20% risk of developing ASD, significantly higher than the population prevalence of 1.5% (Christensen 2016; Ozonoff et al. 2011; Sandin et al. 2014; Szatmari et al. 2016), and thus the high-risk design allows us to study the early manifestations of the condition and understand the behavioural, cognitive and neural mechanisms that precede the clinical onset of ASD (Jones et al. 2014). Yet, these studies have mainly focused on average differences between infants who later develop ASD and their typically developing peers, measuring group differences by means of p-values. Convergent evidence supports the emergence of overt behavioural markers for ASD by the end of the second year of life, such as atypical eye contact, visual tracking, disengagement of visual attention and orienting to name (Elsabbagh and Johnson 2010; Gliga et al. 2014; Jones et al. 2014; Ozonoff et al. 2010; Rogers 2009; Yirmiya and Charman 2010; Zwaigenbaum et al. 2015). Objectively measured behavioural signs for ASD emerging before 12 months include a fall in fixation to the eye region between 2 and 6 months (Jones and Klin 2013), reduced gaze fixation to people at 6 months (Chawarska et al. 2013), and vocal atypicalities (Paul et al. 2011). But early markers for ASD are not limited to social domains. By 14 months, high-risk siblings developing ASD performed significantly worse than unaffected siblings on all scales of the Mullen Scales of Early Learning (MSEL) except for Visual Reception (Landa and Garrett-Mayer 2006), and impairments in verbal skills (particularly receptive language) (Barbaro and Dissanayake 2012), and motor skills were associated with later diagnosis of ASD (Chawarska et al. 2007; Landa and Garrett-Mayer 2006; Landa et al. 2013; MacDonald et al. 2013), even already by the age of 7 months (Leonard et al., 2014; Libertus et al. 2014). Few studies so far have conducted analyses that combined measures from different domains. Estes and colleagues (2015) investigated trajectories of developmental abilities, as measured by the MSEL; adaptive functioning, as measured by the Vineland Adaptive Behavior
Scales (VABS); and early ASD symptoms, as measured by the Autism Observational Scale for Infants (AOSI), in infants at high and low risk for autism in relation to ASD at 24 months, showing a pattern of symptoms starting in the sensorimotor domain at 6 months and moving to the social-communication domain after 12 months. Thus, prediction of autism may require a multi-measure approach. Although previous findings on group differences between infants who later develop ASD and their typically developing peers are valuable in terms of finding relevant biomarkers for the disorder, there is often substantial overlap between groups in individual variation, making prediction for individual infants difficult. The aim of individual prediction of outcome is to automatically classify each individual into one group (e.g. ASD vs. nonASD outcome), and performance is usually measured by accuracy or Area Under the Curve (AUC). The AUC is a measure of predictive accuracy computed as the area under the Receiver Operating Characteristic (ROC) curve, which is a plot of true positive rate vs. false positive rate for the model under evaluation (Metz 1978). Prediction at chance level results in 50% AUC, while prediction has moderate accuracy for AUC above 70%. Few studies have used behavioural measures to predict individual outcome of ASD, individual prediction being more focused on neuroimaging data (Arbabshirani et al. 2017; Emerson et al. 2017; Libero et al. 2015; Zhou et al. 2014). Macari and colleagues (2012) employed a decision-tree nonparametric learning algorithm to classify typical versus ‘atypical’ high-risk infants using as features measures from the Autism Diagnostic Observation Schedule (ADOS) at 12 months. ‘Atypicality’ included but was not limited to ASD, and was based on clinical evaluation at 24 months. Despite promising results, the study was considered preliminary due to a small sample size (n = 84) and the lack of a confirmatory diagnosis at 36 months. Chawarska and colleagues (2014) used the same methods to predict ASD outcome at 36 months in a cohort of high-risk siblings at 18 months. The aim was to identify the individual items of the ADOS-G at 18 months that best differentiated high-risk siblings who were going to develop ASD from typically developing siblings or siblings with other developmental disorders. The combination of six behavioural features (i.e. repetitive behaviours, eye contact, intonation, gestures, giving objects and spontaneous pretend play) allowed the identification of ASD with high accuracy (83%), while poor eye contact or limited gestures alone did not provide good prognostic value for ASD. This suggests that the interaction between (or combination of) individual behaviours must be considered to enhance predictive value for an early identification of later ASD outcome. Prior to our study, classification of ASD from behavioural measures before 12 months has not been reported, while it would be
Journal of Autism and Developmental Disorders
crucial to enable early intervention. Furthermore, previous studies only looked at items from the ADOS, but did not investigate whether different measures of developmental skills and functioning can increase predictive power for ASD at an early age. The aim of the present study was to investigate predictive longitudinal differences from 8 to 36 months between infants at low and high familial risk for autism with different developmental outcomes (typical, ASD, atypical). Further, we investigated whether we could predict ASD or atypical development at 36 months at an individual level within the HR group from data collected at 8 and 14 months. Extending the approach adopted in previous studies, we integrated measures from ASD symptoms, developmental and adaptive functioning, and we compared classifiers based on different combinations of measures to identify which combination is most predictive. We tested the hypothesis that integration of information about symptoms, developmental ability and everyday functioning can improve prediction of ASD compared to prediction from ASD-specific symptoms alone, capturing pervasiveness and addressing the high heterogeneity of ASD. Prediction was also made taking into consideration the dynamics of development by adding the change of scores between 8 and 14 months to cross-sectional measures at 8 months. This allowed us to test our second hypothesis that integration of measures from multiple time-points adds value to prediction of ASD from measures at early age compared to prediction from measures at single time-points.
Data presented in the current paper were collected as part of a large longitudinal study, to which 247 infants participated in one of two phases of longitudinal assessments (104 in Phase 1 and 143 in Phase 2). Data from 232 infants (161 [69.4%] high-risk siblings [HR] and 71 [30.6%] low-risk controls [LR]) were included in this study; ten infants were excluded because they did not receive an ADOS evaluation and/or a clinical outcome evaluation at 36 months; five infants were excluded because they did not attend at least one of the visits. HR infants were at increased familial risk because they had an older biological sibling with ASD, while LR controls had an older full sibling with typical development. The sample was balanced in gender (116 males and 116 females), and 85/161 HR siblings (53%) and 31/71 LR controls (44%) were males. We used imputation through expectation maximization to handle missing data (see Supplemental Material for details). Analyses were
performed on SPSS (http://www.ibm.com/analy tics/us/en/ techn ology /spss).
All infants, irrespective of diagnosis and risk group, were followed longitudinally on four visits from an intake evaluation at 8 months [mean = 8.1; standard deviation, SD = 1.2] with further assessments at 14 months [mean = 14.5; SD = 1.3], 24 months [mean = 25.4; SD = 3.1] and 36 months [mean = 38.4; SD = 2.3]. At each assessment, infants were evaluated on the MSEL and VABS. Autism symptoms were assessed through the AOSI at 8 and 14 months, while the ADOS was used at 24 and 36 months. The Autism Diagnostic Interview—Revised (ADI-R; (Kim et al. 2013)), a structured parent interview, was also used to assess autism symptoms at 36 months. Experimenters were aware of infants’ risk status, but assessments were blind to clinical outcome. At the time of enrolment, none of the infants had been diagnosed with any developmental condition.
Verbal and non-verbal cognitive development was measured at each visit by the MSEL (Mullen 1995), a standardized developmental measure used to assess cognitive functioning between birth and 68 months. Scores are obtained in 5 scales and 2 main functional domains: the gross motor scale (GM), and the cognitive scales. The cognitive scales are visual reception (VR), fine motor abilities (FM), receptive (RL) and expressive language (EL). The Mullen Scale provides normative scores for each specific scale (average T-score = 50, standard deviation SD = 10) and a single composite score representing general intelligence (Early Learning Composite, ELC; average standard score = 100, SD = 15). The T-scores from the five MSEL scales were included in this study.