26570
Outcome Evaluation of Personalized Multidimensional Interventions on Children with Low-Functioning ASD through an Innovative Machine Learning System: A Proof of Concept Study.
Evaluating treatments outcome in children with low-functioning autistic disorder requires the utilization of specific but manageable instruments, both for patients and for their environment (parents, educators, doctors). Few studies so far have focused how use multidimensional data for outcome assessment in residential health settings.
Objectives:
The aim of this study is to highlight the possible outcome prediction of personalized plans of intervention for low-functioning ASD subjects using innovative machine learning systems enabling also to understand which treatment factors are significantly involved.
Methods:
In this pilot observational study, twelve consecutive new patients with low-functioning autism (range of age 3-13 years) have been enrolled between November 2015 and October 2016. Four complementary assessment instruments (Vineland Adaptive Behavior Scales, a 540-item questionnaire which evaluate personal and social autonomy, communication and motoric competences; SDQ-Strenghts and Difficulties Questionnaire-, a 25-item questionnaire useful to screen emotional, behavioral and social problems in children aged 4-16 years; The HoNOSCA-Health of Nation Outcome Scale for Children and Adolescents-, a 15-item clinical assessment scale used as part of the routine outcome monitoring in mental health services, which measures global functioning in patients aged 3-18 years through 4 different areas: behavioral, impairment, symptoms, social functioning; DC-GAS -Disability Child Global Assessment Scale-, a dimensional scale used by the clinician to evaluate global functioning in disabled children and adolescents) have been used at the patients first access in the neuropsychiatric clinic and after 6 months of intensive personalized treatment. Vineland Scales and SDQ questionnaire have been completed by educators and parents; HoNOSCA and DC-GAS have been completed by the clinician. Increase of at least 4 points in the DC-GAS total score between baseline evaluation and the assessment after 6 months the has been individuated as main outcome measure. Ninety four variables related to demographic, familiar, therapeutic, pharmacological, medical and checkup information represented the input for preprocessing. An evolutionary algorithm (TWIST system-Semeion) has been used to subdivide the dataset into training and testing set and select features yielding the maximal amount of information. After this pre-processing, 21 input variables were selected and different machine learning systems have been used to develop a predictive model based on a training-testing crossover procedure.
Results:
Eight out of twelve subjects have shown an improved global functioning at the end of the follow-up. The best machine learning system (three-layers feed- forward neural network with 8 hidden nodes) obtained a global accuracy of 83.3% ( 91.7 % sensitivity and 75% specificity ) with a ROC of 0.89. The variables selected for the predictive model included previous pharmacological and non-pharmacological treatments, actual treatment plan, and baseline scores of different subscales of Vineland, SDQ, HoNOSCA and DC-GAS.
Conclusions:
Machine learning systems shows a promising potential in predicting the outcome of personalized multidimensional interventions in low-functioning ASD subjects. Accurate data collection, considering multidimensional aspects, and the use of a complex and complete statistical analysis, as the machine learning systems, could be useful in order to highlight predictable positive treatment factors.