28747
A Multimedia Screening System to Predict Later ASD Symptoms in Diverse Community Settings: A Machine-Learning Design for Infants and Toddlers
There is a need to create more efficient, accessible autism screening systems so as to reach more diverse populations, decrease parent burden, and provide a roadmap for incorporating advances in digital and machine learning technologies to facilitate state-of-the art prediction of developmental outcomes. The Yale Adaptive Multimedia Screeners (YAMS-One and YAMS-Toddler) were designed to meet these needs.
Objectives:
To describe the development of the YAMS-One system, designed to predict autism symptoms at 18 months from questions asked to parents when their child is 12-months-old, and the YAMS-Toddler, designed to predict concurrent levels of autism symptoms via questions asked when children are 18- to 24-months of age. Results described are based on machine-learning results acquired from n=386 infants and toddlers.
Methods:
Phenotypic data from participants seen at a research clinic fed machine learning algorithms. Participants were [YAMS-One] high-risk infant siblings of children with ASD seen at 12- and 18-months of age (n=76), and [YAMS-Toddler] toddlers seen between 15- and 27-months of age (n=310: ASD, n=102). Machine learning algorithms were tuned to predict ADOS Module 1 scores from questionnaire and assessment items at either 12-months [YAMS-One] or their ADOS-concurrent visit [YAMS-Toddler]. From over 400 items in both versions, spanning autism-specific, developmental, and temperament/regulatory-focused instruments, manual data cleaning and participant culling was followed by: (1) variable imputation (kNN, n=5, median); (2) variable selection (elastic net); (3) machine learning fitting. 100 repeats of 10-fold cross validation were used to estimate prediction accuracies.
Results:
[YAMS-Toddler] Naïve approaches using linear regression and no variable selection were inadequate for predicting ADOS totals (r=.02). Support vector regression was a benchmark for machine learning (r=.641), but required all items. Tree-based CART provided adaptive item parsimony, but suffered from generalization problems (r=.440). Random-forests, like SVM, showed good performance (r=.603) but also required all items. A hybrid sparse-forest variant of CART was developed to create sets of 5 trees that together showed robustness comparable to random forest (r=.548). Results were similar for [YAMS-One], r=.511.
Discussion:
Items unsuited for video-based translation were rejected and training restarted. Final items were examined for redundancy and collapsed into composites via forum review by clinicians and developmental scientists. Items were translated into scripts enacted by mothers and their children (reflecting diverse demographics) for developing professionally-filmed content. Video clips were incorporated into a tablet-based system alongside digital consent, demographic surveys, and implementations of text-based [YAMS-One] CSBS-ITC (Wetherby & Prizant, 2002) or [YAMS-Toddler] M-CHAT-R/F (Robins et al., 2014) for validation. Clinical review screens were developed so health providers could act on validated CSBS/M-CHAT results. Handshaking with backend servers facilitated data monitoring. YAMS was recently piloted in diverse pediatric and research clinics, and showed promising results (e.g. YAMS-Toddler & ADOS scores: r(46)>.60, p<.001).
Conclusions:
This study reflects the development of a multimedia screener reflecting the multidisciplinary expertise of clinical, developmental, and computer science. The resultant screening system was based on machine learning results adapted for parsimony in delivery and reflected necessary tradeoffs based on clinical and practical needs. Results of the deployment of the YAMS-systems are ongoing.