Rule-out test for autism using machine-learning analysis of molecular temporal dynamics in hair - a multicenter study
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
BACKGROUND
Early intervention can improve autism-related outcomes. However, no valid biosignature test exists yet for detecting or excluding autism. In addition, most behavior-based assessments of autism are developed for children aged 18 months and older. We developed a hair-strand-based biomarker test (diagnostic aid) to assist clinicians in ruling out autism in children aged 1 month and older.
METHODS
In a multi-national sample of 1697 (from California (two studies, n= 1112), New York City (n= 123), Sweden (n= 306), Japan (n= 110), and Mexico City (n= 46), with 97% below 21 years-of-age), autism was assessed using DSM-5 criteria for autism spectrum disorder or gold standard diagnostic instruments (ADOS-2 and/or ADI-R). A single hair-strand from children collected at 1 month or older was analyzed using laser ablation-inductively coupled plasma-mass spectrometry to sample down the shaft, generating time-series data at a resolution of ∼800 timepoints (on average) for 12 elemental intensities. We employed a multi-layered machine learning architecture to leverage the temporality of elemental intensities and optimized the test negative predictive value (NPV) and sensitivity. Models were trained, ensembled, and tuned on participants from California and Sweden, then tested on 580 participants (within-population replication in California and Sweden, and external population testing in New York, Mexico City, and Japan).
RESULTS
The diagnostic aid showed an AUC of 0.75 (95%CI: 0.70-0.79), 97% NPV (95%CI: 0.92-0.99), and 96% sensitivity (95%CI: 0.91-0.98), respectively (estimated prevalence for NPV set at 14%). Moreover, for those 36 months or younger, the AUC was 0.79 (95%CI: 0.73-0.85), with 97% NPV (95%CI: 0.87-0.99), and 96% sensitivity (95%CI: 0.88-0.99), respectively. Compared to the baseline odds of autism before taking the test (pre-test odds), those who test negative are, on average, ∼82% lower in odds of autism diagnosis, whereas those who test positive are ∼25% higher.
CONCLUSIONS
By estimating autism likelihood as early as 1 month after birth, early intervention can be delivered with higher precision to young children with developmental support needs. Clinicians may use this diagnostic aid to support early autism diagnosis, improve clinical workflows, and significantly reduce wait times for services.
