Dennis Wall and his colleagues used machine learning to study gold-standard instruments used for assisting the diagnosis of autism. They learned that the number of behaviors that clinicians must assess to arrive at an accurate diagnosis may be far smaller than what is used today. Wall and his group hypothesized that this small number of behaviors, when run through a machine-learning classification tool, would provide a ‘digital phenotype’ that clinicians could use to better manage children on their waiting lists with risk for developmental delay including autism.
The researchers tested this hypothesis in a cohort of 222 individuals at Boston Children’s Hospital. They created a mobile tool that families could use in advance of the clinical visit that enabled parents to provide their opinion of their child’s level of development in seven behavioral areas. These parent measurements were then run through a machine-learning system to generate a risk score. The researchers determined that this score, when measured against the doctor’s official diagnosis, was 90 percent accurate. The electronic platform, brief administration time and automatic scoring suggest potential for widespread use as an autism screening tool if further studies support these positive findings.