Duke University researchers are studying whether AI can be used to screen for autism in real-world settings, such as primary care. They want to develop an objective and scalable autism early detection tool that does not rely on parental reporting. The American Academy of Pediatrics recommends autism screening for all children at 18 and 24 months. Currently, the most common screening method is to have parents fill out a 20-question questionnaire asking about their child's behaviors associated with autism, such as response to name. Pediatricians may conduct interviews based on parents' answers and decide whether to refer the child for diagnostic evaluation.
However, the parental questionnaire screening method also has some limitations. Some parents may have difficulty understanding and answering the questions accurately. The questionnaire may not be provided in the parents' native language. Pediatricians may also not conduct interviews in a timely manner. These factors can affect the accuracy of the questionnaire. In real-world settings, such as primary care in pediatric clinics, large-scale studies have found that the accuracy of the questionnaire is not high enough. Children of color and girls are particularly prone to being overlooked.
In other areas of healthcare, doctors use multiple sources of information to assess a person's risk of having a certain condition. For example, if you are concerned about having a heart problem, your doctor will ask about your symptoms, such as increased fatigue, shortness of breath, or chest pain. Your self-report is an important part of your doctor's screening for heart problems. Your doctor may also perform objective measurements of heart function, such as an electrocardiogram, blood pressure test, and cholesterol test. Your doctor integrates multiple sources of information to make an evidence-based decision about your risk of having a heart problem. We do not have any objective tests for autism, such as a blood test. Autism is a behavioral condition diagnosed through behavioral observation. So, can AI, a computer rather than a person, detect and objectively measure early behavioral signs of autism?
A research team at Duke University has developed a digital application that can be downloaded on smartphones or tablets for autism screening. The app works by playing a series of short and engaging movies that are carefully designed to elicit behaviors associated with autism, such as gaze, orienting to name call, and facial expressions. As the child watches the movies, the camera in the smartphone or tablet records the child's behavioral responses. Then, a technology called "computer vision analysis" automatically and accurately measures the child's behavioral responses. The computer can measure whether the child is attending to social elements, such as people, or more focused on non-social elements, such as toys or other objects. It can measure the child's facial expressions, blink rate, orienting to name, and other body movements. The app then uses machine learning to integrate all these behavioral signals and determine the likelihood of the child having autism. All of this can be done using a smartphone or tablet in less than 10 minutes.
One advantage of using computers to measure behavior (called "digital phenotyping") is its high resolution and accuracy. Computers can measure subtle and unique behaviors that are imperceptible to the human eye. For example, an early sign is the child's blink rate. Typically developing children reduce their blink rate when looking at people, while children with autism reduce their blink rate when looking at objects. The computer can measure whether the child turns their head when called by name, but it can also tell how fast the child turns their head. Children with autism are about one second slower in orienting to name compared to typically developing children. Clinical doctors may miss this early sign.
In a recent study, a digital autism screening app was used in pediatric health check-ups for 475 children, of whom 49 were later diagnosed with autism and 98 were diagnosed with non-autistic developmental delay. The app showed a sensitivity of 87.8% in autism detection, meaning it correctly identified most children with the condition. Its specificity, the percentage of non-autistic children who screened negative, was 80.8%. The app was equally accurate in identifying autism in children of color and girls. Therefore, this more objective approach may help reduce disparities in early detection, which is the first step to accessing early intervention services.
Using computer vision to detect early signs of autism is just one example of how AI is changing healthcare. Computers can never fully replace the role of humans in providing high-quality, compassionate medical care. However, when AI is used ethically and responsibly, it offers many benefits and has the potential to increase access to healthcare services and deliver them more effectively and equitably.
