Stanford researchers use AI to diagnose autism from home videos

A Stanford University research team used machine learning to quickly and accurately diagnose autism in children through short home videos, according to a research article published in PLOS Medicine. After seeing promising results, the method could greatly impact how soon autism patients get access to treatment after being diagnosed.

“These results support the hypothesis that the detection of autism can be done effectively at scale through mobile video analysis and machine learning classification to produce a quantified indicator of autism risk quickly,” Qandeel Tariq, research data scientist with the Stanford University School of Medicine, et al. wrote. “Such a process could streamline autism diagnosis to enable earlier detection and earlier access to therapy that has the highest impact during earlier windows of social development.”

When diagnosing autism, physicians typically use several behavioral exams and evaluate between 20 and 100 behaviors that can take several hours to complete. These standard practices contribute to long wait times for a diagnosis, delayed access to therapy for patients and strain on the current healthcare system, researchers said.  

“Wait times for a diagnostic evaluation can reach or exceed 12 months in the US, and the average age of diagnosis in the US remains near 5 years of age, with underserved populations’ average age at ASD diagnosis as high as 8 years,” Tariq et al. wrote. “The high variability in availability of diagnostic and therapeutic services is common to most psychiatry and mental health conditions across the US, with severe shortages of mental health services in 77 percent of US counties.”

Researchers believe using machine learning to analyze home videos could speed up the diagnosis process, while still being accurate. For the study, researchers used eight machine learning models to analyze 162 home videos of children with and without autism. The models were then tested on its ability to reliably detect autism on mobile platforms.

Each of the machine-learning models had a sensitivity above 94.5 percent, while only three of them had a specificity above 50 percent, according to the study. The top-performing model had an accuracy of 88.9 percent, sensitivity of 94.5 percent and specificity of 77.4 percent. The next best-performing models had accuracies of 85.4 percent and 84.8 percent.

Specifically, one model, the LR5 (a five-feature logistic regression classifier), had the highest accuracy on all age ranges, except with children over six years old.

“This model performed best on children between the ages of 4 and 6 years, with sensitivity and specificity both above 90 percent,” the authors wrote. “The three raters agreed unanimously on 116 out of 162 videos (72 percent) when using the top-performing classifier, LR5.”

While further testing is needed, researchers believe the approach could potentially reduce the “geographic and financial burdens associated with access to diagnostic resources and provide more equal opportunity to underserved populations, including those in developing countries” based on the findings.