AI Model Can Identify Toddlers at Risk of Autism With 80% Accuracy, Study Finds

Recent research has demonstrated the potential of artificial intelligence (AI) to assist in identifying toddlers at risk of autism, with an accuracy rate of about 80% for children under two.

Artificial intelligence (AI) has shown promise in aiding the identification of toddlers at risk of autism, with an accuracy rate of about 80% for children under two, according to recent research.

A team of researchers from the Karolinska Institutet in Sweden developed a machine learning-based screening system. While the AI model cannot replace traditional diagnostic methods, it could help identify children early on who may need further clinical evaluation.

“Using [the] AI model, it can be possible to use available information and identify individuals with an elevated likelihood for autism so that they can get earlier diagnosis and help,” said Dr. Kristiina Tammimies, a study co-author.

However, she cautioned that the model should not be viewed as a standalone diagnostic tool, reiterating that the final diagnosis should be conducted through standard clinical methods.

The AI model was developed using data from the U.S.-based Spark study, which provided information on 15,330 children diagnosed with autism and an equal number of children without the condition.

From medical and background questionnaires, the researchers selected 28 measures that could be easily obtained before children reach 24 months of age, such as age at first smile, eating behaviors and age at first construction of longer sentences.

Using machine learning to analyze patterns in the data, the research team compared the identified patterns between autistic and non-autistic children to build four different models, selecting the most effective one for further testing.

When applied to a separate dataset of 11,936 participants, the model correctly identified 78.9% of the children as either autistic or non-autistic. Specifically, the accuracy was 78.5% for children aged up to two years, 84.2% for those aged two to four years and 79.2% for those aged four to ten years.

An additional test using a dataset of 2,854 autistic individuals resulted in a lower accuracy rate of 68%, which the researchers attributed to differences in the dataset, including some missing parameters.

The study identified several key measures that significantly influenced the AI model's prediction of autism, including problems with eating certain foods, the age at which a child first constructed longer sentences, the age at which a child achieved potty training and the age at which a child first smiled.

These factors, according to the research team, played a crucial role in the model’s ability to differentiate between autistic and non-autistic children.

Further analysis revealed that the model tended to identify autism more accurately in individuals who exhibited more severe symptoms and broader developmental issues. This finding suggests that the model might be more effective at recognizing cases with more noticeable developmental challenges accompanying autism.

Despite the promising results, some experts expressed concerns about the model's ability to correctly identify non-autistic children. With an 80% accuracy rate, the model could potentially lead to overdiagnosis and unnecessary stress for families, as 20% of non-autistic children might be incorrectly flagged as possibly autistic.

Professor Ginny Russell from the University of Exeter sounded a note of caution regarding the push for early diagnosis, especially in very young children.

“It can be hard to tell the difference between a toddler who has a severe impairment and one who is simply developing more slowly but will eventually ‘catch up.’ I would not recommend applying psychiatric labels to children under the age of two on the basis of a limited range of behavioral indicators, such as whether they eat certain foods,” Russell said.