Early diagnosis of autism is vital to improving outcomes and typically occurs around age 5 years. A new artificial intelligence system has now been shown to predict autism earlier in life by using minimal health information.
Researchers used a machine learning model to correctly identify 79% of children with autism, according to a new study published in JAMA Network Open. The model, which analyzes easily obtainable information from medical records and background history, shows promise as a screening tool for children even younger than 2 years old.
“Explanatory [machine learning] models can inform clinicians about the underlying factors leading to ASD detection,” the study states. “Further, they can assist in targeted intervention and follow-up.”
The need for early diagnosis tools has led researchers and the autism treatment industry to develop a variety of novel screening mechanisms, from gut microbiome markers to eye-tracking technology.
Machine learning is a subset of AI that involves programming computers with data and learning models to allow computers to train themselves to find patterns or make predictions.
Researchers for this study utilized machine learning to screen a database of 30,660 children, half of whom were diagnosed with autism and half who were not.
The machine learning model analyzed 28 basic medical and background history items. Of these, researchers found that developmental milestones and eating behaviors were particularly key predictors of autism.
Researchers describe the model’s success as “robust,” with 21% of patients incorrectly identified. Those who were incorrectly identified tended to have significantly different traits from those who were correctly identified, suggesting that the model best identifies children with more symptoms of autism, especially those that relate to communication skills and social functioning.
Currently, clinics often utilize questionnaire-based screening tools, which can struggle when handling diversity in age, sex, race and ethnicity. The machine learning model effectively handled these factors, researchers said, and performed consistently across all groups.
The model particularly shines in the early diagnosis of autism. Among children younger than 2 years old, the model achieved an area under the ROC curve (AUROC) score of 0.868. AUROC scores summarize the accuracy of a machine learning model’s diagnoses. A score of 0.8 to 0.9 is considered excellent.
“The challenges associated with the heterogeneity of autism, the applicability of different screening instruments in various settings, and the quantification and scaling abilities of digital tools may warrant using a combination of such multimodal screening tools at scale for reliable and robust prediction accuracy,” researchers wrote. “Our developed model has the potential for clinical use as a noninvasive [autism spectrum disorder] (ASD) screening tool.”
Alternative diagnostic tools can aid in bottlenecks to autism therapy access, which are worsened by factors including shortages in clinicians trained to diagnose neurodiverse conditions. Early detection of autism is key for early intervention, which can improve developmental outcomes and adaptive skills.
