Nio hinder för klinisk diagnostik av autismspektrumtillstånd med artificiell intelligens: en systematisk litteraturstudie

Sammanfattning: This systematic literature review has tried to answer how eye-tracking and adjacent technologies in combination with machine learning (n = 13 studies) or deep learning (n = 6 studies) respectively can be used for the purpose of autism spectrum diagnostics (ASD). Included articles have been published in peer-reviewed journals. Most studies have sample sizes below 100 participants (n = 14 studies), 101-160 participants (n = 3 studies), 161-1000 participants (n = 2 studies). Studies include toddlers (16 months) to adults. Machine learning tends to be less accurate (59-93%) than deep learning (81-95%) in sorting individuals with or without ASD. The highest accuracy, precision (positive predicative value) and specificity (reliability) in the included studies were achieved by comparing eye-scanning and EEG data (95%, 95% and 95% respectively). A list of 9 obstacles that scientific studies would need to address before these AI technologies could be implemented in clinical practice as a tool for identifying ASD are reviewed.