A machine learning algorithm can detect signs of anxiety and depression in young children’s talks, allowing for the rapid diagnosis of situations that are difficult to deal with and often overlooked in young people, according to a new study published in the Journal of Biomedical and Health Informatics.
“We need fast, objective tests to identify children who are suffering”
says Helen McGinney, a clinical psychologist at the Vermont Center for Children, Youth and Families at the University of Vermont Medical Center and lead author of the research.
“Most children under eight are not diagnosed”,
adds; given that children of such ages are not able to reliably express their emotional problems.
Early diagnosis is particularly important because children respond better to treatments while their brains are still developing, however if they are not treated, there are increased risks of suicide, substance use etc. later. The established method of diagnosis is a 60-90 minute interview with experts, but Helen McGinney and Ryan McGinney, a biomedical engineer at the University of Vermont and senior research author, were looking for ways to harness artificial intelligence and machine learning for faster and more reliable diagnoses.
Researchers used a specially tailored version of a Trier-Social Stress Task designed to trigger anxiety feelings in the subject. 71 children aged 3-8 were asked to create a three-minute story improvisation and told them that they would be judged on the basis of how interesting it would be; the judge would maintain a strict attitude in the conversation, either in neutral or negative reactions. After 90 seconds, and while there were 30 seconds left, a bell sounded, and the judge was saying how long it was.
“Exercise is designed to cause anxiety, and to make them think that someone is judging them”
Children were diagnosed through conventional methods (clinical interview, questionnaire to parents). Also, a machine learning algorithm was used to analyze the statistical characteristics of the recordings of the stories narrated by the children, followed by correlation with the children’s diagnoses. As we have seen, the algorithm had high success rates in diagnostics.
“The algorithm was capable of detecting children with an internalized type diagnosis of 80% accuracy, and in most cases this could be compared with the accuracy of the parent questionnaire”,
says Ryan McGinney. It also produced results very quickly: The algorithm only takes a few seconds to process its data.
The algorithm identified eight different sound features in children’s speeches, but in particular three stood out (low-key voices, repetitions, and a bell-to-bell response), with Helen McGinney pointing out that these are features that one would expect to see in a person suffering from depression. According to the researcher, the next step will be to further develop the algorithm into a more general-purpose tool for clinical use; perhaps through a smartphone application that will record and analyze the results directly.