1) Original Question and Solution
For our original question, we asked, “How can specific language impairment (SLI) be detected and prevented before young children begin showing signs of the disorder through their speech?”
Before consulting the experts on this issue, our solution looked like this:
Create a device that selectively records babies’ babbling and can detect if the infant is on track to have normal speech and comprehension abilities or whether the infant will develop SLI. There will also be an associated phone app for parents to monitor likeliness of their child developing SLI. Lastly, this app will be used to allow infants to get extra practice responding to different speech-related stimuli.
2) Expert Feedback and Solution Adjustments
a) In our conversation with Dr. Mark Liberman, a Penn professor in the Department of Linguistics and the Department of Computer and Information Sciences, we discussed the difficulty of acquiring enough data to build a strong SLI-prediction model. According to Dr. Liberman, a broader problem in biomedical research is that there is often a scarcity of data for projects like ours. Instead of there being shared databases for disorders like SLI, most of the data is typically non-shared data possessed by a small number of researchers. With this in mind, we decided to expand our question and solution to include three other types of childhood speech disorders besides SLI: stuttering and fluency disorders, childhood apraxia of speech, and speech sound disorders. By incorporating these other disorders into our model, we will have more data to work with, which will increase the likelihood that our model is able to detect an irregularity in infants’ babbling. While scarcity of relevant data to build the model is still a potential issue, we believe that this solution is the best way to combat that problem.
b) We also talked with Dr. J. Bruce Tomblin, a language and communication researcher and an expert on the epidemiology and genetics of developmental language disorders at the University of Iowa. One of the questions he posed was how we would be able to differentiate between SLI-specific babbling compared to babbling that might be related to other language disorders. The aforementioned idea to incorporate more genetic speech disorders and having the AI simply check if a child might have any of them would help solve this issue, because we will not have to pinpoint a disorder but instead present possible disorders for the child, and an expert could use current diagnostic tools to figure out the exact disorder. This way we are still contributing to early intervention without having to know the exact disorder.
c) Another new element we are adding is an input value for the age of the child. This was advised to us by Dr. Tomblin as a way to make it more efficient to evaluate the data. A child’s early development advances rapidly, so knowing the age of the child ahead of time would allow the data analysis to be streamlined and create a groundwork for comparison. For example, if the child’s age was unknown, then the symptoms of SLI, or any other speech disorder, would be difficult to detect since there would be no age value to center the data around. The implementation of this feature will drastically increase the accuracy of the design.