Life for a child who is suffering from an undiagnosed language-based learning disability, such as dyslexia, can be tough.
Kids can feel a sense of isolation as they struggle with something everyone else seems to understand. Teachers might not realize a student has a problem and accuse them of being lazy or unfocused. Parents might worry about their child’s intelligence, and classmates may make take to ridiculing a child for being different.
Even when it is discovered that a child may have a language-based learning disability, moving through the process to provide accommodations for a student can be tedious and take years to implement, critical time that can have a long-term impact on a child’s life.
But one undergraduate at the School of Informatics, Computing, and Engineering is developing a method for identifying students with possible language-based learning disabilities and possibly getting those students the help they need.
Katie Spoon, a senior in computer science who also is pursuing her master’s degree in C.S. through SICE’s Accelerated Master’s program, knows the difficulties children face. Her mother has built a career working with students with learning disabilities, and when given the opportunity to take on a meaningful project as part of the Proactive Health Lab at SICE, she didn’t hesitate.
“About 20 percent of students are estimated to have a language-based learning disability,” Spoon said. “I think only 10 percent of those students are actually detected as being possibly dyslexic by the end of the second grade. That’s important, because most school districts have a very long wait list for students to be diagnosed with a learning disability.”
It can take 3-4 years for a student to be diagnosed by an educational psychologist provided by a school system due to a long waitlist, and the expense for a diagnosis outside the system can be prohibitive.
“Unless students are diagnosed by the end of elementary school and have their accommodations in place, they perform at a lower rate vs. other students,” Spoon says. “They are less likely to graduate from high school. It’s really hard to get accommodations in college if you haven’t previously been diagnosed, so students who are diagnosed late really face an uphill battle against the issue.”
Spoon, who worked as a research assistant in the Computer Vision Lab and took a graduate-level computer vision class with Associate Professor David Crandall, used lessons learned from those experiences to design the architecture for a neural network that analyzed writing samples from students both with and without dyslexia. By splitting images of writing samples into individual lines then further dividing those lines into 50 patches, Spoon created more than 37,000 samples of writing to feed into the neural network.
“The cool part is they weren’t actual letters,” Spoon says. “We weren’t trying to split lines into individual letters and look for spelling mistakes. We were looking at the features of the handwriting itself, things like connections between the letters and the way a student writes. Then, two convolutional neural networks share information between the layers in a multi-stream approach.”
The result is a significant improvement from just 10 percent of students with possible language-based learning disabilities being recognized to more than 60 percent. Once a possible issue is detected, the process for diagnosis can begin.
“We’re not trying to diagnose students,” Spoon says. “We’re just trying to raise a flag to alert parents, teachers, and administrators that a student has some signs that there may be an issue, and they need to look into the matter further in case accommodations would be appropriate. Early detection would allow students to get into the cue, and it would allow them to be diagnosed if there is an issue by the sixth grade. It’s a stepping stone, but it’s an important one.”
Spoon is in the process of gathering more data to refine the system, and she hopes to soon have a front-end tool developed that would allow teachers and parents to take a picture of a student’s handwriting and receive a risk score that would show if a sample had features that could be indicative of a learning disability.
They also intend to use the effort to teach kids about machine learning by collecting samples from kids writing the same paragraph on the same paper.
“We’re going to make sure they understand how machine learning works as part of the process,” says Spoon, who enjoyed an internship with IBM Research last summer where she worked on artificial intelligence hardware accelerators. “We want to introduce the idea of data privacy to the kids and teach them that big corporations collecting data is different than academic groups collecting data.”
Associate Professor Katie Siek is the program director of Proactive Health Informatics. She first conceived the project when her daughter was being evaluated for a language-based learning disability.
“I tried to figure out why my child wasn't reading ‘at grade level,’” Siek says. “We had been pointing out to teachers that she continued writing letters and numbers backwards, spelled phonetically, and struggled decoding words. Each year, her teacher would tell us that they would keep an eye out for it because they wanted some time to assess her on their own. They spent the first half of the year assessing her along with her 29 other classmates, then they would start prepping them for standardized tests for the spring. They would then meet with us at the end of the year saying, ‘Yes, she is behind, but we'll see how much she progresses over the summer.’”
Siek took her daughter to an educational psychologist, who asked to see examples of her daughter’s handwriting since she started writing.
“As I was sifting through all of her writing samples and categorizing them by timeframe in each grade, it occurred to me this is what her teachers needed to see,” Siek says. “Teachers typically teach a child for a year before handing them off to the next teacher. This handoff each year is critical for children, especially for those with reading, cognitive, or developmental disabilities. How long has Jaimie been struggling writing b, p, q, d? A teacher does not have time to look through a binder to trace back every instance of b, p, q, d for every child in their class.
“When teachers want to assess the child on their own, they are relying on their own memory during those first few months of school to see if there are issues (on top of learning about each child, teaching an intense curriculum, etc.). But what if computer could automate this? What if as a teacher graded, a system could notify the teacher that Jaimie has been struggling with these letters in every assignment since kindergarten? Then the student may be able to get help sooner.”
Siek’s experience prompted her to suggest the project to Crandall, opening the door for Spoon to get involved.
“We were fortunate Katie decided to try it,” Siek says. “The first thing we learned is that teaching a computer how to identify a child’s handwriting is more difficult than we anticipated. Katie's research has the potential to help us identify children who may have reading and writing disabilities sooner. In Indiana, this is critical since the IREAD test is in the third grade and can impact a child’s ability to move into the fourth grade.”