Using Neural Networks to Reinforce Absence of Gender Bias in Dyslexia Screenings

This study explores the potential of neural networks to detect the gender of a student given a student's score on the Carolina Automated Reading Evaluation (CARE), a collection of thirteen assessments of reading disabilities created by researchers at the University of South Carolina. Data from prior administrations of the CARE at two elementary schools in South Carolina was parsed and read into dictionaries that contained a randomly created ID number, the student's gender and grade level, and their scores on the individual assessments they were administered. A neural network was created in Python using prominent machine learning libraries, taking test scores as inputs and attempting to train gender as an output [16]. Since not every student was administered every assessment, only assessments that had been taken by more than 60% of the students were used in the neural network (see Fig. 1). These assessments were used to train the neural network to detect gender. After repeated attempts to train the network to this data set failed, a data set of similar size from the University of California, Irvine machine learning repository was used to test the network built to determine if there was a structural defect in its construction [15]. As shown in Figure 3, the constructed network was able to train to an appropriate degree on the data set from UC Irvine but despite repeated attempts was unable to train on the data with respect to gender from the CARE administrations. From this, it was interpreted that the CARE data follows similar patterns to previous studies of other tests of reading disabilities, which concluded that students in lower grades show insignificant differences in reading ability by gender.