Continuously-Encoded Deep Recurrent Networks for Interpretable Knowledge Tracing in Speech-Language and Cognitive Therapy

Intelligent Tutoring Systems (ITS), developed over the last few decades, have been especially important in delivering online education. These systems use Knowledge Tracing (KT) to model a student’s understanding of concepts as they perform exercises. Recently, there have been several advancements using Recurrent Neural Networks (RNNs) to develop Deep Knowledge Tracing (DKT) that eliminates the need for manually encoding the student knowledge space. In online education, these models are crucial for predicting student performance and designing personalized curricula (sequence of courses and exercises). In this paper we develop a novel Knowledge Tracing model, called Continuously-encoded Deep Knowledge Tracing (CE-DKT) to automatically encode the user’s knowledge space, when the user’s skill in a given task is continuous-valued instead of binary. We then apply Knowledge tracing, specifically CE-DKT, to the context of digital therapy. Specifically, patients suffering from various neurological disorders such as aphasia, traumatic brain injury, or dementia are often prescribed speech, language and cognitive therapy exercises to perform from a set of predefined workbooks that are not personalized for the patient. We use CE-DKT to automatically encode a patient’s skill level across different tasks, and predict how the patient will perform on unseen tasks. We use data from the digital therapy platform, Constant Therapy, to train a CE-DKT model and demonstrate its high degree of accuracy in predicting a patient’s performance in a digital therapy application. We also demonstrate how to extract interpretable confidence intervals from this model and how to trace predictions to previous tasks using time-step level feature importance. Finally, we describe how this model can be applied to significantly enhance future digital therapy platforms and online student learning systems.