Integrating Complementary Learning Principles in Aphasia Rehabilitation via Adaptive Modeling

More information about this project can be found here

Motivation

Aphasia is a language disorder commonly caused by stroke and other acquired brain injuries that affects over two million people in the US and has a large negative effect on quality of life. Anomia (i.e., word-finding difficulty) is a primary frustration for people with aphasia, and naming treatments for anomia are both widely researched and commonly used in clinical practice. For naming treatments to make a meaningful impact on the lives of people with aphasia, they must produce durable gains in word-finding which generalize beyond the treatment context. However, most theoretically-motivated naming treatment research fails to address the long- term retention of trained words and their generalization to connected speech, limiting their clinical impact. Prevailing learning theory suggests that “desirable difficulty” improves treatment retention and generalization. The current proposal therefore seeks to improve the durability and context generalization of computer-based naming treatment by incorporating model-based algorithms to adaptively maintain desirable difficulty. We will test two distinct models in parallel clinical trials. Our central premise is that these models will facilitate a balance between what have historically been framed as contrasting learning approaches: errorless learning vs. effortful retrieval (Study 1) and massed vs. distributed practice (Study 2). Instead, our models will integrate these approaches by replacing extreme static contrasts with continuous task components which can be adaptively modified based on ongoing patient performance. Study 1 will adaptively balance effort and accuracy using speeded naming deadlines based on a model we have developed which characterizes individuals’ speed-accuracy tradeoffs in picture naming over time. Study 2 will manipulate trial spacing using an adaptive scheduling and memory decay model built into widely available, open-source flashcard software. In both studies, we predict that when compared to matched traditional non-adaptive treatment conditions, our adaptive conditions will produce more successful retention of trained words 3 and 6 months post-treatment on naming probes (Aims 1a, 2a), and better context generalization to connected speech when tested on complex scene descriptions containing untrained exemplars of trained words (Aims 1b, 2b). We also predict that adaptive trial spacing in Study 2 will successfully train many more words than is possible in current standard care. In addition, data generated in Studies 1 and 2 will be used to develop the next generation of adaptive timing models (Aims 1c and 2c), spurring future innovations in personalized medicine. Successful clinical trial outcomes will demonstrate that adaptive computer-based naming treatments provide a novel way to produce large, durable, and generalizable treatment gains, and positive Study 2 findings could be immediately implemented in clinical practice at scale using free open-source software. Successful modeling outcomes will lead to even more effective interventions and lay the groundwork for a transformative research agenda that could ultimately lead to comprehensive adaptive learning systems for aphasia rehabilitation.