Evidence-based modeling approaches to customizing treatments for acquired dyslexias

This proposal tests the ability of data-driven computational cognitive modeling of reading and its impairment to match people with aphasia and resulting reading problems (acquired dyslexia) to the treatment that will most benefit their particular patterns of reading difficulties. Roughly 2 million Americans are currently living with aphasia, with approximately 180,000 acquiring aphasia each year. This chronic language disorder is often extremely debilitating, severely limiting daily life function, social interactions, and vocational participation. Most individuals with aphasia suffer from impaired reading, with 68% meeting strict criteria for acquired dyslexia. Such difficulties are particularly incapacitating in today’s society where so much critical communication is by text, email, and social media. Structured interviews of people with aphasia show that reduced literacy represents a considerable loss in quality of life. Much has been learned about the cognitive basis of reading and its damage. Yet these advances have not led to corresponding advances in treating aphasia-related dyslexias. To formulate more individualized and effective treatments, we must build on models of typical reading and reading impairments to develop a model of reading rehabilitation. Currently there is no single standard of care for acquired dyslexia treatment. Treatments range from approaches focused on lists of whole- words and their associated pictures and spoken forms (e.g., iReadMore) to approaches such as phonomotor treatment (PMT) and semantic feature analysis (SFA) that target more specific processes. Improvements are often seen on words trained during treatment, but gains rarely generalize to untrained words. PMT and SFA have shown modest generalization to untrained words. Dyslexias occur in different patterns, reflecting different sources of impairment. What is needed is a method for matching a particular pattern of acquired dyslexia with the treatment that targets the components of reading affected by that impairment (personalized therapy). We propose using a neural network (NN) modeling approach that develops mappings between letter string inputs (orthography), spoken word forms (phonology), and meanings (semantics). We extend this model to cases of stroke-induced dyslexia by lesioning the model and allowing it to re-train, simulating aspects of brain damage and recovery. We show how disruption of processing can lead to different kinds of dyslexia. Critically, we extend it to model rehabilitation. We hypothesize that by simulating cognitive sources of reading impairment and their rehabilitation on an individual level, tailored treatment will lead to greater reading improvements than less tailored approaches. The specific aims are: (1) Develop model-based accounts of post-stroke dyslexias and their rehabilitation at the group and individual levels. (2) Determine whether computational model-based personalized treatment improves reading. Successful completion of the proposed studies will enhance our knowledge of the cognitive basis of reading, enabling personalized therapy to potentially improve quality of life for stroke survivors with aphasia and associated dyslexia.