Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Poster Slams
Simulating individual differences in response to alexia therapies
Poster C58 in Poster Session C, Friday, October 7, 10:15 am - 12:00 pm EDT, Millennium Hall
Ryan Staples1, Olga Boukrina2, Amanda Khoudary2,3, Nicole Giordano2, Dima Karim2,4, Elizabeth B. Madden5, William W. Graves1; 1Rutgers University - Newark, 2Kessler Foundation, 3Hackensack Meridian Health, Nutley, New Jersey, 4Rutgers University Biomedical and Health Sciences, 5Florida State University
Acquired reading disorders are common after stroke, and while there are effective reading therapies, patient outcomes are highly variable. Reasons for this variability include the severity and location of stroke-related brain damage, as well as the specific therapy utilized by clinicians. This is compounded by the lack of a formal understanding of which therapy is ideal for a given patient’s specific impairment. Computational models of reading offer a potential solution to this problem. A full range of possible reading impairments can be simulated, and therapies can be tested against each other to determine which promotes the greatest recovery. We compared the performance of two treatments in their capacity to promote recovery using a computational model of reading damaged to reflect a variety of acquired impairments. The first treatment was a partial, model-adapted implementation of phonomotor therapy (PMT), a multimodal therapy focused on improving phonology. The second treatment was whole-word training. We hypothesized that PMT, with its focus on phonology, would promote recovery following orthography-phonology damage. Since PMT provides no semantic information, we hypothesized that whole-word training would promote greater recovery following semantic damage. We lesioned a feedforward neural network model in four locations: the connections into and out of the hidden layers mediating between orthography and phonology (OP), between orthography and semantics (OS), between semantics and phonology (SP), and the connections into and out of semantics (Sem). Five healthy models were independently damaged three times at each of 25%, 50%, and 75% of connections (80%, 90%, 95% for SP lesions) removed. The lesioned models were retrained using either model adapted PMT or by retraining the model with whole words. Model adapted PMT consisted of 100 epochs of training on anchor vowel and consonant grapheme-phoneme correspondences, 100 epochs of training on anchor/intermediate vowels and consonants, 100 epochs of all vowel and consonants, and finally 60 epochs of words. Normal recovery consisted of 100 epochs of training on words. The number of training items was matched for the treatments (PMT: 303,280, whole word: 299,800). The models were assessed using sets of high and low frequency/consistency words. We find main effects of lesion location, F(2,702 = 1431.26, p < 0.001), severity F(2,702 = 424.31, p < 0.001), and therapy F(1,702 = 524.86, p < 0.001), as well as a three-way interaction F(4,702 = 59.89, p < 0.001). Word reading is best promoted by whole-word retraining across all lesion severities following any semantic lesion, and after 25% and 50% OP lesions. However, following 75% OP lesions, PMT caused better reading recovery. Our results provide a proof-of-concept for tackling individual differences in alexia recovery using computational models. Consistent with our hypotheses, semantic lesions are better served by whole-word training, while severe OP lesions were best remediated with PMT. We suggest that careful use of behavioral testing to identify an individual patient’s impairments can inform the location and severity of damage to models of reading. Therapies can then be tested on the individualized model to identify which will promote the greatest recovery for a given patient.
Topic Areas: Computational Approaches, Disorders: Acquired