Abstract
Background: Demographic and clinical predictors of aphasia recovery have been identified in the literature. However, little attention has been devoted to identifying and distinguishing predictors of improvement for different outcomes, e.g., production of treated vs. untreated materials. These outcomes may rely on different mechanisms, and therefore be predicted by different variables. Furthermore, treatment features are not typically accounted for when studying predictors of aphasia recovery. This is partly due to the small numbers of cases reported in studies, but also to limitations of data analysis techniques usually employed. Method: We reviewed the literature on predictors of aphasia recovery, and conducted a meta-analysis of single-case studies designed to assess the efficacy of treatments for verb production. The contribution of demographic, clinical, and treatment-related variables was assessed by means of Random Forests (a machine-learning technique used in classification and regression). Two outcomes were investigated: production of treated (for 142 patients) and untreated verbs (for 166 patients). Results: Improved production of treated verbs was predicted by a three-way interaction of pre-treatment scores on tests for verb comprehension and word repetition, and the frequency of treatment sessions. Improvement in production of untreated verbs was predicted by an interaction including the use of morphological cues, presence of grammatical impairment, pre-treatment scores on a test for noun comprehension, and frequency of treatment sessions. Conclusion: Improvement in the production of treated verbs occurs frequently. It may depend on restoring access to and/or knowledge of lexeme representations, and requires relative sparing of semantic knowledge (as measured by verb comprehension) and phonological output abilities (including working memory, as measured by word repetition). Improvement in the production of untreated verbs has not been reported very often. It may depend on the nature of impaired language representations, and the type of knowledge engaged by treatment: it is more likely to occur where abstract features (semantic and/or grammatical) are damaged and treated.