Predicting effects of computer-based intervention on structure and content of aphasic patients’ spoken language q Denise McCall a , Marcia C. Linebarger b , Rita Sloan Berndt a, * a University of Maryland, School of Medicine, MD 21201, USA b Psycholinguistic Technologies, Inc., Jenkintown, PA, USA SentenceShaperä is a computer program designed to assist language production in aphasia. Previous research has identified two types of posi- tive effects from the system: (1) ‘‘Aided Effects’’: the superiority, in mea- sures of structure and/or informativeness, of spoken narratives created on the system (aided ) as compared to the same narratives produced with- out the system (Linebarger, Schwartz, Romania, Kohn, & Stephens, 2000; Bartlett, Fink, Schwartz, & Linebarger, 2007); and (2) ‘‘Treatment Effects’’: improvements from baseline in structure and/or content of unaided spoken narratives after a period of largely independent home use of SentenceShaper (e.g., Linebarger, McCall, Virata, & Berndt, 2007). These effects have varied across subjects, with some showing no effects and others making quite marked gains in structure, content and speech rate. Identifying speakers who are likely to benefit from SentenceShaper use is an important research goal. Demonstrating treatment effects can require a substantial time commitment, while aided effects can be detected after limited practice with the system. If aided effects are predictive of treatment effects, then the extent and type of aided effects may guide decisions about whether and how to employ SentenceShaper for treatment. McCall, Virata, Linebarger, and Berndt (2006) evaluated the relationship between aided and treatment effects in a small number of subjects who presented with pervasive structural impairments, and found that aided production was predictive of unaided gains in two structural measures: mean sentence length (MSL) and proportion of words that occur in sentences. The goal of the current study was to address this issue with speakers demonstrating a wider range of impairments. Subjects Eleven subjects differing in the severity and nature of their aphasic lan- guage symptoms participated in the same SentenceShaper treatment pro- tocol. Treatment gains for six of these subjects have been previously reported. All subjects were at least 1.5 years post-onset (range 1.5–9.5 years), demonstrated relatively preserved auditory comprehension, decreased fluency, and limitations in the content of their narratives as indexed by the proportion of Correct Information Units (CIUs) to total words produced (Nichols & Brookshire, 1993). Prior to treatment, nine subjects showed marked impairments in at least one of the two structural measures noted above. Procedures Unaided narratives describing a short, silent video were elicited, after which subjects were trained to use SentenceShaper. Following training, narrative production was assessed in two conditions: unaided descriptions of a wordless picture book followed during a later session by aided retel- lings of the same stories. Subjects then used SentenceShaper at home for a period of home practice, after which unaided production was reassessed using the same Baseline video. The first 150 narrative words of each story in each condition (n = 4 per subject) were scored using the Quantitative Production Analysis (QPA) (Saffran, Berndt, & Schwartz, 1989) to assess structural elaboration, and using the CIU Analysis to evaluate narrative content. CIU procedures were modified to assure comparability of aided and unaided samples for the purposes of this study. The ‘‘aided’’ condition allows speakers to ‘‘clean up’’ their narratives (rejecting repairs, repeti- tions, false starts, etc.) before Content Units/total words is computed, whereas ‘‘unaided’’ narratives are normally scored by comparing number of CIUs to total words produced. For this comparison, therefore, CIU analysis in the unaided condition was performed on the QPA Narrative Words only , from which these unintended words had already been removed. While this modification obscures a functionally important aspect of SentenceShaper’s aided effect, i.e., the opportunity to produce more flu- ent messages lacking such false starts and errors, it may afford a clearer comparison between the aided and unaided productions as reflections of the speaker’s ‘‘best effort,’’ which may be relevant to predicting treatment effects. Analyses (v 2 for data expressed as proportions, and paired t-test on the number of words produced in sentences) compared: (1) aided and unaided performance immediately following training to assess ‘‘aided effects’’ on the QPA and CIU measures, (2) unaided production at Baseline and Post-treatment to assess ‘‘treat- ment effects’’ on these same measures, (3) number of patients showing consistent effects on both measures (consistently negative or positive), to assess the predictive value of the aided assessment. q Supported by NIH-NIDCD R01DC05629. doi:10.1016/j.bandl.2007.07.118 * Corresponding author. Fax: +1 410 706 0324. E-mail address: rberndt@umaryland.edu (R.S. Berndt). www.elsevier.com/locate/b&l Brain and Language 103 (2007) 8–249