Final Paper Number 413.019
Background: Structural and pragmatic language deficits are core symptoms of Autism Spectrum Disorder (ASD) and predict long-term outcomes. Although language proficiency is a treatment target in early intensive behavioral interventions, measurement is cumbersome and costly. Better language outcome measures are needed. By analyzing language transcripts, our study generated Automated Discourse Measures (ADMs) that we tested for their convergent and discriminant validity across clinical groups.
Objectives: 1. examine language differences between three clinical groups (ASD, ADHD, and TD); 2. analyze the convergent validity of these measures by calculating correlations between the ADMs and standardized language measures; 3. investigate the accuracy of each individual ADM in predicting ASD status; and 4. examine if gains in accuracy would be obtained by combining all ADMs together in predicting ASD status.
Methods: 169 participants (96 ASD, 28 Typically Developing (TD), 45 Attention Deficit Hyperactivity Disorder (ADHD) ages 7 to 17 were evaluated with the Autism Diagnostic Observation Schedule (ADOS-2), module 3. ADOS tasks were transcribed following SALT guidelines. Transcripts of one task were automatically analyzed with novel software to generate seven ADMs for each participant: Mean Length of Utterance in Morphemes (MLUM), Number of Different Word Roots (NDWR), um proportion, content proportion, unintelligible proportion, c-units per minute (CPM), and repetition proportion.
Results: With the exception of repetition proportion (p = .07), nonparametric ANOVAs (Kruskal-Wallis) showed significant group differences (p<.01). The TD and ADHD groups did not differ from each other in post-hoc analyses for the seven ADMs.The ASD group showed significantly lower language skills. The highest effects sizes were found for content proportion and CPM. The ADMs were correlated with standardized clinical evaluations of ASD, which provided support for their convergent validity. In logistic regression analyses adjusted on age and IQ, four ADMs were found to significantly predict ASD versus nonASD status with accurate classification ranging from 67.9% to 75.5%. When combined together in one model, an overall correct classification rate of 82.4% was achieved. We estimated ROC curves using the class prediction probabilities drawn from the logistic regression models. The combined model using all seven ADMs had the highest AUC of 0.9223, followed by the individual model for content proportion with an AUC of 0.8149.
Conclusions: All seven ADMs show an improved accuracy of ASD prediction over a baseline model using only age and IQ; a combined model achieves a highly improved prediction model for ASD diagnosis using easily classifiable language measurements. These ADMs offer a promising approach for generating novel outcome measures.