Impact of Spanish Dialect on Artificial Intelligence in Assessing Expressive Language

Abstract Objective AcceXible is a new Artificial Intelligence (AI) tool that identifies speech biomarkers for cognitive impairment that was developed in Spain. Good sensitivity (85.7%) and specificity (92.8%) in distinguishing patients with mild cognitive impairment from individuals with normal functioning have been established with populations in Spain, and preliminary data are similar in the United States and the United Kingdom. In this study, we assessed transcription accuracy for Spanish speakers in the US originating from Latin American countries (LAC) and the effect it may have on the performance of Accexible’s AI model. Method AcceXible involves participants completing several verbal tasks via computer interface, including Animal Naming (AN). Transcription accuracy for 16 Spanish-speaking patients from LAC in Texas was assessed by comparing data obtained from hand scoring of AcceXible AN to data from different customized expressive language transcription tools used in AcceXible’s platform. The transcription tool was updated based on inconsistencies found. Results AcceXible’s customized transcription tool was initially found to miss 34% of total items on AN due to differences in vocabulary and pronunciation between LAC and Spain. The transcription tool was modified to recognize the LAC vocabulary and pronunciations, which resulted in 25.84% increased accuracy for an overall transcription accuracy of 91%. Conclusions AcceXible is a useful tool for efficiently assessing expressive language, and the current findings highlight the need to consider not just language, but also nationality when adapting tests across cultures.