S-119
Theoretical and Computational Neuroscience
Automated speech and language markers of childhood autism in restricted interest based conversational tasks
Brenda Evelyn Soria1,2, María Luz Gonzalez-Gadea1,2, Tomás Torres Barbero3, Franco Javier Ferrante1,2,4, Alejandro Sosa Welford1, Alexia Rattazzi5, Esteban Vaucheret Paz6, Adolfo Martín García1,7,8
1. Centro de Neurociencias Cognitivas, Universidad de San Andrés, Buenos Aires, Argentina.
2. National Scientific and Technical Research Council (CONICET), Argentina.
3. Universidad Argentina de la Empresa, Buenos Aires, Argentina.
4. Facultad de Ingeniería, Universidad de Buenos Aires, Argentina.
5. PANAACEA, Buenos Aires, Argentina.
6. Servicio de Neurología Infantil. Hospital Italiano de Buenos Aires. Buenos Aires, Argentina.
7. Global Brain Health Institute (GBHI), University of California, San Francisco, CA, USA and Trinity College Dublin, Dublin, Ireland.
8. Departamento de Lingüística y Literatura, Facultad de Humanidades, Universidad de Santiago de Chile, Santiago, Chile.
Presenting Author:
brendasoria6@gmail.com
Automated speech and language analysis has demonstrated potential for identifying the linguistic and acoustic characteristics of language in autistic children. However, most studies use traditional assessment tools to extract markers, which are resource-intensive. This study used brief conversation to elicit natural communication about general and restricted interests (GI and RI) in 72 participants (35 autistic and 37 non-autistic, age between 6-15 years old). Acoustic and linguistic markers were then extracted and analyzed using group-level comparisons and machine-learning classification. Results revealed group differences in pitch variability, disfluencies, fluency, lexical diversity, and third-person pronoun use. Most alterations were found in RI rather than GI interviews. Machine-learning models integrating speech and language features achieved high classification accuracy (AUC = 0.88 vs. 0.82). The RI-based conversational tasks are promising for ASD diagnosis due to their objectivity, affordability, scalability, and sensitivity and could complement traditional assessment.
