6th Edition of Neurology World Conference 2026

Abstract

Voice prosody has emerged as a promising non-invasive biomarker for identifying Autism Spectrum Disorder (ASD). This raises questions about the generalizability of voice-based biomarkers across different languages. Given the strong genetic basis of ASD, we predict that voice-based attributes should remain consistent across languages. We further hypothesize that cross-lingual training can help prevent machine learning models from overfitting to the characteristics of any specific language. Using data from three datasets (two in Danish and one in English) that include 74 autistic children and 58 neurotypical children, and leveraging the eGeMAPS feature set, we demonstrate the benefits of cross-lingual training. We utilize Extreme Gradient Boosted Decision Trees (XGBoost) algorithms, which have shown high accuracy in classifying medical datasets, to compare the performance of monolingual training (Danish only and English only) to cross-lingual training. Our findings reveal that cross-lingual learning outperforms monolingual training by 5 percentage points in F1 score and 4 percentage points in AUC-ROC. Moreover, our model trained and tested using both languages achieved an F1 score of 0.83 and an AUC-ROC score of 0.86. These results highlight the potential of using vocal non-invasive biomarkers to improve the efficiency and scalability of ASD diagnosis.