Machine Learning Identifies Infant Speech Preference as an Important Predictor for Early Autism Risk Detection
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
While conventional ASD risk assessments rely on later-emerging behavioral features, early vocal preferences offer infant-available predictive markers. Therefore, multimodal measures integrating behavioral features with vocal preference may enhance autistic risk prediction accuracy. This study developed machine learning models to identify high autistic risk infants using multimodal measures, including gaze duration of sound contrasts (speech vs. non-vocal; speech vs. monkey calls) and learning/motor scores. Analysis of 9/12 month infant data (56 high-risk, 82 low-risk) showed SVM performed best after optimization (71.43% accuracy, AUC=0.46). Average model feature importance analysis identified that key predictors are 9-month human-over-monkey preference, 9-month speech-over-nonspeech preference, and 12-month receptive language percentiles. Findings demonstrate that gaze-based speech preference assessment addresses critical limitations of traditional behavioral markers, and when integrated with predictive modeling, opens new pathways for significant advances in early autism risk detection.
Fields of Study & Subjects
Autism Spectrum Disorder Research
Language Development and Disorders
Child Nutrition and Feeding Issues
Publication details
- Journal / source
- Institutional Repositories DataBase (IRDB)
- Pages
- 41–56
- Publication date
- 1 March 2026
- Language
- EN