Classifying Autism From Crowdsourced Semistructured Speech Recordings: Machine Learning Model Comparison Study.

Area of Science:

• Computational linguistics and artificial intelligence applied to developmental disorder diagnostics.
• Speech signal processing for identifying neurodevelopmental markers.

Background:

• Autism spectrum disorder (ASD) is a neurodevelopmental condition affecting communication and social interaction.
• Increasing prevalence of ASD necessitates efficient and accessible diagnostic tools.
• Abnormalities in speech prosody are recognized indicators of ASD in children.

Purpose of the Study:

• To evaluate machine learning (ML) models for automated autism detection using child speech.
• To assess the efficacy of ML in analyzing self-recorded audio from children's home environments.
• To compare different ML approaches for classifying autism spectrum disorder (ASD) versus neurotypical (NT) status.

Main Methods:

• Development of a novel dataset of cellphone-recorded child speech from the Guess What? mobile game.
• Implementation and training of three ML models: Random Forests, Convolutional Neural Networks (CNNs), and fine-tuned Wav2Vec 2.0.
• Audio features and spectrograms were used for Random Forests and CNNs, respectively; Wav2Vec 2.0 utilized transformer-based speech recognition.

Main Results:

• The Convolutional Neural Network (CNN) model achieved the highest accuracy at 79%.
• The fine-tuned Wav2Vec 2.0 model demonstrated 77% accuracy in classifying ASD status.
• The Random Forest classifier reached 70% accuracy, with all models evaluated using 5-fold cross-validation.

Conclusions:

• Machine learning models can effectively predict autism status from varied home-recorded speech data.
• These findings suggest the potential for automatic autism detection using readily available technology.
• The study highlights the feasibility of using speech analysis for non-specialized, real-world autism screening.