Robust and Accurate Features for Detecting and Diagnosing Autism Spectrum Disorders

Area of Science:

• Speech Signal Processing
• Biomedical Engineering
• Developmental Psychology

Background:

• Autism Spectrum Disorder (ASD) detection and classification remain challenging.
• Existing speech feature extraction algorithms have limitations.
• Accurate identification of ASD subtypes is crucial for targeted interventions.

Purpose of the Study:

• To evaluate a novel speech feature extraction algorithm for the Interspeech 2013 Autism Challenge.
• To improve the accuracy of detecting children with ASD and classifying them into subtypes.
• To compare the performance of the new algorithm against baseline methods.

Main Methods:

• Applied a harmonic model to estimate fundamental frequency (f0) and derived features like Harmonic-to-Noise Ratio (HNR), shimmer, and jitter.
• Utilized these novel features alongside standard acoustic features (energy, cepstral, spectral).
• Employed regression for ASD detection and classification for subtype identification.

Main Results:

• The novel speech features improved ASD detection by 2.3% in unweighted average recall (UAR).
• Classification of ASD into four subtypes was enhanced by 2.8% UAR.
• The algorithm demonstrated superior performance over baseline methods in both subtasks.

Conclusions:

• The developed speech feature extraction algorithm shows significant promise for improving ASD detection and classification.
• Harmonic model-based features offer valuable insights into speech characteristics relevant to ASD.
• This approach has potential applications in early diagnosis and personalized treatment strategies for ASD.