Speech-based concussion detection in athletes using Mel-spectrograms and convolutional neural networks

Area of Science:

• Neurology
• Artificial Intelligence
• Digital Health

Background:

• Neurological disorders can cause subtle changes in motor control, cognition, and speech.
• Conventional diagnostic tools and imaging (CT, MRI) often miss these functional abnormalities.
• Non-invasive digital health approaches are gaining traction for neurological assessment.

Purpose of the Study:

• To investigate the efficacy of speech-based deep learning for detecting concussion-related dysfunction.
• To analyze structured speech tasks for identifying subtle neurological alterations.
• To explore the potential of convolutional neural networks (CNNs) in speech analysis for mild traumatic brain injury (concussion).

Main Methods:

• Audio recordings of 225 athletes performing eight structured speech tasks were collected.
• Recordings were converted to Mel spectrograms and analyzed using a CNN with data augmentation.
• Gradient-weighted class activation mapping (Grad-CAM) was used to interpret model behavior and task relevance.

Main Results:

• The CNN model demonstrated strong classification performance, with accuracy up to 96% for specific speech tasks.
• Tasks involving dynamic articulation and rhythmic control (e.g., sentence reading, rapid syllable repetition) yielded the highest accuracy.
• Sustained vowel tasks showed lower performance and less focused model attention due to limited acoustic variability.

Conclusions:

• CNN-based analysis of dynamic speech features is effective for detecting neurological dysfunction.
• Task selection is crucial for optimizing speech-based digital diagnostic tools.
• This framework shows potential for future speech-based neurological assessment, with concussion as an initial application.