Continuous TBI Monitoring From Spontaneous Speech Using Parametrized Sinc Filters and a Cascading GRU

Area of Science:

• Neuroscience
• Speech-Language Pathology
• Artificial Intelligence

Background:

• Traumatic Brain Injury (TBI) affects cognitive and communication functions, leading to speech disorders.
• Continuous TBI monitoring is crucial for rehabilitation and preventing regression, with over 80,000 US individuals affected by long-term disabilities.
• Existing TBI speech monitoring methods rely on burdensome, episodic scripted tasks, lacking longitudinal insights.

Purpose of the Study:

• To develop and evaluate a continuous TBI monitoring system using passive, spontaneous speech analysis.
• To leverage AI and speech processing for unobtrusive, longitudinal assessment of TBI-related speech changes.
• To improve TBI detection accuracy and facilitate personalized rehabilitation strategies.

Main Methods:

• Extraction of low-level acoustic features using parametrized Sinc filters (pSinc) from spontaneous speech.
• Classification of TBI using a cascading Gated Recurrent Unit (cGRU) model incorporating prediction history.
• Passive speech data collection via smartphones for continuous, unobtrusive monitoring.

Main Results:

• The proposed cGRU method achieved 83.87% balanced accuracy in TBI classification on conversational speech.
• Outperformed prior TBI detection methods on patient-therapist discourse data.
• SHapley Additive exPlanations (SHAP) identified unique predictive words, and a correlation was found with coordination deficits.

Conclusions:

• Continuous TBI monitoring from spontaneous speech is feasible and effective using AI-powered acoustic feature analysis.
• The cGRU model offers a promising approach for unobtrusive, longitudinal TBI assessment.
• This technology can significantly aid in TBI rehabilitation and management by providing continuous patient insights.