CognoStroke: Automated Cognitive and Mood Assessment on the Hyper-Acute Stroke Unit

Area of Science:

• Neurology
• Artificial Intelligence
• Psychiatry

Background:

• Cognitive and mood impairments are prevalent in stroke survivors, negatively impacting outcomes and quality of life.
• Current assessment methods are time-consuming and impractical for hyper-acute stroke units (HASU).
• Automated assessment tools are needed for efficient and timely evaluation in HASU settings.

Purpose of the Study:

• To introduce and evaluate CognoStroke, an automated system for assessing cognitive and mood impairments in stroke survivors within HASU.
• To determine the efficacy of large language models in analyzing speech for mood and cognitive state classification.
• To identify barriers to the adoption of automated assessment tools in HASU.

Main Methods:

• CognoStroke utilizes conversational AI and speech analysis via large language models (GPT2, BART, RoBERTa) to classify cognitive (MoCA) and mood (GAD-7, PHQ-9) thresholds.
• Performance was measured using Macro F1-scores (MFSs).
• Patient-reported barriers to using CognoStroke were collected.

Main Results:

• CognoStroke achieved a best MFS of 0.783 for MoCA thresholding using optimized prompt combinations.
• MFSs of 0.686 for PHQ-9 and 0.617 for GAD-7 were obtained, with improved scores using fewer prompts.
• 75 out of 151 stroke survivors completed the assessment, with common barriers including computer access and technology anxiety.

Conclusions:

• CognoStroke demonstrates potential for classifying stroke survivors based on cognitive and mood status, aiding early intervention.
• Optimizing conversational prompts can enhance the accuracy of automated assessments.
• Addressing challenges like patient computer access and technology anxiety is crucial for widespread implementation.