Improving anxiety assessment in Stroke and Acquired Brain injury: the validation of the Anxiety Intensity Scale Circles (AISCs), the Yale-anxiety, and Natural Language Processing (NLP) models

Background: Assessment of anxiety in stroke services often involves the use of brief screening measures, but many of these tools are unsuitable for people with cognitive or communication difficulties.

Common adaptations for stroke include minimal use of language and centre alignment, to support with language impairments and hemispatial neglect, respectively.

Few anxiety measures possess such adaptations, highlighting an issue of inclusivity for people with cognitive communication impairments.

Aims: We aim, therefore, to explore the validity of two brief anxiety screens: the Anxiety Intensity Scale Circles (AISCs) and the Yale-anxiety (Yale-a). The AISCs is a one-item visual analogue scale, and the Yale-a is a simple closed question.

We will also investigate whether Natural Language Processing (NLP) models are able to identify mood disorders and can highlight previously unknown predictive linguistic features.

Methods: One hundred and forty-two stroke inpatients will be consecutively screened across five sites in the East of England. Participants more than two weeks post-stroke capable of giving informed consent will be included.

Consenting participants will complete the AISCs, Yale-a, the Depression Intensity Scale Circles (DISCs), Yale-depression (Yale-d), Patient Health Questionnaire-9 (PHQ-9), Generalised Anxiety Disorder-7 (GAD-7), and Western Aphasia Battery-Revised (WAB-R).

Participants will then complete a semi-structured interview delivered by a trained assessor, which will act as a reference standard by identifying cases of clinically significant anxiety or depression. The AISCs, DISCs, Yale-a, and Yale-d will be re-administered during a third contact to assess test-retest reliability.

Contacts will be voice recorded to enable NLP analyses.

Analysis: Convergent and divergent validity of the AISCs will be assessed via correlations with the GAD-7 and PHQ-9, respectively.

Classification accuracy of the AISCs and Yale-a against the reference standard will be calculated via a Receiver Operating Characteristic analysis. Test-retest reliability will be calculated using Intraclass Correlation Coefficients. Voice recordings will be transcribed using the Artificial Intelligence programme Whisper.

An existing NLP model will be trained using half of the recorded sample.

The predictive value of NLP will be evaluated using the second half and feature analysis used to identify relevant linguistic features. Timelines for delivery: The project will take place over two years. Data collection will commence within ten months of award and last approximately eight months. Analysis and write-up will be completed shortly after, and first papers submitted for publication within 20 months.

Impact and dissemination: These accessible measures will be prompted for integration into NHS services to improve access to anxiety assessment for those with cognitive communication impairments. The AISCs and Yale-a will be made freely available and published online. Findings will be disseminated via journal publications, clinician networks, charities, and academic conferences.

Accessible summaries and posters will be produced and shared widely.