Large Language Models for Stuttering Assessment and Therapy

Istanbul Gelisim University5 enrolled

Overview

This observational study aims to compare the quality of responses generated by four large language models (ChatGPT, Claude, Google Gemini, and Grok-4.20) to questions about stuttering. The study focuses on three main areas: general information about stuttering, clinical assessment, and therapy. A total of nine questions were developed based on evidence-based clinical practice guidance, including the American Speech-Language-Hearing Association (ASHA) Practice Portal. Each question is presented to each language model in separate sessions, and the generated responses are recorded for evaluation. Five speech-language therapists with clinical experience in stuttering independently evaluate the model-generated responses. Each response is rated for relevance, accuracy, clarity, completeness, and consistency using a 5-point Likert scale. The responses are also compared with guideline-based reference information. The study does not involve any clinical intervention or treatment of patients. The aim is to determine how closely large language model responses align with current clinical practice guidance and to identify their potential strengths and limitations when used as informational or clinical support tools in the field of stuttering.

Study Type

OBSERVATIONAL

Enrollment

Conditions

Stuttering

Interventions

ChatGPT-Generated Responses

Claude-Generated ResponsesOTHER

Claude responses to nine standardized questions on stuttering information, assessment, and therapy were generated in separate sessions and recorded for independent expert evaluation. Questions were repeated under comparable conditions to allow assessment of response consistency.

Google Gemini-Generated ResponsesOTHER

Google Gemini responses to nine standardized questions on stuttering information, assessment, and therapy were generated in separate sessions and recorded for independent expert evaluation. Questions were repeated under comparable conditions to allow assessment of response consistency.

Grok-4.20-Generated ResponsesOTHER

Grok-4.20 responses to nine standardized questions on stuttering information, assessment, and therapy were generated in separate sessions and recorded for independent expert evaluation. Questions were repeated under comparable conditions to allow assessment of response consistency.

Eligibility

Sex: ALLMin age: 18 YearsHealthy volunteers:

Medical Language ↔ Plain English

Inclusion Criteria: * Having an undergraduate or postgraduate degree in speech and language therapy. * Actively practicing clinically in the field of stuttering/fluency disorders. * Voluntarily agreeing to participate in the study. * Completing the evaluation form in full. Exclusion Criteria: * Not having active clinical experience in the field of stuttering. * Incomplete completion of the evaluation form. * Withdrawal of voluntary participation during the evaluation process.

Outcomes

Primary Outcomes

Expert-Rated Quality of Large Language Model Responses

Responses generated by four large language models (ChatGPT, Claude, Google Gemini, and Grok-4.20) to nine standardized questions about stuttering were independently evaluated by five speech-language therapists with clinical experience in stuttering. Each response was rated using a 5-point Likert scale (1 = strongly disagree to 5 = strongly agree) across five criteria: relevance, accuracy, clarity, completeness, and consistency. Responses were evaluated against guideline-based reference information derived from evidence-based clinical practice guidance. Higher scores indicate better response quality and closer alignment with clinical practice guidance. Mean scores were calculated for each large language model and each evaluation criterion.

Time frame: During the single cross-sectional expert evaluation conducted over approximately 1 month