AI-Assisted Electrode Contact Configuration Mapping for Epidural Electrical Stimulation in Spinal Cord Injury: A Comparative Evaluation of Large Language Models
Trial Snapshot
- Phase
- Not Applicable
- Status
- Active, not recruiting
- Sponsor
- Enrollment
- 20
- Locations
- 1
- Primary Endpoint
- Clinical Accuracy Score of Large Language Model Responses
Study Overview
Brief Summary
This observational and methodological study aims to compare the performance of large language models in generating electrode contact configuration recommendations for epidural electrical stimulation in spinal cord injury.
Five standardized synthetic spinal cord injury scenarios will be presented to four large language models: ChatGPT-4o, Claude, Grok 3, and Gemini 2.5 Pro. Each model will receive the same standardized prompt. The generated responses will be anonymized and evaluated independently by experts with experience in spinal cord injury rehabilitation and epidural electrical stimulation.
The responses will be assessed in five main areas: clinical accuracy, technical feasibility, safety awareness, consistency with current clinical guidance, and completeness of the response. Agreement between expert evaluators will also be examined.
No real patients, human participants, clinical interventions, or personal health data are included in this study. The study is designed to explore the potential and current limitations of large language models as artificial intelligence-based clinical decision-support tools in neurorehabilitation.
Study Design
- Study Type
- Observational
- Observational Model
- Other
- Time Perspective
- Cross Sectional
Eligibility Criteria
- Sex
- All
- Accepts Healthy Volunteers
- No
Inclusion Criteria
- •Responses generated for one of the five predefined standardized synthetic spinal cord injury scenarios.
- •Responses generated using the identical standardized prompt specified in the study protocol.
- •Responses generated by one of the four prespecified large language models.
- •Complete responses available for expert evaluation.
Exclusion Criteria
- •Responses generated using prompts that differ from the standardized study prompt.
- •Incomplete, interrupted, or technically corrupted model outputs.
- •Duplicate responses or outputs not corresponding to a predefined synthetic scenario.
- •Any response generated using real patient-identifiable or personal health information.
Arms & Interventions
Gemini 2.5 Pro
Responses generated by Grok 3 for five standardized synthetic spinal cord injury scenarios using the same standardized prompt. The responses will be evaluated for clinical accuracy, technical feasibility, safety awareness, guideline consistency, and completeness.
Intervention: Gemini 2.5 Pro Large Language Model (Other)
Claude
Responses generated by Claude for five standardized synthetic spinal cord injury scenarios using the same standardized prompt. The responses will be evaluated for clinical accuracy, technical feasibility, safety awareness, guideline consistency, and completeness.
Intervention: Claude Large Language Model (Other)
Grok 3
Responses generated by Grok 3 for five standardized synthetic spinal cord injury scenarios using the same standardized prompt. The responses will be evaluated for clinical accuracy, technical feasibility, safety awareness, guideline consistency, and completeness.
Intervention: Grok 3 Large Language Model (Other)
ChatGPT-4o
Responses generated by ChatGPT-4o for five standardized synthetic spinal cord injury scenarios using the same standardized prompt. The responses will be evaluated for clinical accuracy, technical feasibility, safety awareness, guideline consistency, and completeness.
Intervention: ChatGPT-4o Large Language Model (Other)
Outcomes
Primary Outcomes
Clinical Accuracy Score of Large Language Model Responses
Time Frame: At the time of expert evaluation, within 1 week after study initiation
Clinical accuracy of the epidural electrical stimulation electrode contact configuration recommendations generated by each large language model will be independently evaluated by expert reviewers using a 5-point Likert-type rating scale. Higher scores indicate greater clinical accuracy of the generated recommendations.
Secondary Outcomes
- Technical Feasibility Score of Large Language Model Responses(At expert evaluation, within 1 week after study initiation)
- Safety Awareness Score of Large Language Model Responses(At expert evaluation, within 1 week after study initiation)
- Clinical Guideline Consistency Score of Large Language Model Responses(At expert evaluation, within 1 week after study initiation)
- Response Completeness Score of Large Language Model Responses(At expert evaluation, within 1 week after study initiation)
Investigators
Görkem Açar
Director
Istanbul Gelisim University