Assessing large language model responses to pediatric depression FAQs: a cross-sectional study on readability,...

Area of Science:

• Artificial Intelligence in Mental Health
• Pediatric Psychiatry
• Digital Health Information Quality

Background:

• Pediatric depression symptoms vary by age, complicating diagnosis and treatment.
• Parents and adolescents increasingly seek mental health information online, including from AI.
• The quality of online information (readability, accuracy, completeness) is crucial.

Purpose of the Study:

• To compare the suitability of three contemporary large language models (LLMs) as informational tools for pediatric depression.
• To assess the readability, factual accuracy, and completeness of LLM responses to frequently asked questions about pediatric depression.

Main Methods:

• A cross-sectional study analyzed responses from ChatGPT-5, Microsoft Copilot GPT-5, and DeepSeek 3.1V to 15 standardized questions on pediatric depression.
• Readability was assessed using seven indices, accuracy and completeness were scored (0-6 scale), and sentiment analysis was performed.
• Statistical analysis included one-way ANOVA with Tukey post hoc tests.

Main Results:

• Readability varied: DeepSeek 3.1V (Flesch Reading Ease 54-55, Grade Level ~9.5) was easiest to comprehend.
• ChatGPT-5 had intermediate readability (Grade Level ~10.5), while Copilot-5 had the lowest (Grade Level ~10.8).
• Copilot-5 demonstrated the highest accuracy, ChatGPT-5 offered the greatest completeness, and DeepSeek 3.1V excelled in linguistic accessibility.

Conclusions:

• LLMs present diverse capabilities in providing pediatric depression information.
• DeepSeek 3.1V offers better readability, Copilot-5 stronger accuracy, and ChatGPT-5 more comprehensive content.
• AI chatbot information on pediatric mental health requires careful human review and understanding before application.