How large language model responds to common depression questions: A comparative analysis of ChatGPT-4.0, DeepSeek,...

To evaluate the accuracy, comprehensiveness and readability of responses generated by four widely used large language models (LLMs) - ChatGPT-4.0, DeepSeek, Google Gemini and Perplexity - when addressing common depression-related questions.

As patients frequently turn to digital tools for health information, reliable LLMs could play a supportive role in primary care and mental health education. However, their performance in providing accurate and accessible responses to depression-related questions remains underexplored.

Cross-sectional analysis.

Thirty-five depression-related questions (covering pathogenesis, risk factors, clinical presentation, diagnosis, prevention, treatment, prognosis and nursing) were collected from seven authoritative websites. Responses from each LLM were independently evaluated by three psychiatric nurses in a blinded manner, focusing on accuracy and comprehensiveness. R Software was employed for the analysis of readability (Flesch-Kincaid Grade Level, Gunning Fog Index and Flesch Reading Ease Score).

All four LLMs achieved high mean accuracy ratings (ChatGPT-4.0 = 4.67, DeepSeek = 4.62, Google Gemini = 4.65, Perplexity = 4.04). DeepSeek produced the highest proportion of very comprehensive responses (73.3 %), followed by ChatGPT-4.0 (44.8 %), Google Gemini (36.2 %) and Perplexity (6.7 %). Significant differences in readability scores were observed, with DeepSeek and Google Gemini performing less favorably compared with ChatGPT-4.0 (p < 0.05).

LLMs, particularly DeepSeek, show potential as supplementary resources for depression-related health education in primary care and mental health contexts. Nevertheless, further research is needed to confirm their clinical utility, address readability challenges and evaluate their impact on real-world patient outcomes.