Impact of Query Language on the Structure and Guideline Alignment of AI-Generated Rehabilitation Programs in Chronic Kidney Disease

Background: Artificial intelligence (AI) is increasingly used in nephrology, including rehabilitation planning for patients with chronic kidney disease (CKD). However, most AI systems are predominantly trained on English-language data, which may influence the quality and clinical relevance of the generated recommendations. Objective: To evaluate the impact of query language on the quality and clinical applicability of AI-generated exercise programs for CKD patients undergoing renal replacement therapy. Methods: We conducted a structured qualitative comparison using predefined evaluation criteria based on KDIGO and ERA rehabilitation guidelines. Outputs were assessed for structure, clinical detail, safety framing, and adaptability. Identical prompts were formulated in Ukrainian and English. Generated exercise programs were assessed for alignment with international guidelines (KDIGO, ERA), level of clinical detail, progression, safety considerations, and adaptability. Results: All AI systems produced safe exercise programs incorporating aerobic, resistance, flexibility, and relaxation components. However, significant differences were observed depending on the query language. Ukrainian-language outputs were simpler and focused on general well-being, with limited progression and monitoring. In contrast, English-language outputs demonstrated greater clinical depth, including structured progression, intradialytic adaptations, and the use of validated monitoring tools (e.g., Borg RPE scale). Copilot provided the highest clinical precision, ChatGPT delivered structured programs, and Gemini emphasized safety and motivation. English-language prompts produced more detailed and guideline-aligned outputs, whereas Ukrainian-language prompts generated simpler, wellness-oriented recommendations. Conclusions: Query language influences the structure and clinical completeness of AI-generated rehabilitation programs. English-language prompts currently yield more detailed and guideline-aligned outputs. Further multilingual model development is needed. English-language queries currently yield more clinically robust outputs. Development of multilingual AI systems and standardized prompt frameworks is essential to ensure equitable access to AI-assisted healthcare.