AI-powered personalized diverticulitis dietary advisory using localized large language models

Archives of Orthopaedic and Trauma Surgery · Springer Nature · 2026-10-28 · DOI: 10.1007/s43538-026-00967-9

Аннотация

Management of diverticulitis dietary guidelines remains challenging due to the heterogeneous composition of real-world meals, where mixed ingredients and regional variations make accurate assessment difficult for patients. This work proposes a fully offline, privacy-preserving multimodal pipeline that integrates high-resolution meal image understanding with clinically grounded dietary reasoning for personalized diverticulitis management. The proposed system combines a 12-billion-parameter vision–language model (Gemma3:12b) for fine-grained food item detection with a 4-billion-parameter instruction-tuned medical large language model (MedGemma-4b-it) for evidence-based dietary classification. A key design contribution is the incorporation of a user-in-the-loop correction mechanism, enabling refinement of detected items prior to clinical reasoning, thereby improving robustness for complex and region-specific meals. Experimental evaluation on a workstation-class NVIDIA RTX 5000 Ada GPU demonstrates that the average total execution time is approximately 15,911 ms for the vision–language model and 12,925 ms for the medical LLM, with model load durations of 7,197 ms and 2,166 ms, respectively. Prompt evaluation latency is significantly lower for the medical LLM (391 ms) compared to the vision model (1,158 ms), while token generation throughput reaches 92 tokens/s versus 38 tokens/s. Statistical analysis (ANOVA, $$p < 0.001$$ for key metrics) confirms significant differences in load time, prompt latency, and throughput, while total duration differences remain statistically non-significant due to compensatory effects across pipeline stages. While the total end-to-end execution time includes model loading overheads, the interactive inference latency after model initialization remains within a few seconds, enabling responsive user interaction. The system translates detected food items into structured Safe, Caution, and Unsafe categories with clinically interpretable rationales, providing actionable dietary guidance. By operating entirely offline, the framework preserves patient privacy, supports low-connectivity environments, and offers a scalable foundation for future clinical validation and extension to broader gastrointestinal applications. The implementation is available at https://github.com/ParthaPRay/diverticulitis_ollama_LLM .

Читать у издателя →

Похожие статьи