MedFastReason: LoRA-Adapted LLMs with Chain-of-Thought for Disease Detection, Diagnostic and Treatment Reasoning

In today’s world, the integration of Large Language Models (LLMs) in healthcare continues to expand in the medical sector. Nevertheless, prior research has revealed critical limitations including model hallucinations, limited clinical precision, and inadequate integration between disease detection and treatment planning systems. This study introduces an integrated medical question answering framework that combines disease detection and treatment reasoning in a two-stage architecture. The system leverages state-of-the-art large language models (LLMs) with specialized optimization techniques to address gaps in comprehensive medical AI applications. The Disease Finding Model, built upon Llama 3(8B) with quantization and LoRA adaptation. The model achieved its best performance using the BioMedical-Llama-3-8B configuration enhanced with LoRA, 4-bit quantization, and one-shot prompting. It achieves ROUGE-L scores of 0.75, BERTScore of 0.86, and F1-Token scores of 0.85, while maintaining decent inference times of 85.4ms. Qwen-based Reasoning Medical Model named II Medical leverages Chain-of-Thought prompting to generate stepwise, guideline-compliant treatment plans. It achieves guideline compliance scores of 0.80 and reasoning coherence scores of 0.79. Despite these advances, the system faces limitations, including error propagation from disease misdiagnoses to treatment recommendations. This research advances medical AI by demonstrating that parameter-efficient optimization techniques can maintain clinical-grade performance while enabling deployment on standard hardware. It also brings intelligent clinical decision support systems closer to practical use in real-world healthcare settings.