QLoRA-Based Fine-Tuning of LLMs on Multiple Medical Reasoning Tasks to Enhance their Comprehension of Clinical Notes

Clinical notes pertinent to patient discharge summaries are crucial for support and decision-making during medical ministrations. LLMs (Large Language Models) have demonstrated exceptional capabilities in natural language processing. Therefore, exploiting their prowess in coherent knowledge retrieval is desired for clinical decision-making assistance. This study aims to enhance the performance of LLMs in comprehending the intricacies of clinical notes by fine-tuning them on the following medical reasoning tasks: 1) Summarization; 2) Relation Extraction; 3) Coreference Resolution; 4) Temporal Information Extraction; 5) Question-Answering; 6) Abbreviation Expansion; 7) Paraphrasing; 8) Named Entity Recognition. We have investigated the ability of 3 state-of-the-art LLMs to grasp medical reasoning efficiently when fine-tuned by computationally inexpensive QLoRA (Quantised Low-Rank Adaptation) technique. Among Mistral-2 (7B), Llama-3 (8B), and Llama-2 (13B), Llama-3(8B) proves to be adept for medical reasoning by outperforming the other LLMs on average over all the tasks when assessed with ROUGE (Recall Oriented Understudy for Gist Evaluation) and BLEU (Bilingual Evaluation Understudy) metric. Further, we also compare Llama-3’s efficacy post-fine-tuning and pre-fine-tuning on the ability to recall and be precise in generating relevant answers to questions concerning all tasks.