Enhancing Clinical Efficiency through LLM:
Discharge Note Generation for Cardiac Patients

Discharge notes are essential in the healthcare sector, serving as a comprehensive summary of a patient’s hospital stay, including diagnosis, treatments, and follow-up care recommendations. These documents are crucial for ensuring smooth transitions between care settings, enhancing communication among healthcare providers, and supporting effective patient management after hospitalization. Furthermore, they play a significant role in reducing readmission rates by facilitating the proper management of ongoing care plans, thus maintaining the quality and safety of patient care.

With the advancement of AI, especially natural language processing (NLP) algorithms, there has been a growing interest in their application across various domains, including healthcare. NLP technologies have shown capability in automating tasks that require understanding and generating human language, transforming the way we interact with data and technology. In healthcare, these applications range from automating clinical documentation to enhancing patient interaction with care providers through conversational agents.

LLM has found significant applications in the medical field, including creating clinical notes, interpreting lab results, and anonymizing patient data (Yang et al., 2023a). Their ability to generate human-like text suggests to increase both efficiency and accuracy in medical documentation, potentially reducing the administrative workload for healthcare professionals and allowing them more time for patient care. However, implementing LLM in healthcare faces challenges, notably the linguistic diversity and medical jargon prevalent in multi-cultural hospital settings. This complexity demands LLM to comprehend and interpret medical terminology across different languages and understand the nuances of clinical communication.

Our study underscores the pivotal role of deploying specialized LLM, particularly Mistral-7B, in automating the creation of discharge notes within cardiology, utilizing actual hospital patient data. By expertly navigating the complexities of genuine patient records, Mistral-7B significantly elevates documentation efficiency and continuity. Endorsed by cardiology professional for its clinical relevance and utility, Mistral-7B’s ability to streamline healthcare documentation with real-world data showcases a promising avenue towards integrating AI-driven tools in enhancing patient care and optimizing medical record accuracy. This advancement marks a crucial step forward in leveraging cutting-edge AI to refine healthcare documentation practices directly from the front lines of patient care.

The application of NLP in the medical field has been experiencing a steady increase. This trend is reflected in numerous studies and projects that leverage NLP techniques to extract valuable insights from medical texts, enhance patient care, and facilitate medical research. As such, our work contributes to this growing body of research, further demonstrating the potential of NLP in healthcare.

Clinical Notes Generation The generation of clinical notes, particularly radiology reports (Yang et al., 2023b), represents a critical intersection between NLP and clinical practices. Traditional methods of generating radiology reports have been both time-consuming and tedious for radiologists, prompting the exploration of automated systems. Recent advances have seen the development of multi-modal approaches that combine images, disease labels, and textual data to generate comprehensive radiology reports.

Patient Summary Generation Generating patient summaries (Shing et al., 2021) is an area where NLP is making strides, addressing the need for efficient tools to concisely capture crucial patient information. Recent work proposes an extractive-abstractive summarization pipeline that extracts key sentences from clinical notes and abstracts them into coherent summaries. This approach aids in handling large text volumes while ensuring faithfulness and traceability to original documents. NLP applications in clinical note and patient summary generation can enhance accuracy, efficiency, and comprehensiveness of medical documentation, supporting better patient outcomes and alleviating healthcare professionals’ workload.

Unlike previous study, we directly generate comprehensive discharge records, including chief complaints, medical history, hospital course, discharge status, and follow-up instructions, using large language models. Additionally, we have developed a model specialized in cardiac-related medical terminology and document writing, utilizing a large dataset from a cardiology department of actual hospital.

Large Language Model (LLM) Recent years have witnessed the rise of powerful LLM that have revolutionized various natural language processing tasks, including text generation, summarization, and question answering. These models, trained on vast amounts of textual data, have demonstrated remarkable capabilities in understanding and generating human-like text.

One of the pioneering open source models in this domain is Llama (Touvron et al., 2023), developed by Meta AI. Llama is a family of models ranging from billions to trillions of parameters, trained on a vast corpus of online data. These models have exhibited impressive performance on a wide range of tasks, including open-ended generation, question answering, and code generation. Llama’s scalability and adaptability have made it a popular choice among researchers and developers.

Another notable open soruce LLM is Mistral (Jiang et al., 2023), developed by Mistral AI. This model has demonstrated remarkable performance across numerous benchmarks, outperforming all other models of the same size on many tasks. Notably, it has outperformed the larger-sized model of Llama across various benchmarks. With its impressive performance across various benchmarks, innovative architectural design, and specialization for conversational tasks, Mistral has emerged as a powerful and versatile language model, poised to make significant contributions in natural language processing and its applications.

The findings of this study emphasize the significant potential of fine-tuned LLM, particularly Mistral-7B, in automating the generation of discharge notes for cardiac patients. The quantitative results, evaluated across multiple metrics such as ROUGE, BLEU, BERT Score, and Perplexity, collectively demonstrate the models’ ability to generate accurate, coherent, and clinically relevant discharge notes from patient medical records. While no single model excelled across all metrics, the overall performance highlights the promise of this approach in enhancing healthcare documentation efficiency and continuity of care.

Notably, the qualitative assessment by an expert in cardiology further reinforces the practical utility of the generated discharge notes. The Mistral-7B model exhibited a high degree of accuracy in capturing patients’ medical conditions, treatment processes, and follow-up recommendations. The generated notes were deemed complete, including all essential medical information, while maintaining excellent readability and comprehensibility for the intended audience, including patients, guardians, and healthcare professionals. Additionally, the consistent quality across various patient records and the overall utility in supporting clinical decision-making and care planning underscore the model’s potential for real-world implementation in healthcare settings.

Limitations While the quantitative metrics employed, such as ROUGE, BLEU, BERT Score, and Perplexity, provide valuable insights into the models’ performance, they are not specifically designed to evaluate medical documentation. These metrics may not fully capture the nuances and critical aspects of healthcare documentation, such as adherence to clinical guidelines, appropriate use of medical terminology, and patient safety considerations. The lack of domain-specific, clinically-oriented evaluation metrics poses a challenge in comprehensively assessing the generated discharge notes’ quality and suitability for healthcare applications. Additionally, the lack of standardization in the formatting and style of progress notes poses a challenge for model training and performance. Progress notes are often written in a free-form manner by different physicians, leading to inconsistencies in aspects such as date formatting, abbreviations, and terminology usage. For instance, some physicians may write dates as ”210304,” while others use formats like ”2021/03/04.” Such variations can introduce confusion and inconsistencies during the model’s learning process, potentially impacting its ability to accurately interpret and generate discharge notes. Establishing guidelines or implementing preprocessing steps to standardize the input data could mitigate these issues and improve the model’s performance.

Future works Firstly, the development and integration of domain-specific, clinically-oriented evaluation metrics tailored for assessing medical documentation is crucial. These metrics should capture essential aspects such as adherence to clinical guidelines, appropriate use of medical terminology, and patient safety considerations, enabling a more comprehensive and reliable evaluation of the generated discharge notes.

Another crucial direction for future research is the extension of this approach to other medical specialties beyond cardiology. While this study focused on generating discharge notes for cardiac patients, the methodologies and insights gained could be adapted and applied to other domains, such as oncology, neurology, or pediatrics. By expanding the scope of the dataset to encompass diverse medical conditions and specialties, the models could be fine-tuned to generate discharge notes tailored to the specific requirements and terminologies of each field. This would not only broaden the applicability of the AI-driven documentation approach but also contribute to a more comprehensive and unified system for streamlining medical documentation across various healthcare domains.

Furthermore, expanding the dataset to include data from multiple healthcare facilities and diverse patient populations could significantly enhance the models’ generalizability and robustness, ensuring their applicability across different clinical settings and patient demographics. This could involve establishing collaborations with other healthcare institutions or leveraging existing large-scale medical datasets.

The exploration of multi-modal approaches, incorporating not only textual data but also medical images, lab results, and other relevant patient information, presents a promising direction. By leveraging these diverse data sources, the models could gain a more holistic understanding of patients’ conditions, enabling the generation of more accurate and well-rounded discharge notes. Additionally, the integration of advanced techniques such as attention mechanisms and multimodal fusion could further improve the models’ ability to effectively utilize and combine information from multiple modalities.

Conclusion This study represents a significant step towards leveraging the power of large language models in revolutionizing healthcare documentation practices. The promising results demonstrate the potential for automating discharge note generation, reducing administrative burdens on healthcare professionals, and facilitating seamless continuity of care for patients. However, continued research and development, including the incorporation of domain-specific evaluation metrics and the exploration of multi-modal and knowledge-augmented approaches, are crucial for further refining and advancing these AI-driven solutions in the healthcare domain.