Optimizing discharge summary generation: fine-tuning LLMs by DoRA and iterative self-evaluation for enhanced medical text generation

Optimizing discharge summary generation: fine-tuning LLMs by DoRA and iterative self-evaluation for enhanced medical text generation | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Optimizing discharge summary generation: fine-tuning LLMs by DoRA and iterative self-evaluation for enhanced medical text generation Wenbin Li, Hui Feng, Chao Hu, Minpeng Xu, Longlong Cheng This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7814494/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 17 Jan, 2026 Read the published version in Scientific Reports → Version 1 posted 12 You are reading this latest preprint version Abstract Background The generation of discharge summaries in healthcare is crucial but labor-intensive. Purpose This study introduces a novel framework that enhances the automation of this process using advanced natural language processing (NLP) techniques. Methods We combine the Decomposed Low Rank Adaptation (DoRA) fine-tuning method with a unique self-evaluation mechanism to improve large language models (LLMs) tailored for medical text generation. Our methodology leverages DoRA to efficiently adapt pre-trained LLMs to the specialized medical domain, demonstrating superior performance over traditional methods like LoRA and QLoRA across multiple metrics. We further incorporate a self-evaluation mechanism inspired by cognitive psychology principles, enabling iterative refinement of model outputs to enhance accuracy and completeness. Results This approach is rigorously compared against popular few-shot prompting and Chain of Thought (CoT) methods. Extensive experiments yield significant improvements in both quantitative (ROUGE, BLEU, BERT score) and qualitative (accuracy, completeness, relevance, consistency, and utility) assessments compared to baseline techniques. Conclusions Our results demonstrate substantial enhancements in the quality and consistency of generated discharge summaries while markedly reducing the time required for their creation. This research underscores the potential of AI-driven tools in healthcare documentation, significantly reducing the time required for generating discharge summaries while improving their quality and consistency. The findings indicate promising prospects for automating medical documentation that adheres to high standards of accuracy and relevance. Biological sciences/Computational biology and bioinformatics Health sciences/Health care Physical sciences/Mathematics and computing Natural Language Processing Large Language Models Discharge Summaries Figures Figure 1 Figure 2 I. Introduction Discharge summaries play a critical role in healthcare by providing essential information to both healthcare providers and patients [ 1 ]. For clinicians, these summaries offer a comprehensive overview of the patient’s hospitalization, supporting continuity of care and facilitating future consultations. For patients and their families, discharge summaries serve as valuable documents that clarify the medical care received and outline post-discharge instructions. Despite their importance, the task of writing discharge summaries often receives limited attention due to clinicians’ primary focus on patient care and clinical duties. Consequently, the time constraints faced by healthcare providers frequently lead to discharge summaries that lack sufficient detail and depth, potentially compromising the quality of the medical record [ 2 ]. In fast-paced clinical environments, key details or changes in a patient’s condition maybe inadvertently omitted, undermining the completeness of the medical documentation. Additionally, the absence of standardized formats and guidelines often results in inconsistencies in how medical information is organized, making it challenging for readers to quickly extract critical insights [ 3 ]. The creation of discharge summaries requires experienced clinicians to manually integrate diverse data sources from the Hospital Information System (HIS), including test results, medical orders, pathology reports, and other relevant information, which further complicates the process. Recent advancements in Artificial Intelligence (AI) have rapidly transitioned from theoretical research to practical applications in daily life. Large language models (LLMs), crucial in natural language processing, have shown particular promise in the medical field [ 4 ]. Their capacity for human-like conversational interactions [ 5 ], [ 6 ] has led to successful applications in paper reviewing [ 7 ], lab result interpretation [ 8 ], and patient data anonymization [ 9 ]. However, the implementation of LLMs in healthcare faces significant challenges, primarily due to the disconnect between their training on general linguistic data and the specialized terminology of medicine. While fine-tuning methods like Low Rank Adaptation (LoRA) [ 10 ] have been explored to bridge this gap, they have shown limited effectiveness in adapting to the intricacies of medical texts. Our study addresses these limitations by employing the DoRA (Decomposed Low Rank Adaptation) [ 11 ] fine-tuning method, which better aligns the model with medical- specific terminology and improves fine-tuning efficiency. Furthermore, we introduce a novel self-evaluation mechanism, a pioneering approach in this domain that enables the model to iteratively assess and enhance its output, significantly improving the quality and consistency of generated discharge summaries. This research presents an innovative end-to-end workflow for automating the creation of discharge notes from unstructured breast and thyroid surgery Hospital Information System (HIS) data. By leveraging fine-tuned LLMs and our unique self-evaluation mechanism, we have developed a system that not only reduces the time required by clinical workers but also surpasses existing methods in accuracy and efficiency. Our approach has received endorsement from healthcare professionals for its clinical relevance and utility, underscoring its potential to enhance patient care and improve the precision of medical records. This advancement represents a significant leap forward in integrating AI-driven tools within patient care settings, offering a transformative solution to refine healthcare documentation practices. By addressing the challenges of applying LLMs to specialized medical contexts and introducing a self-improving mechanism, our work paves the way for more accurate, efficient, and context-aware AI applications in healthcare. II. Methodology II.A. Dataset This study utilizes a dataset derived from the Thyroid Surgery Department at a major tertiary hospital in China. The dataset comprises information from 6,214 patients, retrospectively collected from January 1, 2018, to December 31, 2022. The data were accessed and extracted from the hospital’s information system between March 15, 2023, and May 20, 2023 for research purposes. These data are characterized by their fully unstructured and complex nature, presenting significant challenges for analysis and processing. Data collection and processing were conducted under strict adherence to privacy protocols within the hospital’s intranet infrastructure. This approach ensures the preservation of patient confidentiality, with robust measures in place to prevent any potential leakage of sensitive information. The use of this dataset was approved by the Medical Research Ethics Committee of Renmin Hospital of Wuhan University (Ethics Approval Number: WDRY2024-K038), under the research protocol titled 'Integrated Multi-Omics and Artificial Intelligence Analysis in Thyroid Nodule Patients: Prediction, Subtype Identification, and Exploration of Recurrence Factors'. The Clinical Research Ethics Committee of Renmin Hospital of Wuhan University approved the conduct of the clinical study in accordance with the research protocol and the waiver of informed consent application. As illustrated in Fig. 1 , the process of transforming unstructured data from the HIS into structured discharge summaries involves several critical steps. These steps include integration, deduplication, privacy filtering, and normalization. Data within the HIS are distributed across multiple tables, such as the Doctor’s Advice Information Table, Clinical Lab Information Table, Nursing Information Table, Physical Signs Information Table, Blood Glucose Information Table, Inspection Information Table, Diagnosis Table, Pathology Information Table, and Progress Notes Table. Given the interrelated nature of these data sources and the presence of duplicative entries, deduplication is a necessary step to ensure data consistency and integrity. We employ a MinHash algorithm for efficient deduplication, effectively identifying and removing redundant data. Privacy concerns are addressed through the use of regular expression matching techniques, which filter out sensitive information from the data sets. Subsequent steps involve preparing the textual data for analysis, which includes spelling correction, acronym expansion, and the standardization of medical terminology. These pre-processing tasks are essential for enhancing the quality and usability of the data, ultimately facilitating the summarization task performed by LLMs. II.B. Model Recent advancements in language models have shown significant potential for medical applications. Mistral [ 12 ], as discussed in [ 13 ], has been particularly noted for its effectiveness in generating discharge summaries [ 13 ]. Similarly, Llama 3 has demonstrated robust performance across a variety of LLM benchmarks [ 14 ]. Due to our focus on Chinese medical data and the limited availability of Chinese corpus in previously used models, we also consider incorporating Qwen2 [ 15 ], [ 16 ] to better address linguistic and regional data specifics. II.C. Parameter Efficient Fine Tuning (PEFT) The generation of discharge summaries typically adheres to a strict output format, necessitating fine-tuning of the LLM to produce standardized summaries. However, strict data protection regulations within hospital environments often restrict the ability to conduct extensive model training on external systems. Consequently, hospitals typically lack the necessary computational resources, such as GPUs, to perform full-parameter fine-tuning of large language models. This limitation underscores the importance of Parameter Efficient Fine Tuning (PEFT) in these settings. Previous studies have extensively discussed methods such as LoRA and QLoRA, which utilize the low-rank factorization of weight matrices: W0 = W0 + B *AT However, these LoRA-based methods possess inherent drawbacks. For example, they tend to increase the general magnitude of the weight matrix components differently compared to traditional full-parameter fine-tuning [ 11 ]. This discrepancy can lead to suboptimal fine-tuning outcomes, particularly in fields like medicine, where precise terminology is crucial. To address these issues, we employ the Weight-Decomposed Low-Rank Adaptation (DoRA) method, which decomposes weights into direction and magnitude components. This approach more closely approximates full-parameter fine-tuning, enhancing the model’s ability to accurately handle specialized medical terminology and context. II.D. Self-Evaluation Mechanism In cognitive psychology, rapid, intuitive judgments are often likened to Daniel Kahneman’s 140 ‘System 1 thinking,’ which operates subconsciously. In contrast, tasks that require deliberate and methodical thought, such as mathematical calculations and logical reasoning, are categorized as ‘System 2 thinking.’ Similarly, the output of Large Language Models (LLMs), especially in tasks like generating discharge summaries, can be initially rapid and intuitive, mirroring ‘System 1 thinking.’ However, ensuring the reliability and completeness of these outputs often necessitates a more deliberate review process, akin to ‘System 2 thinking.’ Despite this, our research demonstrates that LLMs are capable of self-evaluating their outputs to identify and correct omissions. This capability is facilitated by our novel self-evaluation mechanism, the general procedure of which is depicted in Fig. 2 . Initially, the LLM processes the diverse information extracted from HIS data to generate a preliminary discharge summary. Concurrently, information segmentation tools decompose the original input into distinct entities and events. Upon generating the initial summary, this output, along with segments of the original data, is re-fed into the LLM. This process allows the model to determine whether it can locate the original data’s details within the discharge summary. If details are found to be missing, the model identifies this as an omission and adjusts the summary accordingly. This iterative process continues for each data segment. Finally, the most recent summary and the complete original input are reintegrated into the LLM to assess if the enhanced summary satisfactorily represents the final, accurate discharge report. Algorithm 1 Self-Evaluation Mechanism for Discharge Summary Generation Require Extracted HIS notes X, Target Model M, Prompt for discharge generation P, Evaluation Prompt Ep R ← M(P ∥ X) ▷ Generate initial result isComplete ← False while not isComplete do E ← Information__Decompose(X) ▷ Decompose input for each ei in E do complete ← M(Ep ∥ ei ∥ R) ▷ Evaluate if not complete then R ← M(R ∥ ei ) ▷ Update result end if end for complete ← M(Ep ∥ X ∥ R) ▷ Final evaluation if complete then isComplete ← True end if end while return R ▷ Final discharge summary II.E. Experiment Our experimental setup involved the deployment of several pre-trained models to evaluate the efficacy of different fine-tuning approaches. We used the Qwen2-7B model with LoRA fine-tuning and direct few-shot prompting as our baseline for comparison. Additionally, we explored the performance of other pre-trained models, specifically Mistral 7B and Llama3 8B, to understand their behavior under similar experimental conditions. To provide a comprehensive comparison, we employed DoRA fine-tuning on the same dataset used with the baseline model. The specifics of our experimental configuration, including the details of fine-tuning parameters, are documented in Table 1 . For evaluation, we utilized several methods to assess the effectiveness of the fine-tuned models. These included: • Direct few-shot prompting, which tests the model’s ability to generate summaries based on a limited number of example inputs. • Chain of Thought (CoT) prompting [ 17 ], designed to simulate a step-by-step reasoning process in generating the output. The detailed structure follows the [ 18 ]. • Our novel approach termed ‘Self-Evaluation Generation’, which incorporates an iterative feedback mechanism to refine the model’s output continuously. These methods were chosen to explore different aspects of model performance, particularly focusing on their capability to generate accurate and comprehensive discharge summaries under varied prompting conditions Table 1 Training parameters LoRA Parameters LoRA__alpha LoRA__dropout LoRA rank bias 16 0.1 64 None DoRA Parameters DoRA__alpha DoRA__dropout DoRA rank bias 16 0.1 64 None Other Parameters epoch batch size warmup ratio weight decay 5 64 5% 0.05 II.F. Evaluation Metrics To rigorously assess the effectiveness of the fine-tuning techniques and the quality of the generated discharge summaries, we employed a suite of quantitative and qualitative metrics. These metrics were chosen to provide a comprehensive view of both linguistic accuracy and content relevance. II.F.1. Quantitative Metrics The “self-evaluation” method consistently achieves the highest scores in several critical metrics: We utilized the following standardized metrics to evaluate the linguistic and semantic quality of the model outputs: ROUGE (Recall-Oriented Understudy for Gisting Evaluation) This metric measures the overlap of n-grams between the generated text and the reference texts. It is particularly useful for assessing the extent to which important information is captured by the generated summaries. The ROUGE score is given by • BLEU (Bilingual Evaluation Understudy) : This metric evaluates the correspondence of machine-generated text to human-written text, focusing on precision of n-grams. It is computed as: where BP is the brevity penalty, are the weights assigned to each n-gram, and are the n-gram precisions. • BERT Embedding Similarities : We measure the cosine similarity between the embeddings of the generated text and the reference text, derived from a pre-trained BERT model. This metric assesses the semantic similarity beyond surface lexical matching. • Perplexity : Typically used to measure the likelihood of the sequence of words in the generated text, given by: where is the model’s probability distribution over the entire text. II.F.2. Qualitative Metrics In addition to automated metrics, we also conducted human evaluations to assess the practical utility of the generated summaries. Trained evaluators rated the outputs based on the following five dimensions: • Accuracy : The factual accuracy of the information presented. • Completeness : The extent to which all relevant information is included. • Relevance & Clarity (R&C) : How relevant and clear the information is with respect to the patient’s medical history and treatment. •Consistency : The internal consistency of the information within the summary itself. •Utility : The practical usefulness of the summary for continuing care and patient understanding. These metrics together enable a holistic evaluation of our model’s performance, encompassing both the technical efficacy of the fine-tuning methods and the practical applicability of the generated texts in real-world medical settings. III. Results III.A. Analysis of Fine-Tuning Methods The results, as depicted in Table 2 , highlight significant differences in performance across various fine-tuning methods applied to different models. With the same LoRA rank and training epochs, the DoRA method generally outperforms both LoRA and QLoRA across most evaluation metrics. This trend is consistent across different models, underscoring the robustness of the DoRA method in optimizing model performance. III.A.1. Model-Specific Performance Qwen2-7B Model The Qwen2-7B model, when fine-tuned with DoRA, achieved competitive results in almost all metrics, with particularly noteworthy improvements in BERT score and Perplexity, indicating enhanced semantic understanding and lower model confusion, respectively. The LoRA fine-tuning method showed a slight advantage in Rouge1 scores, which may be attributed to the model’s intrinsic handling of Chinese corpus. However, DoRA consistently demonstrated superior overall performance, suggesting that its fine-tuning mechanism is particularly effective for comprehensive text analysis. Mistral 7B Model For the Mistral 7B model, DoRA fine-tuning resulted in substantial improvements over both LoRA and QLoRA methods. Although the gains were less pronounced compared to the Qwen2-7B model, DoRA still enhanced the BLEU score and BERTscore significantly, pointing to improved translation fidelity and semantic accuracy. Llama3 8B Model The Llama3 8B model exhibited a different pattern, with LoRA performing relatively better than QLoRA but still underperforming compared to DoRA in terms of BERT score and Perplexity. This indicates that DoRA is more effective in reducing perplexity and enhancing semantic coherence in outputs, which is crucial for medical summary applications. III.A.2. Comparative Evaluation The detailed results in Table 2 underline the efficacy of DoRA fine-tuning across multiple models and metrics. The DoRA method not only consistently yields lower perplexity scores, indicating a better fit to the data—but also achieves higher scores in both Rouge2 and RougeL, which are critical for capturing finer nuances and ensuring the logical flow of generated texts. Furthermore, DoRA’s ability to outperform in terms of BLEU and BERT score across different models reinforces its applicability in enhancing the accuracy and relevance of generated medical texts. In summary, these results validate the hypothesis that DoRA fine-tuning significantly enhances the quality of discharge summaries generated by LLMs. The consistent outperformance of DoRA across diverse models and metrics emphasizes its potential in medical text generation, particularly in contexts demanding high accuracy and coherence. Table 2 Influence of different fine-tuning methods Method Rouge1 Rouge2 RougeL BLEU BERTscore Perplexity Qwen2-7B + LoRA 0.461 0.343 0.372 0.14 0.851 1.321 Qwen2-7B + QLoRA 0.209 0.191 0.193 0.12 0.829 1.64 Qwen2-7B + DoRA 0.463 0.350 0.391 0.15 0.866 1.278 Mistral 7B + LoRA 0.222 0.195 0.223 0.07 0.753 1.933 Mistral 7B + QLoRA 0.210 0.174 0.221 0.07 0.722 2.434 Mistral 7B + DoRA 0.242 0.221 0.241 0.12 0.831 1.792 Llama3 8B + LoRA 0.397 0.191 0.202 0.08 0.745 1.992 Llama3 8B + QLoRA 0.377 0.163 0.194 0.06 0.721 2.370 Llama3 8B + DoRA 0.239 0.186 0.224 0.11 0.822 1.839 III.B. Analysis of Generating Methods The comparative analysis of different generating methods is captured in Table 3 . This analysis demonstrates the superior performance of the “self-evaluation” method over traditional few-shot and Chain of Thought (CoT) prompting methods across several metrics. III.B.1. Quantitative Performance The “self-evaluation” method consistently achieves the highest scores in several critical metrics: • Rouge Scores : It surpasses other methods in Rouge1 and Rouge2 scores, indicative of better extraction and reproduction of key phrases and sentences from the source text. • BLEU and BERTscore : It shows notable improvements in BLEU scores, reflecting better translation fidelity and syntactic coherence. The BERTscore, which assesses the semantic similarity between the generated texts and the references, is also highest for the “self-evaluation” method, underlining its effectiveness in maintaining semantic integrity. III.B.2. Qualitative Performance The “self-evaluation” method consistently achieves the highest scores in several critical metrics: • Rouge Scores : It surpasses other methods in Rouge1 and Rouge2 scores, indicative of better extraction and reproduction of key phrases and sentences from the source text. • BLEU and BERTscore : It shows notable improvements in BLEU scores, reflecting better translation fidelity and syntactic coherence. The BERTscore, which assesses the semantic similarity between the generated texts and the references, is also highest for the “self-evaluation” method, underlining its effectiveness in maintaining semantic integrity. In addition to automated metrics, the “self-evaluation” method demonstrates remarkable performance in human -evaluated dimensions: • Completeness and Accuracy : Achieving the highest ratings in these categories, the method proves its capability in generating comprehensive and accurate summaries. Table 3 Comparison of different generating methods Method Rouge1 Rouge2 RougeL BLEU BERTscore Accuracy Completeness R & C Consistency Utility Qwen2-7B + few-shot 0.463 0.350 0.391 0.15 0.866 4.0 3.9 3.2 4.1 4.1 Mistral-7B + few-shot 0.242 0.221 0.241 0.12 0.831 3.9 3.5 3.2 4.1 4.2 Llama3-8B + few-shot 0.239 0.186 0.224 0.11 0.822 3.8 3.5 4.2 3.8 4.3 Qwen2-7B + CoT 0.487 0.373 0.446 0.20 0.897 4.2 4.2 4.2 4.2 4.3 Mistral-7B + CoT 0.399 0.343 0.450 0.19 0.837 4.3 4.3 4.2 4.4 4.3 Llama3-8B + CoT 0.343 0.321 0.462 0.18 0.854 4.1 4.2 4.2 4.0 4.5 Qwen2-7B + self-evaluation 0.512 0.398 0.451 0.24 0.923 4.7 4.9 4.5 4.2 4.5 Mistral-7B + self-evaluation 0.424 0.350 0.451 0.25 0.889 4.6 4.9 4.6 4.3 4.4 Llama3-8B + self-evaluation 0.356 0.344 0.462 0.22 0.882 4.7 4.8 4.5 4.3 4.5 • Utility : With top scores in utility, the method validates its practical applicability in real-world settings, providing outputs that are highly useful for medical professionals and patients alike. The self-evaluation method not only performs consistently across different LLMs (Qwen2-7B, Mistral-7B, Llama3-8B) but also shows robustness in handling diverse datasets and operational contexts. This is evidenced by its superior performance in both linguistic and content-specific evaluations, suggesting its adaptability and effectiveness across different medical textual environments. III.B.3. Implications The superior results of the self-evaluation method, as documented in the table, underscore its potential as a significant advancement in medical text generation. Its ability to iteratively refine the output using an internal feedback mechanism allows for a level of precision and relevance that traditional methods struggle to achieve. This is particularly valuable in medical settings where accuracy and detail are paramount. The detailed results presented in Table 3 validate the effectiveness of the self-evaluation method in generating high-quality medical texts, making it a promising approach for enhancing automated medical documentation systems. IV. Discussion IV.A.Summary of Key Findings This study yields several significant findings regarding the enhancement of medical text generation through advanced fine-tuning methods and iterative self-evaluation mechanisms: The DoRA method shows marked advantages over traditional fine-tuning approaches such as LoRA and QLoRA, particularly in addressing the specialized linguistic and structural demands of medical texts. DoRA achieves superior performance across multiple metrics (ROUGE, BLEU, BERTscore), indicating enhanced adaptation to the unique challenges posed by medical terminology and context. This method’s fine-tuning mechanism enables more efficient integration of domain-specific knowledge, resulting in more coherent, relevant, and semantically accurate discharge summaries. The introduction of a self-evaluation mechanism significantly elevates the quality of generated medical texts. This iterative review and refinement process substantially reduces omissions and factual inconsistencies. The mechanism not only enhances the completeness and relevance of discharge summaries but also increases overall model robustness by ensuring better alignment between generated content and original input data. Substantial improvements in both automated and human-evaluated metrics, including accuracy, consistency, and utility, underscore the effectiveness of this approach. Among the models tested, the Qwen2 7B model consistently outperforms others in this domain, primarily due to its superior proficiency in Chinese language processing. This finding highlights the importance of language-specific capabilities in medical text generation tasks. IV.B. Interpretation and Significance of Results Analysis of training results, particularly perplexity metrics, reveals that the DoRA method, which incorporates magnitude of weight training, facilitates more comprehensive learning compared to LoRA methods. This leads to substantial improvements, especially for models like Llama and Mistral, which initially possessed less knowledge of Chinese content. The addition of weight magnitude training in DoRA appears to enhance the model’s ability to adapt to unfamiliar domains, suggesting a more robust fine-tuning approach for specialized applications. The effectiveness of our self-evaluation mechanism aligns with recent findings in the field. As noted by [ 19 ], self-evaluation capabilities in Large Language Models (LLMs) can serve as a defense mechanism against adversarial attacks, indicating LLMs’ ability to discern true from false information. Furthermore, [ 20 ] has demonstrated the potential of LLMs as critics for their own outputs. Our research provides comprehensive empirical evidence supporting the feasibility and efficacy of LLM self-evaluation and self-improvement in the context of medical text generation. The combined application of DoRA fine-tuning and self-evaluation mechanisms represents a significant advancement in automating medical documentation. This approach not only improves the efficiency of discharge summary generation but also enhances the quality and reliability of the generated content. Such improvements have the potential to significantly reduce the administrative burden on healthcare professionals while maintaining high standards of accuracy and completeness in medical records. IV.C. Limitations of the Study While this study demonstrates significant improvements in the generation of medical texts, there are several limitations that should be acknowledged: Limitations of the dataset The dataset used in this study is specific to a single hospital and focuses primarily on discharge summaries from the Breast and Thyroid Surgery Department. As a result, the model’s performance may not generalize well to other medical departments or hospitals with different patient populations, medical practices, or clinical documentation formats. This specificity limits the scope of the findings and highlights the need for broader evaluations across diverse datasets and medical specialties to validate the robustness of the proposed approach. Potential biases or errors in models and methods The fine-tuning and self-evaluation mechanisms employed in this study are dependent on the quality and structure of the training data. Any inherent biases or errors present in the dataset, such as inconsistencies in documentation or variations in medical terminology, could lead to biased model outputs. Furthermore, the iterative self-evaluation mechanism, while improving text quality, may introduce compounding errors if the model’s initial outputs contain inaccuracies that are reinforced during the refinement process. These limitations underline the importance of rigorous data preprocessing and error mitigation strategies. Considerations of computational resources and time costs The DoRA fine-tuning method and the self-evaluation mechanism, while efficient, still require considerable computational resources, particularly in a hospital environment where high-end hardware, such as GPUs, may not be readily available. The iterative nature of the self-evaluation process increases time costs, which may limit the real-time applicability of the approach in settings where rapid document generation is needed. Balancing the trade-off between improved accuracy and resource constraints remains a challenge, particularly for smaller healthcare institutions with limited computational infrastructure. IV.D. Future Research Directions Building upon the findings and limitations of this study, several future research directions could further enhance the field of medical text generation: Extension to other types of medical texts: Future research could explore the application of the DoRA fine-tuning method and self-evaluation mechanism to generate other types of medical documents, such as clinical progress notes, radiology reports, or surgical summaries. Expanding the scope to include various forms of medical documentation could help verify the generalizability of the proposed approach and uncover additional domain- specific challenges. Exploration of multimodal inputs Incorporating multimodal data, such as medical images, audio recordings, or lab results, alongside textual inputs could provide a richer context for generating more accurate and comprehensive summaries. Research into how large language models can integrate and process multimodal information effectively would represent a significant advancement in automated medical documentation. Further optimization of the self-evaluation mechanism The current self-evaluation process could be optimized to reduce computational costs and time requirements. Future work could focus on developing more efficient evaluation strategies that maintain high levels of accuracy while improving the speed of iteration. Additionally, integrating external validation sources, such as medical guidelines or clinical databases, could help the model better assess its outputs. Study of long-term impacts on medical practice Longitudinal studies could examine the practical implications of integrating AI-generated discharge summaries into real -world clinical workflows. Research could focus on how such systems affect clinician workload, patient care quality, and documentation accuracy over time, as well as the potential legal and ethical considerations of relying on AI-generated medical texts. Development of adaptive fine-tuning methods Investigating new fine-tuning techniques that allow models to adapt dynamically to evolving medical knowledge and practices could further improve the utility of LLMs in healthcare. Adaptive methods could help maintain model relevance in the face of new medical guidelines or treatments without requiring full retraining. V. Conclusions This study demonstrates significant advancements in the generation of medical discharge summaries by integrating large language models with the Decomposed Low-Rank Adaptation fine-tuning method. Compared to traditional techniques like LoRA and QLoRA, DoRA excels in handling the complexities of medical terminology and context, achieving superior performance in terms of accuracy, coherence, and semantic consistency across various evaluation metrics. Moreover, the introduction of the self-evaluation mechanism allows the model to iteratively assess and refine its outputs, minimizing omissions and improving the completeness and accuracy of the generated texts. This iterative process ensures that the final discharge summaries are more aligned with the original input data and meet higher standards of clinical relevance and utility. The findings of this study suggest that AI-driven approaches can significantly reduce the time required for generating high-quality medical documentation, while also enhancing the consistency and reliability of these documents. This research paves the way for future exploration in automating various types of healthcare documentation, improving both efficiency and quality in clinical workflows. Declarations Conflict of Interest Statement The authors have no relevant conflicts of interest to disclose. Funding Declaration The authors did not receive support from any organization for the submitted work. Data Availability Declaration Due to ethical restrictions, the raw data cannot be made publicly available. However, de-identified data may be obtained from the first author upon reasonable request. Author Contribution Wenbin Li conceptualized the research idea and wrote the main manuscript text. Hui Feng and Minpeng Xu conducted the literature review and contributed to drafting and editing portions of the manuscript. Chao Hu and Longlong Cheng compiled and analyzed the latest developments in machine learning techniques, wrote the technical methods section, and assisted in final manuscript revisions. All authors reviewed and approved the final version of the manuscript. References Wilson, S., Ruscoe, W., Chapman, M., Miller, R. & Chen General practitioner–hospital communications: a review of discharge summaries, Journal of quality in clinical practice 21, 104–108 W.-K. Linear Networks and Systems. Belmont, CA: Wadsworth, 1993, pp. 123–135. (2001). van Walraven, C. & Rokosh, E. What is necessary for high-quality discharge summaries? Am. J. Med. Qual. 14 , 160–169 (1999). Neira, R. A. Q., de Vries, G. J., Caffarel, J. & Stretton, E. Extraction of data from a hospital information system to perform process mining, in MEDINFO 2017: Precision Healthcare through Informatics pages 554–558 (IOS, 2017). Thirunavukarasu, A. J. et al. Large language models in medicine. Nat. Med. 29 , 1930–1940 (2023). Brown, T. B. Language models are few-shot learners, arXiv preprint arXiv:2005.14165(2020). Floridi, L., Chiriatti, M., GPT-3. & Its nature, scope, limits, and consequences. Mind. Mach. 30 , 681–694 (2020). Zhou, R., Chen, L. & Yu, K. Is LLM a Reliable Reviewer? A Comprehensive Evaluation of LLM on Automatic Paper Reviewing Tasks, in Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation 430 (LREC-COLING 2024), pages 9340–9351, (2024). He, Z. et al. Quality of Answers of Generative Large Language Models Versus Peer Users for Interpreting Laboratory Test Results for Lay Patients: Evaluation Study. J. Med. Internet. Res. 26 , e56655 (2024). Wiest, I. C. et al. Anonymizing medical documents with local, privacy preserving large language models: The LLM-Anonymizer, medRxiv, 2024–06 (2024). Hu, E. J. et al. Lora: Low-rank adaptation of large language models, arXiv preprint arXiv:2106.09685(2021). Liu, S. Y. et al. Dora: Weight-decomposed low-rank adaptation, arXiv preprint arXiv:2402.09353(2024). Jiang, A. Q. et al. Mistral 7B, arXiv preprint arXiv:2310.06825 (2023). Jung, H. et al. Enhancing Clinical Efficiency through LLM: Discharge Note Generation for Cardiac Patients, arXiv preprint arXiv:2404.05144 (2024). Dubey, A. et al. The llama 3 herd of models, arXiv preprint arXiv:2407.21783 (2024). Bai, J. et al. Qwen technical report, arXiv preprint arXiv:2309.16609 (2023). Yang, A. et al. Qwen2 technical report, arXiv preprint arXiv:2407.10671 (2024). Wei, J. et al. Chain-of-thought prompting elicits reasoning in large language models. Adv. Neural. Inf. Process. Syst. 35 , 24824–24837 (2022). Tang, A. Q., Zhang, X. & Dinh, M. N. Ignition Innovators at Discharge Me! Chain-of-Thought Instruction Finetuning Large Language Models for Discharge Summaries, arXiv preprint arXiv:2407.17636 (2024). Brown, H., Lin, L., Kawaguchi, K. & Shieh, M. Self-Evaluation as a Defense Against Adversarial Attacks on LLMs, arXiv preprint arXiv:2407.03234 (2024). McAleese, N. et al. Llm critics help catch llm bugs, arXiv preprint arXiv:2407.00215 (2024). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 17 Jan, 2026 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 06 Nov, 2025 Reviews received at journal 31 Oct, 2025 Reviews received at journal 24 Oct, 2025 Reviewers agreed at journal 24 Oct, 2025 Reviews received at journal 23 Oct, 2025 Reviewers agreed at journal 23 Oct, 2025 Reviewers agreed at journal 23 Oct, 2025 Reviewers invited by journal 23 Oct, 2025 Editor assigned by journal 23 Oct, 2025 Editor invited by journal 14 Oct, 2025 Submission checks completed at journal 11 Oct, 2025 First submitted to journal 11 Oct, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7814494","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":538662218,"identity":"2e194ee3-9a18-49ba-978a-916563936b98","order_by":0,"name":"Wenbin Li","email":"","orcid":"","institution":"Tianjin University","correspondingAuthor":false,"prefix":"","firstName":"Wenbin","middleName":"","lastName":"Li","suffix":""},{"id":538662219,"identity":"2c34d5f2-ad89-4be3-9636-5f6725df078d","order_by":1,"name":"Hui Feng","email":"","orcid":"","institution":"Renmin Hospital of Wuhan University","correspondingAuthor":false,"prefix":"","firstName":"Hui","middleName":"","lastName":"Feng","suffix":""},{"id":538662220,"identity":"84649218-d993-4594-9b34-674f8273c32f","order_by":2,"name":"Chao Hu","email":"","orcid":"","institution":"China Electronics Cloud Technology Co., Ltd","correspondingAuthor":false,"prefix":"","firstName":"Chao","middleName":"","lastName":"Hu","suffix":""},{"id":538662221,"identity":"bbb99029-1155-4b4f-8a21-1e423bb104e7","order_by":3,"name":"Minpeng Xu","email":"","orcid":"","institution":"Tianjin University","correspondingAuthor":false,"prefix":"","firstName":"Minpeng","middleName":"","lastName":"Xu","suffix":""},{"id":538662222,"identity":"46d2e517-5e9d-42f1-87b0-d80a175cd97e","order_by":4,"name":"Longlong Cheng","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA2klEQVRIie3RsQrCQAyA4SsnmaJdW4T6BMJBQeliX6VFcBJRXLp5UNBF36APIQjFMXLQSXF1FcHJwe6Crjr13ATv3z+SEMZMph/M5px25QNm4SIlPeIuIFaubEQCi0iPiCP6VJdeJJyh0NxMoSB32xl12bA83VjPa8sKYaUwpsl+MA3kYRNkrO93qIJwztfkQGHJ3SpvIqM4ryLAmSCEpyUVXvUI8pqg+hzidYGgRxwOkXL24LtL8INMaNwSHpUq7wl4dutyPt2SnldJPkei7mveyLfCZDKZ/qIXLt1D62etfzwAAAAASUVORK5CYII=","orcid":"","institution":"Tianjin University","correspondingAuthor":true,"prefix":"","firstName":"Longlong","middleName":"","lastName":"Cheng","suffix":""}],"badges":[],"createdAt":"2025-10-09 07:53:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7814494/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7814494/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-026-35552-z","type":"published","date":"2026-01-17T16:31:12+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":95224413,"identity":"ed50a79e-39e0-44cd-9e98-d8edbf7fd5fb","added_by":"auto","created_at":"2025-11-05 16:23:42","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":329107,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/8fbde0b927793a2592513b99.docx"},{"id":95225681,"identity":"43eb67ed-a078-402b-89fd-a2e6e8cfc16f","added_by":"auto","created_at":"2025-11-05 16:25:23","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7003,"visible":true,"origin":"","legend":"","description":"","filename":"bf1d4cc8d0b940ac92ef22c76cceedef.json","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/8bbbf623fa46d4aa2caa1683.json"},{"id":95112825,"identity":"c031e78a-3655-48de-81dd-67baa0644386","added_by":"auto","created_at":"2025-11-04 12:20:53","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":89412,"visible":true,"origin":"","legend":"","description":"","filename":"bf1d4cc8d0b940ac92ef22c76cceedef1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/376523cae71cbfecab3b8193.xml"},{"id":95224863,"identity":"92ea3175-9bbb-482e-a54b-2df7041da7b6","added_by":"auto","created_at":"2025-11-05 16:24:25","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":112800,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/31764226ed3d0f39a204570f.png"},{"id":95112823,"identity":"8f1a76ec-9e32-41aa-ad33-384aefb1a4c3","added_by":"auto","created_at":"2025-11-04 12:20:53","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":68,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/e76a5875cd2a7d6c2cb3c564.png"},{"id":95226182,"identity":"fc2d0882-4411-4a61-bd3d-2791c6dd036e","added_by":"auto","created_at":"2025-11-05 16:30:34","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":113642,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/61b8ddee0d3c50c0d527b590.jpeg"},{"id":95224167,"identity":"17abf0a7-469a-43d5-beb0-9a0215fdda6c","added_by":"auto","created_at":"2025-11-05 16:23:28","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":27210,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/07f6175284098c9c116b030f.png"},{"id":95115487,"identity":"dd4f61e2-f09e-48d0-ae7b-848ff9463017","added_by":"auto","created_at":"2025-11-04 12:52:22","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11797,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/4f006f75207ea0f1d9ec01df.png"},{"id":95112828,"identity":"9c16b6ff-f19d-4ceb-a291-1357dc541139","added_by":"auto","created_at":"2025-11-04 12:20:53","extension":"jpeg","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6342,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/1c3d9cfc961f6b08e1458717.jpeg"},{"id":95224878,"identity":"5080c6c3-f288-4476-8219-e2ee91d4546f","added_by":"auto","created_at":"2025-11-05 16:24:25","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":24534,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/d6996b6aba9ced82c5c7585d.png"},{"id":95112834,"identity":"c8c12e27-f413-44d9-814b-42fb7d595a94","added_by":"auto","created_at":"2025-11-04 12:20:53","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":308,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/14a2d8381467ed0ffb108828.png"},{"id":95112837,"identity":"717223b5-c918-45a5-b84b-32590b72eae4","added_by":"auto","created_at":"2025-11-04 12:20:53","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":86448,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/a15331429fe8da6bfb4a29b3.png"},{"id":95112832,"identity":"e5def32e-1bf8-4164-a302-fafe119967fd","added_by":"auto","created_at":"2025-11-04 12:20:53","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3605,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/4bb5ee9b22fb1ab921a6ce55.png"},{"id":95224541,"identity":"a386de1a-067b-4b9a-9f25-cacd237ce860","added_by":"auto","created_at":"2025-11-05 16:23:53","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1986,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/d6f0aaf7b3bfefbeed2a5f33.png"},{"id":95115486,"identity":"5cc045e7-5cfe-4bfc-acda-30ebd7c0e348","added_by":"auto","created_at":"2025-11-04 12:52:21","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2063,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/a860b1c86f8a1d0272e4383f.png"},{"id":95112839,"identity":"5995ae96-d897-4a25-8a0b-30a9e1a7179d","added_by":"auto","created_at":"2025-11-04 12:20:53","extension":"xml","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":85769,"visible":true,"origin":"","legend":"","description":"","filename":"bf1d4cc8d0b940ac92ef22c76cceedef1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/36c6d47e9b92991f631d8a1e.xml"},{"id":95112838,"identity":"33920004-e45c-438e-8d0c-a436d90d5b4c","added_by":"auto","created_at":"2025-11-04 12:20:53","extension":"html","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":97109,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/560bf732cbf31a5b79d82284.html"},{"id":95112821,"identity":"727411ac-5774-4636-b00a-118720b02ff2","added_by":"auto","created_at":"2025-11-04 12:20:53","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":172383,"visible":true,"origin":"","legend":"\u003cp\u003eData Processing Pipeline for Discharge Summary Generation from Unstructured HIS Data.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/28798d1efea6f09429679ac0.png"},{"id":95112824,"identity":"d138a4aa-7759-49e0-a792-3f349bd79162","added_by":"auto","created_at":"2025-11-04 12:20:53","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":619639,"visible":true,"origin":"","legend":"\u003cp\u003eThe general sketch of self-evaluation mechanism\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/14820c0fc33c532398503a20.png"},{"id":100614953,"identity":"3462b49f-86f4-4653-88eb-74ba84da5118","added_by":"auto","created_at":"2026-01-19 17:28:52","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1890801,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7814494/v1/cf948a75-abcb-40b8-b836-67cf8892ef37.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Optimizing discharge summary generation: fine-tuning LLMs by DoRA and iterative self-evaluation for enhanced medical text generation","fulltext":[{"header":"I. Introduction","content":"\u003cp\u003eDischarge summaries play a critical role in healthcare by providing essential information to both healthcare providers and patients [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. For clinicians, these summaries offer a comprehensive overview of the patient\u0026rsquo;s hospitalization, supporting continuity of care and facilitating future consultations. For patients and their families, discharge summaries serve as valuable documents that clarify the medical care received and outline post-discharge instructions. Despite their importance, the task of writing discharge summaries often receives limited attention due to clinicians\u0026rsquo; primary focus on patient care and clinical duties. Consequently, the time constraints faced by healthcare providers frequently lead to discharge summaries that lack sufficient detail and depth, potentially compromising the quality of the medical record [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eIn fast-paced clinical environments, key details or changes in a patient\u0026rsquo;s condition maybe inadvertently omitted, undermining the completeness of the medical documentation. Additionally, the absence of standardized formats and guidelines often results in inconsistencies in how medical information is organized, making it challenging for readers to quickly extract critical insights [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. The creation of discharge summaries requires experienced clinicians to manually integrate diverse data sources from the Hospital Information System (HIS), including test results, medical orders, pathology reports, and other relevant information, which further complicates the process.\u003c/p\u003e\u003cp\u003eRecent advancements in Artificial Intelligence (AI) have rapidly transitioned from theoretical research to practical applications in daily life. Large language models (LLMs), crucial in natural language processing, have shown particular promise in the medical field [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Their capacity for human-like conversational interactions [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] has led to successful applications in paper reviewing [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], lab result interpretation [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], and patient data anonymization [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. However, the implementation of LLMs in healthcare faces significant challenges, primarily due to the disconnect between their training on general linguistic data and the specialized terminology of medicine.\u003c/p\u003e\u003cp\u003eWhile fine-tuning methods like Low Rank Adaptation (LoRA) [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] have been explored to bridge this gap, they have shown limited effectiveness in adapting to the intricacies of medical texts. Our study addresses these limitations by employing the DoRA (Decomposed Low Rank Adaptation) [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] fine-tuning method, which better aligns the model with medical- specific terminology and improves fine-tuning efficiency. Furthermore, we introduce a novel self-evaluation mechanism, a pioneering approach in this domain that enables the model to iteratively assess and enhance its output, significantly improving the quality and consistency of generated discharge summaries.\u003c/p\u003e\u003cp\u003eThis research presents an innovative end-to-end workflow for automating the creation of discharge notes from unstructured breast and thyroid surgery Hospital Information System (HIS) data. By leveraging fine-tuned LLMs and our unique self-evaluation mechanism, we have developed a system that not only reduces the time required by clinical workers but also surpasses existing methods in accuracy and efficiency. Our approach has received endorsement from healthcare professionals for its clinical relevance and utility, underscoring its potential to enhance patient care and improve the precision of medical records.\u003c/p\u003e\u003cp\u003eThis advancement represents a significant leap forward in integrating AI-driven tools within patient care settings, offering a transformative solution to refine healthcare documentation practices. By addressing the challenges of applying LLMs to specialized medical contexts and introducing a self-improving mechanism, our work paves the way for more accurate, efficient, and context-aware AI applications in healthcare.\u003c/p\u003e"},{"header":"II. Methodology","content":"\u003cp\u003eII.A. Dataset\u003c/p\u003e\u003cp\u003eThis study utilizes a dataset derived from the Thyroid Surgery Department at a major tertiary hospital in China. The dataset comprises information from 6,214 patients, retrospectively collected from January 1, 2018, to December 31, 2022. The data were accessed and extracted from the hospital\u0026rsquo;s information system between March 15, 2023, and May 20, 2023 for research purposes. These data are characterized by their fully unstructured and complex nature, presenting significant challenges for analysis and processing. Data collection and processing were conducted under strict adherence to privacy protocols within the hospital\u0026rsquo;s intranet infrastructure. This approach ensures the preservation of patient confidentiality, with robust measures in place to prevent any potential leakage of sensitive information. The use of this dataset was approved by the Medical Research Ethics Committee of Renmin Hospital of Wuhan University (Ethics Approval Number: WDRY2024-K038), under the research protocol titled 'Integrated Multi-Omics and Artificial Intelligence Analysis in Thyroid Nodule Patients: Prediction, Subtype Identification, and Exploration of Recurrence Factors'. The Clinical Research Ethics Committee of Renmin Hospital of Wuhan University approved the conduct of the clinical study in accordance with the research protocol and the waiver of informed consent application.\u003c/p\u003e\u003cp\u003eAs illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, the process of transforming unstructured data from the HIS into structured discharge summaries involves several critical steps. These steps include integration, deduplication, privacy filtering, and normalization. Data within the HIS are distributed across multiple tables, such as the Doctor\u0026rsquo;s Advice Information Table, Clinical Lab Information Table, Nursing Information Table, Physical Signs Information Table, Blood Glucose Information Table, Inspection Information Table, Diagnosis Table, Pathology Information Table, and Progress Notes Table. Given the interrelated nature of these data sources and the presence of duplicative entries, deduplication is a necessary step to ensure data consistency and integrity.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eWe employ a MinHash algorithm for efficient deduplication, effectively identifying and removing redundant data. Privacy concerns are addressed through the use of regular expression matching techniques, which filter out sensitive information from the data sets. Subsequent steps involve preparing the textual data for analysis, which includes spelling correction, acronym expansion, and the standardization of medical terminology. These pre-processing tasks are essential for enhancing the quality and usability of the data, ultimately facilitating the summarization task performed by LLMs.\u003c/p\u003e\n\u003ch3\u003eII.B. Model\u003c/h3\u003e\n\u003cp\u003eRecent advancements in language models have shown significant potential for medical applications. Mistral [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], as discussed in [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], has been particularly noted for its effectiveness in generating discharge summaries [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Similarly, Llama 3 has demonstrated robust performance across a variety of LLM benchmarks [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Due to our focus on Chinese medical data and the limited availability of Chinese corpus in previously used models, we also consider incorporating Qwen2 [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] to better address linguistic and regional data specifics.\u003c/p\u003e\n\u003ch3\u003eII.C. Parameter Efficient Fine Tuning (PEFT)\u003c/h3\u003e\n\u003cp\u003eThe generation of discharge summaries typically adheres to a strict output format, necessitating fine-tuning of the LLM to produce standardized summaries. However, strict data protection regulations within hospital environments often restrict the ability to conduct extensive model training on external systems. Consequently, hospitals typically lack the necessary computational resources, such as GPUs, to perform full-parameter fine-tuning of large language models. This limitation underscores the importance of Parameter Efficient Fine Tuning (PEFT) in these settings. Previous studies have extensively discussed methods such as LoRA and QLoRA, which utilize the low-rank factorization of weight matrices:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eW0\u0026thinsp;=\u0026thinsp;W0\u0026thinsp;+\u0026thinsp;B *AT\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eHowever, these LoRA-based methods possess inherent drawbacks. For example, they tend to increase the general magnitude of the weight matrix components differently compared to traditional full-parameter fine-tuning [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. This discrepancy can lead to suboptimal fine-tuning outcomes, particularly in fields like medicine, where precise terminology is crucial.\u003c/p\u003e\u003cp\u003eTo address these issues, we employ the Weight-Decomposed Low-Rank Adaptation (DoRA) method, which decomposes weights into direction and magnitude components. This approach more closely approximates full-parameter fine-tuning, enhancing the model\u0026rsquo;s ability to accurately handle specialized medical terminology and context.\u003c/p\u003e\n\u003ch3\u003eII.D. Self-Evaluation Mechanism\u003c/h3\u003e\n\u003cp\u003eIn cognitive psychology, rapid, intuitive judgments are often likened to Daniel Kahneman\u0026rsquo;s 140 \u0026lsquo;System 1 thinking,\u0026rsquo; which operates subconsciously. In contrast, tasks that require deliberate and methodical thought, such as mathematical calculations and logical reasoning, are categorized as \u0026lsquo;System 2 thinking.\u0026rsquo; Similarly, the output of Large Language Models (LLMs), especially in tasks like generating discharge summaries, can be initially rapid and intuitive, mirroring \u0026lsquo;System 1 thinking.\u0026rsquo; However, ensuring the reliability and completeness of these outputs often necessitates a more deliberate review process, akin to \u0026lsquo;System 2 thinking.\u0026rsquo;\u003c/p\u003e\u003cp\u003eDespite this, our research demonstrates that LLMs are capable of self-evaluating their outputs to identify and correct omissions. This capability is facilitated by our novel self-evaluation mechanism, the general procedure of which is depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Initially, the LLM processes the diverse information extracted from HIS data to generate a preliminary discharge summary. Concurrently, information segmentation tools decompose the original input into distinct entities and events.\u003c/p\u003e\u003cp\u003eUpon generating the initial summary, this output, along with segments of the original data, is re-fed into the LLM. This process allows the model to determine whether it can locate the original data\u0026rsquo;s details within the discharge summary. If details are found to be missing, the model identifies this as an omission and adjusts the summary accordingly. This iterative process continues for each data segment. Finally, the most recent summary and the complete original input are reintegrated into the LLM to assess if the enhanced summary satisfactorily represents the final, accurate discharge report.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eAlgorithm 1\u003c/strong\u003e\u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eSelf-Evaluation Mechanism for Discharge Summary Generation\u003c/span\u003e\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eRequire\u003c/strong\u003e\u003cp\u003eExtracted HIS notes X, Target Model M, Prompt for discharge generation P, Evaluation Prompt Ep\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e\u003ccolgroup cols=\"2\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eR \u0026larr; M(P ∥ X)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003e▷ Generate initial result\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eisComplete \u0026larr; False\u003c/p\u003e\u003cp\u003e\u003cb\u003ewhile\u003c/b\u003e not isComplete \u003cb\u003edo\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eE \u0026larr; Information__Decompose(X)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e▷ Decompose input\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003efor\u003c/b\u003e each ei in E \u003cb\u003edo\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ecomplete \u0026larr; M(Ep ∥ ei ∥ R)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e▷ Evaluate\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eif\u003c/b\u003e not complete \u003cb\u003ethen\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eR \u0026larr; M(R ∥ ei )\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e▷ Update result\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eend if end for\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ecomplete \u0026larr; M(Ep ∥ X ∥ R)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e▷ Final evaluation\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eif\u003c/b\u003e complete \u003cb\u003ethen\u003c/b\u003e\u003c/p\u003e\u003cp\u003eisComplete \u0026larr; True\u003c/p\u003e\u003cp\u003e\u003cb\u003eend if end while\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003ereturn\u003c/b\u003e R\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e▷ Final discharge summary\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\n\u003ch3\u003eII.E. Experiment\u003c/h3\u003e\n\u003cp\u003eOur experimental setup involved the deployment of several pre-trained models to evaluate the efficacy of different fine-tuning approaches. We used the Qwen2-7B model with LoRA fine-tuning and direct few-shot prompting as our baseline for comparison. Additionally, we explored the performance of other pre-trained models, specifically Mistral 7B and Llama3 8B, to understand their behavior under similar experimental conditions.\u003c/p\u003e\u003cp\u003eTo provide a comprehensive comparison, we employed DoRA fine-tuning on the same dataset used with the baseline model. The specifics of our experimental configuration, including the details of fine-tuning parameters, are documented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\u003cp\u003eFor evaluation, we utilized several methods to assess the effectiveness of the fine-tuned models. These included:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u0026bull; Direct few-shot prompting, which tests the model\u0026rsquo;s ability to generate summaries based on a limited number of example inputs.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u0026bull; Chain of Thought (CoT) prompting [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], designed to simulate a step-by-step reasoning process in generating the output. The detailed structure follows the [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e].\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u0026bull; Our novel approach termed \u0026lsquo;Self-Evaluation Generation\u0026rsquo;, which incorporates an iterative feedback mechanism to refine the model\u0026rsquo;s output continuously.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThese methods were chosen to explore different aspects of model performance, particularly focusing on their capability to generate accurate and comprehensive discharge summaries under varied prompting conditions\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eTraining parameters\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"2\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eLoRA Parameters\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLoRA__alpha\u003c/p\u003e\u003cp\u003eLoRA__dropout\u003c/p\u003e\u003cp\u003eLoRA rank\u003c/p\u003e\u003cp\u003ebias\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e16\u003c/p\u003e\u003cp\u003e0.1\u003c/p\u003e\u003cp\u003e64\u003c/p\u003e\u003cp\u003eNone\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003e\u003cb\u003eDoRA Parameters\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDoRA__alpha\u003c/p\u003e\u003cp\u003eDoRA__dropout\u003c/p\u003e\u003cp\u003eDoRA rank\u003c/p\u003e\u003cp\u003ebias\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e16\u003c/p\u003e\u003cp\u003e0.1\u003c/p\u003e\u003cp\u003e64\u003c/p\u003e\u003cp\u003eNone\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eOther Parameters\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eepoch\u003c/p\u003e\u003cp\u003ebatch size\u003c/p\u003e\u003cp\u003ewarmup ratio\u003c/p\u003e\u003cp\u003eweight decay\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e5\u003c/p\u003e\u003cp\u003e64\u003c/p\u003e\u003cp\u003e5%\u003c/p\u003e\u003cp\u003e0.05\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\n\u003ch3\u003eII.F. Evaluation Metrics\u003c/h3\u003e\n\u003cp\u003eTo rigorously assess the effectiveness of the fine-tuning techniques and the quality of the generated discharge summaries, we employed a suite of quantitative and qualitative metrics. These metrics were chosen to provide a comprehensive view of both linguistic accuracy and content relevance.\u003c/p\u003e\n\u003cdiv class=\"Heading\"\u003eII.F.1. \u003cb\u003eQuantitative Metrics\u003c/b\u003e\u003c/div\u003e\u003cp\u003eThe \u0026ldquo;self-evaluation\u0026rdquo; method consistently achieves the highest scores in several critical metrics:\u003c/p\u003e\u003cp\u003eWe utilized the following standardized metrics to evaluate the linguistic and semantic quality of the model outputs:\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eROUGE (Recall-Oriented Understudy for Gisting Evaluation)\u003c/strong\u003e\u003cp\u003eThis metric measures the overlap of n-grams between the generated text and the reference texts. It is particularly useful for assessing the extent to which important information is captured by the generated summaries. The ROUGE score is given by\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"625\" height=\"82\"\u003e\u003c/p\u003e\u003cp\u003e\u0026bull;\u003cb\u003eBLEU (Bilingual Evaluation Understudy)\u003c/b\u003e: This metric evaluates the correspondence of machine-generated text to human-written text, focusing on precision of n-grams. It is computed as:\u003c/p\u003e\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"387\" height=\"91\"\u003e\u003c/p\u003e\u003cp\u003ewhere BP is the brevity penalty, are the weights assigned to each n-gram, and are the n-gram precisions.\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u0026bull;\u003cb\u003eBERT Embedding Similarities\u003c/b\u003e: We measure the cosine similarity between the embeddings of the generated text and the reference text, derived from a pre-trained BERT model. This metric assesses the semantic similarity beyond surface lexical matching.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u0026bull;\u003cb\u003ePerplexity\u003c/b\u003e: Typically used to measure the likelihood of the sequence of words in the generated text, given by:\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003e\u003cimg src=\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAQAAAAAxCAYAAAA1I3faAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAAFiUAABYlAUlSJPAAABiWSURBVHhe7Zx5VBRXFod/DQ0qYFhcQAICjSMqGjXuGsQAEYNJDKAmggg4xJiQqKNGjctMTDRgNHOG4LhH2zC4dILKaETUoCiLEUWUcWULsjUKKFs3S3ff+SN0TVc1S7MYM6G+c+oc+t53361b9erWrVevEBARgYeHp1uixxXw8PB0H/gEwMPTjeETAA9PN4ZPADw83Rg+AfDwdGP4BMDD043hEwAPTzeGTwA8PN0YPgHw8HRj+ATA0y40F46q/9Z1Mamu7ZqjI7bttdG1va7tngVc39zf7UXALwXm6UqICAKBgCtulY7YdCXPw39X++xof3wFwNOltDUIm7vftGXTEg0NDUALfQJAY2MjV9QsHfXfGXTxqVKpuKIW0aW/5uArAJ5OsWfPHggEAtTV1bEuOD09PRgZGUEgEOC9995j5B29U2lSUFCA9PR0jBkzBjY2Ni32eePGDRQXF2PatGkwNjbmqpGamoqMjAxMmDABL7/8Mlf9XPnpp58AAO7u7lwVQ2NjI+Li4mBlZYXx48dz1TrBJwCeTmFjY4OioiJYWFjA3NycuRvX1dWhuLgYaOUOzaWlC1mTqqoq7N27Fzk5OVi+fDkGDRrEbcIQFxeHEydOwM3NDb6+vhAKhSx9REQEli1bhtWrVyM8PJyl46LLvnUVKSkpiI6OxtChQ/HRRx9x1Qx1dXXYtGkT5HI5QkNDIRKJuE3ahnh4OoFEIiF9fX3av38/S15SUkKWlpbU1UNs//79NGvWLK6YhUqlYv4uLy+nUaNG0c2bN1lt1Lz77rv02WefccXPjYqKCpowYQKlpqZyVS3yz3/+kwICAqixsZGrahN+DoCnU3h7e8Pd3R1r167FgwcPGLmlpSUCAwNZbTXRtSrQpLa2FmvXrsX27du5Khaad2pzc3OsWrUKa9asYbVRo1QquaJO0ZG4NNm5cyfGjh2LcePGcVUtEhISgpycHFy8eJGrahM+AfB0CqFQiJ07d0Imk2HdunWMXCAQYPny5UhLSwMRaV0Y3HKaiDBv3jwsXLgQJSUlWL9+PZycnFil+enTpyESiWBjY8OyvXDhAsaPH4+wsDAAQHR0NCZNmoSMjAwIBAJ4enoiLS0NBQUFLLvmkMvl8Pb2hqOjI5ycnLBlyxZGl5ycjHHjxsHR0RHDhw/H6NGjMWfOHNTW1jJtuHFpohmjVCrVilGlUiE6OhqzZs2Cvr4+Y9dcfJMnT0ZGRgYAwNDQENOnT0d0dDRjoyv/twlAIpEgICAAixYtwokTJ7hqnt8QkUiETz/9FDExMdi3bx8jt7S0xNixYyEQCFq9MABg7969GDRoENLT0xEVFQWRSITp06cjNjYWVVVVAIDz589j0qRJLDu5XI7Lly9j4MCBSEhIQE5ODrKzsxEcHAxLS0sAgIWFBaytrZGWlsay5dLY2IiwsDAoFAqEhobC3d0dYrEYmZmZqKqqQmRkJOzs7BAaGgozMzOUlJTgjTfegIGBAberZtmzZ0+rMRYWFiI3N5c1oaeOz9bWlhVfUFAQEx8AuLq6Ij4+nvmtKx2eBKypqcHRo0ehVCqhUqkgl8tBRNDT00O/fv3wyiuvwM7OjmvWZUgkEqxZswZ5eXnYvHkz1q5dy23SpeTn5yM+Ph7u7u5wdHTkqv9vaGhowOXLl3Hv3j3U19fDwMAAIpEIHh4e6NGjB9BUakdHR6OmpoZrzjBp0iTWxfjkyRO4ubnh6dOnOHv2LP70pz+x2reFWCyGs7MzFixYAD8/P4SGhuLx48fw9PREXFwchg4dCldXV8ybNw+LFy9m7ORyOeLi4tC/f38sWrQIQUFBcHZ2xsyZM1n9z549G6NGjcL69etZ8rlz58LZ2Rl/+9vfkJ+fj+nTpyM8PBze3t6oqKiAt7c3XnvtNfj6+uL111/HkSNHMHHiREgkEmzbtg1JSUkwNDREYWEhLly4gLKyMtjY2ODtt9/WSgzcGD/66CM8evSIibG0tBQ+Pj6oqKhgbNTx9evXD++//36L8eXn58Pe3h4ymQy9evVi6VqFOymgK1KplF566SUCQADIzs6ORCIRWVlZkbm5Oa1atYqePHnCNetSvvrqKzIwMKD4+HiuqsvZvXs3AaCjR49yVVpoTkL9nqioqKDo6GiaPn06OTs7k0gkIhMTExo4cCDFxMQwk0ilpaU0ceJEEolEWpujoyOJRCKKiIjgdk+xsbFkbGxM77//PlelE0lJSTRo0CB68OABERElJyeTlZUVpaWlERHR0KFD6dChQxyrX6moqCA7OzsKDg4mhUJBxDkPCxcupCVLlmhY/MqcOXOYScCkpCRycnKisrIyRh8WFkazZs2itLQ0EolElJeXR0REiYmJNHjwYCouLqb6+noKDw+nqVOnkp2dHdna2tKNGzeYPjThxpiUlMTEePToURo8eDDXhKiF+Lh6AJSfn89VtUqHHwEsLS0hFovx4osvYurUqcjOzkZOTg5SU1MxcuRIfPXVVzh9+jTXTCd0LUpqampgZWWFkSNHclWdhrsPZWVlsLa21rr7KxQK1NbWstq3Ve52Fdx9bIv4+Hj4+/ujX79+yMjIQE5ODvbu3YvCwkIsWrSIeW3Xv39/pKamIicnR2tTn+clS5Zwu8dbb72FGTNmsB4D2kNBQQHGjRvHVA+//PILDAwMmGf+uro6rVd5mlhbW8PLywv6+vpar+2EQiHkcjmrPRehUAiBQID6+npGVl9fD0NDQ2a9wcmTJ1FdXY0zZ87A2toavXv3RllZGU6dOoUDBw4gOzsb9vb2iImJYfWt5uHDh6wY8/PzmRhbik8di2Z8XNR2dXV1XFWrdDgBoGnw19TUwMXFhdkBOzs7ZuFHTk4Ox+J/tDZ4db2AsrKyYG1t3a6SpzW/rbFy5Urcv38fo0aNYsnPnDmD0aNHQyqVsuS/BboeJ03MzMywceNG5nzNnj0bs2bNQkVFBa5du8ZtrhOax1SpVCIiIoKl15Xc3FwMHjyY+R0XF4fRo0czz7qWlpYoLy/XsPif70uXLuHJkyfMmwjusamoqMCAAQOY382NA2qarNS8CIVCIZRKJSwtLbF69WosWbIEFhYWOH78OLZt2wZjY2MMGDAA8fHxsLe3h1AohJubm5Z/Nbm5uazHI80YraystOJDUyyJiYms+NSo41A/NlhZWbH0bdGpBFBaWoqqqirWYgyBQIDevXsDbWSjlg6QrpSXlyM9PR19+vSBiYkJV90iLfnlDghuO0NDQ5iYmLCyb0VFBTZv3ozevXuzBldH4PpvDl3atIanpycSExPh4ODAyIRCIXr27AkiavV8tYb6WInFYujp6cHf35/R6brPcrkciYmJMDQ0BJqqldjYWAQHBzP929jYoKSkhGUnEAiQm5uLxMREzJs3D0lJScjMzMT+/ftZ7YqLi2Fra8uyQ9N5Vd9ARCIRhEIhDh06BDSNsaioKEybNg1lZWW4ceMG0tPTkZmZifPnz2PMmDEQNE1wGhkZQU9PDzU1Nbh48aLWZCUAyGQyXL58mZlrOXv2LGJjYxEUFASBQAB7e3uUlpayKhA0JY1Lly7h3XffRVJSEm7duqUVX2lpKUxMTNp1LQC/nqAOs337djI2Nqb09HSWfP369QSAjhw5wpITEf3rX/+iwMBAmj17Nn3wwQd06dIlRnfixAkKCgqijz/+mB4+fEhERP/4xz9o/vz5tGLFCqqoqGDa3r17l8zMzCg8PJyRadKaHyKib7/9lt555x3auHEjyWQyRr5hwwaaP38+84yWkpJC8+fPp/fee4+OHz9ORERyuZw+/fRTmj59Ounr65OJiQm99dZb5Ovry8wRpKam0oIFC+jDDz9knmGpaWHK4sWLKSQkhD7//HPG9/OcN5gxYwYZGhqy9rO93L9/n6ZMmUKJiYlcFcOJEydo8eLFVFhYSBkZGRQSEkL3798nIqKysjIaM2YM+fj4kK+vL3l4eNDq1atZ80i7du2iadOmafRItG3bNgoKCqJz587R9evXycLCgvz9/eny5ctMm9raWurbty/dvXuXZXv06FEyMzMjBwcHSklJIYVCQdu2baMxY8bQm2++SV5eXuTr60v5+flUWVlJLi4u5OfnR0FBQbRgwQLasGEDFRcXs/rcuXMnhYeHU3l5OUtOTTG+/PLLLcZYW1tLAwcOpKtXrxI1jQnN+K5du0Z9+vTRio+I6JtvvqGZM2eyZLrQ4QpAoVAgPT0dI0aMwOjRoxl5QkICjhw5gvHjx2PKlCmMvLq6Gjt27MCePXvw+PFj1NfXQyKRYMeOHZDJZEhISEB6ejpKSkoQGRmJmzdv4uTJk3j06BESEhKwZ88e5uMPALh9+zYUCoXWGu62/ADA9evXcf36dRQWFuLrr7/GlStXAABHjx5t9l1xcnIy9u7dizt37rDkT58+hVKpxKhRo5g7FwCUlJTg9OnTqK6uhlgsxqFDh1gLTqRSKfbt29fmwo3U1FT8/e9/13lLTU3ldtEmaWlpyMjIwMyZMzFkyBCuWidqa2sRERGBV199FVOnTmXkdXV12Lt3L9B0N71+/TpiYmKQmJiI06dPs6qsp0+folevXswCmDFjxmDlypUwMzNj2nh6eiIvLw9FRUWMDABsbW3h4eEBJycnBAcHw9HRkTX2zp07Bzs7Ozg5ObHsAGDWrFlMW319fcyfPx+enp7o0aMH+vXrh6VLl2LgwIGQSqXo3bs3GhsbUV1dDZlMhri4OBw7dozp6/z586isrERwcDAsLCw0vPxKZWVlqzEaGRnBx8eHqUDUx0cd35AhQxAUFKQVn0qlQmxsLKvy0hluRtCVp0+f0siRI8nDw4NSU1MpOTmZfvjhB3J1daUXX3yRoqKimLYKhYIOHjxI+vr65O/vz2RHb29vMjU1pYKCAvrmm2/o5MmTdPfuXXrhhRfok08+oRUrVlBxcTFFRUVRSEgI6069YsUKMjU1Zc22tubHzMyMCgoKiIho37599OOPP1Jubi717duXli5dSrdv36Y33niDcnJymP7UbN26tdm3DcePHycDAwOt6uLGjRu0ZcsWqq+vp9dff52GDRtGpaWlLL2pqSnt3r2bZcclIiJCaxa+ta25mfnWqKysJD8/P7K1taXk5GSumqG56kRTduzYMZoyZQpLT0R05coVUg+xrKwsio2NpdDQUAoODqYvv/yS1fbMmTPk5eXFkjVHUFAQbdu2jaqqqriqZikuLiY/Pz8Si8VcVbvYs2cPBQQEsGQBAQHk4+NDREQ5OTm0evVq5i2BeqxpEh8f3+ZdOj09nVxdXenOnTvNHncuDQ0N9NNPP9GUKVM69Natwwng4cOHNGDAADIwMCChUEhCoZAMDAzI3d2dTp06xWqbl5dHffv2JTc3N2poaCAiorq6OvL29iYHBwdWGZWcnEy9e/eml156iR4/fqzRy/9QqVTk4+NDAwcOZNn+8ssvLfoRiURa5VpNTQ298sor5ObmRgsWLKB///vfLD01+dqwYQPZ2tqSVCplydetW6cl57Jp0yYyMTGhpKQkRnb48GGyt7dvMb6uprmBpFQq6cCBAyQUCrXOV3uQy+U0ZMgQOnv2LFVVVbG2RYsWMQlATUxMDE2YMIHkcjlLvm7dOvL19SWlUsmSc6mqqiIXFxc6deoUKRSKZmNTU1tbS19++SX9+c9/Zslbs2mJL774giZOnEhFRUVUVVVFRUVFNHnyZAoLCyOFQkGrVq2iqKgoqqmpoXv37tEHH3zA7ULnGL/99lvy8/PTGlfN7fft27fp1VdfZR4b2ku7E4B6J44dO0YCgYDEYjHrpNfW1mp9lBAWFkampqZ05coVIiLKzs6mDz/8kACQWCxmBbZ582YCQAcPHmw2YNK4cN9++21WG66fnJwcCg0NJQB04MABrf5UKhV5e3uTvr4+BQYGau03NfmaOnUqjR8/niorKxl5dXU1ubi40Lhx41hyLocPHyYAFBMTw8i++OILWrNmTZsD4Vly8+ZNsrW1pf379zcbt64sXLiQAJBAIGBuBOoNTWtENPnhhx9o0qRJVFtby5J7e3vT4MGD6c6dOyx5czx+/JgiIiLo559/5qpYiMViOn78OMlkMta5544DXcjLyyNXV1fS19cnoVBI5ubmFBoaSmVlZXT16lUyNjYmAGRgYEB6enpaY5PaEWNDQwOlpqbSpk2buCoWcrmcVq9eTffu3Wt2bYAutDsBqNm6dSv17NlTazKCi0wmI09PTwJA1tbWJBKJyMXFhWbPnk3R0dFaF09gYCCZm5tTVlYWS67JrVu3qH///vTJJ58wsub8TJ06tUU/aqKiosjIyIhOnz7NVZFKpaKysjJycHCgZcuWsS5YtXzp0qWtXsiJiYlkZGREW7duJWoqu728vFr8Ou23IC8vj5ycnGjTpk06l9ItkZWVRVevXmUeAzW3n3/+mTWxePHiRVq5ciUNGTKEkpOTSSKRsPrJzMzUSgwtIZVKqbq6mitmUVpa2uJ5by8KhYIePHhAKSkpTGyFhYVETTeDtLQ05hikpKRQdnY2t4t2xVhXV9fsY4QmSqWS8vPzWx1/bdGhpcD19fUICAjArVu3cO7cOdbrFTS9+lFPYDx69AiTJ0+GmZkZAgICoFQq0a9fP9jb28PFxYVlV1VVBVdXV5iYmOD8+fPM6xIuZ86cgZeXF3bs2MEsC1X7MTU1xYIFC1r1o6ahoQGBgYGQSCS4cOECM4Gluf9JSUl47bXXEBkZiZCQEMY2OTkZHh4eWnIuDx48wKRJkzB79mzs3r0bEokEGRkZ2Lhxo9ZSUS4SiQQSiYQrbpG5c+di7ty5XDELhUIBf39/9OjRA+Hh4bC2tuY2eSZ8/fXXKCwsxJw5c/CXv/wFxsbG+Oyzz1iThjy/PR16C1BfX4/bt2/D3t6+zYUHDQ0NqK6uhrOzM5YuXYrly5cjICAALi4uSEpKYq03l0qlKCgowPDhw1u8+NG0mqpXr14YMWIEI2uPHzU//vgjiAgWFhbMmwBw1gBkZmaisbFR6wu0mzdvQqFQtPm9g5WVFezs7FBQUIDHjx8jISEBfn5+bV78zwKVSoXDhw+jpKQEYWFhrIs/PT29wwuBWoJ7bxk+fDgmT56MRYsWYejQoRg7dixLz/Mc4JYEupCZmUl9+vShefPmcVVaPHr0iBwdHWnGjBnMs2ZDQwOlpKSQv78/a6Lj0qVLZGRkRFu2bNHoQZvly5eTra0tVVRUkFQqpaKionb5ISIqKiqigIAAunXrFk2bNo08PDwYueZa8OXLl5OFhQXdvn2b8UVNEzrGxsbMfENxcXGLJVtwcDCNHDmSdu3aRevWrSPSeA7lPid2ltb6KywsJJFIRB9//DHdvHmT0tLSmM3b2/u5/WOM1vaZ59nS7gpAoVDgu+++Q3l5OeRyeZvrq9VfBmZmZuL69euorq5GRkYGQkJCMHz4cNYnjQ8fPoRMJmvzS7KioiIolUqUlZXh4MGDkEgkOvupr69HdXU1du7ciWHDhmHEiBGwsbHBf/7zHzx48AA7d+7EhQsXWL569uwJQ0NDiMVipiQ3MjKCTCbDvXv3UFxcjO3bt2utzlLj6OiI7OxsfPfdd1i2bBnQVGVoPmp0Fa31d+HCBeTm5iIyMhIjR47EuHHjmO348eNwdnbmmjwzNKuD1vaZ5xnDzQht8f333zOzuwDonXfeYem52VylUtGjR4/ozTffZGaKnZ2dSSwWU319Pautr68vmZubt/lFU1RUFBkaGpKpqSnzvp2aZoc1/QwfPlzLz+eff07m5uYUHh7OyBMTE8nQ0JBZWag5K672ZWZmxvKVl5dHDg4OBIDMzc0pMDCQ9VpP8zgcOXKEwHnjwT1OvwU+Pj6sc8fdDh8+zDXh+YPT7knA8vJyZGdnQ19fHyqVCmZmZqwPOFoiLy8PUqkURARjY2M4ODjghRdeYLXJysqCXC7HsGHDmv0qSk1lZSVyc3PR2NgIOzs7VhXRlp+CggIUFxfD3t6esaurq8Pdu3eb7a8lX0qlEvfv30dVVRUEAgEsLS1hb2/P2Gly8uRJREZG4vvvv4epqSlX3SXoUk1kZWWhsrKSK2ZwdHSEubk5V8zzB6bdCUBXdBmQvxXPc1+kUinWrl2L+fPnw83NjaV71vv1rPtvi5b8tyTn+e1p9xxAV/GM8k6n+u2MrSYJCQkoKSlBfn4+Dh48CAcHB62L//dCV8XM8/9JuyuAZ529O9N/Z2y7CqVSCU9PT7i6uqJPnz5IS0vDgQMHuM105vcQ07OmO8T4e6XdFcCzPlGd6b8ztl2FTCZDbW0t/vrXv+LatWvYtWsXt0m7+D3E9KzpDjH+Xml3BcDDw/PHod0VAA8Pzx8HPgHw8HRj+ATAw9ON4RMAD083hk8APDzdGD4B8PB0Y/gEwMPTjeETAA9PN+a/XRxt/heAuRAAAAAASUVORK5CYII=\" width=\"256\" height=\"49\"\u003e\u003c/p\u003e\u003cp\u003ewhere is the model\u0026rsquo;s probability distribution over the entire text.\u003c/p\u003e\n\u003ch3\u003eII.F.2. Qualitative Metrics\u003c/h3\u003e\n\u003cp\u003eIn addition to automated metrics, we also conducted human evaluations to assess the practical utility of the generated summaries. Trained evaluators rated the outputs based on the following five dimensions:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003e\u0026bull; Accuracy\u003c/b\u003e: The factual accuracy of the information presented.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003e\u0026bull; Completeness\u003c/b\u003e: The extent to which all relevant information is included.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003e\u0026bull; Relevance \u0026amp; Clarity (R\u0026amp;C)\u003c/b\u003e: How relevant and clear the information is with respect to the patient\u0026rsquo;s medical history and treatment.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003e\u0026bull;Consistency\u003c/b\u003e: The internal consistency of the information within the summary itself.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003e\u0026bull;Utility\u003c/b\u003e: The practical usefulness of the summary for continuing care and patient understanding.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThese metrics together enable a holistic evaluation of our model\u0026rsquo;s performance, encompassing both the technical efficacy of the fine-tuning methods and the practical applicability of the generated texts in real-world medical settings.\u003c/p\u003e"},{"header":"III. Results","content":"\u003cp\u003eIII.A. Analysis of Fine-Tuning Methods\u003c/p\u003e\u003cp\u003eThe results, as depicted in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, highlight significant differences in performance across various fine-tuning methods applied to different models. With the same LoRA rank and training epochs, the DoRA method generally outperforms both LoRA and QLoRA across most evaluation metrics. This trend is consistent across different models, underscoring the robustness of the DoRA method in optimizing model performance.\u003c/p\u003e\n\u003ch3\u003eIII.A.1. Model-Specific Performance\u003c/h3\u003e\n\u003cp\u003e\u003cstrong\u003eQwen2-7B Model\u003c/strong\u003e\u003cp\u003eThe Qwen2-7B model, when fine-tuned with DoRA, achieved competitive results in almost all metrics, with particularly noteworthy improvements in BERT score and Perplexity, indicating enhanced semantic understanding and lower model confusion, respectively. The LoRA fine-tuning method showed a slight advantage in Rouge1 scores, which may be attributed to the model\u0026rsquo;s intrinsic handling of Chinese corpus. However, DoRA consistently demonstrated superior overall performance, suggesting that its fine-tuning mechanism is particularly effective for comprehensive text analysis.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eMistral 7B Model\u003c/strong\u003e\u003cp\u003eFor the Mistral 7B model, DoRA fine-tuning resulted in substantial improvements over both LoRA and QLoRA methods. Although the gains were less pronounced compared to the Qwen2-7B model, DoRA still enhanced the BLEU score and BERTscore significantly, pointing to improved translation fidelity and semantic accuracy.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eLlama3 8B Model\u003c/strong\u003e\u003cp\u003eThe Llama3 8B model exhibited a different pattern, with LoRA performing relatively better than QLoRA but still underperforming compared to DoRA in terms of BERT score and Perplexity. This indicates that DoRA is more effective in reducing perplexity and enhancing semantic coherence in outputs, which is crucial for medical summary applications.\u003c/p\u003e\u003c/p\u003e\n\u003ch3\u003eIII.A.2. Comparative Evaluation\u003c/h3\u003e\n\u003cp\u003eThe detailed results in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e underline the efficacy of DoRA fine-tuning across multiple models and metrics. The DoRA method not only consistently yields lower perplexity scores, indicating a better fit to the data\u0026mdash;but also achieves higher scores in both Rouge2 and RougeL, which are critical for capturing finer nuances and ensuring the logical flow of generated texts. Furthermore, DoRA\u0026rsquo;s ability to outperform in terms of BLEU and BERT score across different models reinforces its applicability in enhancing the accuracy and relevance of generated medical texts. In summary, these results validate the hypothesis that DoRA fine-tuning significantly enhances the quality of discharge summaries generated by LLMs. The consistent outperformance of DoRA across diverse models and metrics emphasizes its potential in medical text generation, particularly in contexts demanding high accuracy and coherence.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eInfluence of different fine-tuning methods\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"7\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMethod\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRouge1\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRouge2\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eRougeL\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eBLEU\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eBERTscore\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003ePerplexity\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eQwen2-7B\u0026thinsp;+\u0026thinsp;LoRA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.461\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.343\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.372\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.14\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.851\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e1.321\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eQwen2-7B\u0026thinsp;+\u0026thinsp;QLoRA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.209\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.191\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.193\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.829\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e1.64\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eQwen2-7B\u0026thinsp;+\u0026thinsp;DoRA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e0.463\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e0.350\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e0.391\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e0.15\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.866\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e1.278\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMistral 7B\u0026thinsp;+\u0026thinsp;LoRA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.222\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.195\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.223\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.753\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e1.933\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMistral 7B\u0026thinsp;+\u0026thinsp;QLoRA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.210\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.174\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.221\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.722\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e2.434\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMistral 7B\u0026thinsp;+\u0026thinsp;DoRA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.242\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.221\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.241\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.831\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e1.792\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLlama3 8B\u0026thinsp;+\u0026thinsp;LoRA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.397\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.191\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.202\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.745\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e1.992\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLlama3 8B\u0026thinsp;+\u0026thinsp;QLoRA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.377\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.163\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.194\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.06\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.721\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e2.370\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLlama3 8B\u0026thinsp;+\u0026thinsp;DoRA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.239\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.186\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.224\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.822\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e1.839\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\n\u003ch3\u003eIII.B. Analysis of Generating Methods\u003c/h3\u003e\n\u003cp\u003eThe comparative analysis of different generating methods is captured in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. This analysis demonstrates the superior performance of the \u0026ldquo;self-evaluation\u0026rdquo; method over traditional few-shot and Chain of Thought (CoT) prompting methods across several metrics.\u003c/p\u003e\n\u003ch3\u003eIII.B.1. Quantitative Performance\u003c/h3\u003e\n\u003cp\u003eThe \u0026ldquo;self-evaluation\u0026rdquo; method consistently achieves the highest scores in several critical metrics:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003e\u0026bull; Rouge Scores\u003c/b\u003e: It surpasses other methods in Rouge1 and Rouge2 scores, indicative of better extraction and reproduction of key phrases and sentences from the source text.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003e\u0026bull; BLEU and BERTscore\u003c/b\u003e: It shows notable improvements in BLEU scores, reflecting better translation fidelity and syntactic coherence. The BERTscore, which assesses the semantic similarity between the generated texts and the references, is also highest for the \u0026ldquo;self-evaluation\u0026rdquo; method, underlining its effectiveness in maintaining semantic integrity.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\n\u003ch3\u003eIII.B.2. Qualitative Performance\u003c/h3\u003e\n\u003cp\u003eThe \u0026ldquo;self-evaluation\u0026rdquo; method consistently achieves the highest scores in several critical metrics:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003e\u0026bull; Rouge Scores\u003c/b\u003e: It surpasses other methods in Rouge1 and Rouge2 scores, indicative of better extraction and reproduction of key phrases and sentences from the source text.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003e\u0026bull; BLEU and BERTscore\u003c/b\u003e: It shows notable improvements in BLEU scores, reflecting better translation fidelity and syntactic coherence. The BERTscore, which assesses the semantic similarity between the generated texts and the references, is also highest for the \u0026ldquo;self-evaluation\u0026rdquo; method, underlining its effectiveness in maintaining semantic integrity.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eIn addition to automated metrics, the \u0026ldquo;self-evaluation\u0026rdquo; method demonstrates remarkable performance in human -evaluated dimensions:\u003c/p\u003e\u003cp\u003e\u003cb\u003e\u0026bull; Completeness and Accuracy\u003c/b\u003e: Achieving the highest ratings in these categories, the method proves its capability in generating comprehensive and accurate summaries.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparison of different generating methods\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"11\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c11\" colnum=\"11\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMethod\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRouge1\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRouge2\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eRougeL\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eBLEU\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eBERTscore\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eCompleteness\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eR \u0026amp; C\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c10\"\u003e\u003cp\u003eConsistency\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c11\"\u003e\u003cp\u003eUtility\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eQwen2-7B\u0026thinsp;+\u0026thinsp;few-shot\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.463\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.350\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.391\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.15\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.866\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e4.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e3.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e3.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e4.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e4.1\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMistral-7B\u0026thinsp;+\u0026thinsp;few-shot\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.242\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.221\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.241\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.831\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e3.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e3.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e3.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e4.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e4.2\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLlama3-8B\u0026thinsp;+\u0026thinsp;few-shot\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.239\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.186\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.224\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.822\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e3.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e3.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e4.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e3.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e4.3\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eQwen2-7B\u0026thinsp;+\u0026thinsp;CoT\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.487\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.373\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.446\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.20\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.897\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e4.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e4.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e4.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e4.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e4.3\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMistral-7B\u0026thinsp;+\u0026thinsp;CoT\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.399\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.343\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.450\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.19\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.837\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e4.3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e4.3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e4.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e\u003cb\u003e4.4\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e4.3\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLlama3-8B\u0026thinsp;+\u0026thinsp;CoT\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.343\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.321\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e0.462\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.18\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.854\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e4.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e4.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e4.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e4.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e\u003cb\u003e4.5\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eQwen2-7B\u0026thinsp;+\u0026thinsp;self-evaluation\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e0.512\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e0.398\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.451\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.24\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.923\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e4.7\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e4.9\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e4.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e4.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e\u003cb\u003e4.5\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMistral-7B\u0026thinsp;+\u0026thinsp;self-evaluation\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.424\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.350\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.451\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e0.25\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.889\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e4.6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e4.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e\u003cb\u003e4.6\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e4.3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e4.4\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLlama3-8B\u0026thinsp;+\u0026thinsp;self-evaluation\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.356\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.344\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e0.462\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.22\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.882\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e4.7\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e4.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e4.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e4.3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e\u003cb\u003e4.5\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003e\u0026bull; Utility\u003c/b\u003e: With top scores in utility, the method validates its practical applicability in real-world settings, providing outputs that are highly useful for medical professionals and patients alike.\u003c/p\u003e\u003cp\u003eThe self-evaluation method not only performs consistently across different LLMs (Qwen2-7B, Mistral-7B, Llama3-8B) but also shows robustness in handling diverse datasets and operational contexts. This is evidenced by its superior performance in both linguistic and content-specific evaluations, suggesting its adaptability and effectiveness across different medical textual environments.\u003c/p\u003e\n\u003ch3\u003eIII.B.3. Implications\u003c/h3\u003e\n\u003cp\u003eThe superior results of the self-evaluation method, as documented in the table, underscore its potential as a significant advancement in medical text generation.\u003c/p\u003e\u003cp\u003eIts ability to iteratively refine the output using an internal feedback mechanism allows for a level of precision and relevance that traditional methods struggle to achieve. This is particularly valuable in medical settings where accuracy and detail are paramount.\u003c/p\u003e\u003cp\u003eThe detailed results presented in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e validate the effectiveness of the self-evaluation method in generating high-quality medical texts, making it a promising approach for enhancing automated medical documentation systems.\u003c/p\u003e"},{"header":"IV. Discussion","content":"\u003cp\u003eIV.A.Summary of Key Findings\u003c/p\u003e\u003cp\u003eThis study yields several significant findings regarding the enhancement of medical text generation through advanced fine-tuning methods and iterative self-evaluation mechanisms:\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eThe DoRA method shows marked advantages over traditional fine-tuning approaches such as LoRA and QLoRA, particularly in addressing the specialized linguistic and structural demands of medical texts. DoRA achieves superior performance across multiple metrics (ROUGE, BLEU, BERTscore), indicating enhanced adaptation to the unique challenges posed by medical terminology and context. This method\u0026rsquo;s fine-tuning mechanism enables more efficient integration of domain-specific knowledge, resulting in more coherent, relevant, and semantically accurate discharge summaries.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eThe introduction of a self-evaluation mechanism significantly elevates the quality of generated medical texts. This iterative review and refinement process substantially reduces omissions and factual inconsistencies. The mechanism not only enhances the completeness and relevance of discharge summaries but also increases overall model robustness by ensuring better alignment between generated content and original input data. Substantial improvements in both automated and human-evaluated metrics, including accuracy, consistency, and utility, underscore the effectiveness of this approach.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eAmong the models tested, the Qwen2 7B model consistently outperforms others in this domain, primarily due to its superior proficiency in Chinese language processing. This finding highlights the importance of language-specific capabilities in medical text generation tasks.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e\n\u003ch3\u003eIV.B. Interpretation and Significance of Results\u003c/h3\u003e\n\u003cp\u003eAnalysis of training results, particularly perplexity metrics, reveals that the DoRA method, which incorporates magnitude of weight training, facilitates more comprehensive learning compared to LoRA methods. This leads to substantial improvements, especially for models like Llama and Mistral, which initially possessed less knowledge of Chinese content. The addition of weight magnitude training in DoRA appears to enhance the model\u0026rsquo;s ability to adapt to unfamiliar domains, suggesting a more robust fine-tuning approach for specialized applications.\u003c/p\u003e\u003cp\u003eThe effectiveness of our self-evaluation mechanism aligns with recent findings in the field. As noted by [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], self-evaluation capabilities in Large Language Models (LLMs) can serve as a defense mechanism against adversarial attacks, indicating LLMs\u0026rsquo; ability to discern true from false information. Furthermore, [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] has demonstrated the potential of LLMs as critics for their own outputs. Our research provides comprehensive empirical evidence supporting the feasibility and efficacy of LLM self-evaluation and self-improvement in the context of medical text generation.\u003c/p\u003e\u003cp\u003eThe combined application of DoRA fine-tuning and self-evaluation mechanisms represents a significant advancement in automating medical documentation. This approach not only improves the efficiency of discharge summary generation but also enhances the quality and reliability of the generated content. Such improvements have the potential to significantly reduce the administrative burden on healthcare professionals while maintaining high standards of accuracy and completeness in medical records.\u003c/p\u003e\n\u003ch3\u003eIV.C. Limitations of the Study\u003c/h3\u003e\n\u003cp\u003eWhile this study demonstrates significant improvements in the generation of medical texts, there are several limitations that should be acknowledged:\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eLimitations of the dataset\u003c/strong\u003e\u003cp\u003eThe dataset used in this study is specific to a single hospital and focuses primarily on discharge summaries from the Breast and Thyroid Surgery Department. As a result, the model\u0026rsquo;s performance may not generalize well to other medical departments or hospitals with different patient populations, medical practices, or clinical documentation formats. This specificity limits the scope of the findings and highlights the need for broader evaluations across diverse datasets and medical specialties to validate the robustness of the proposed approach.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003ePotential biases or errors in models and methods\u003c/strong\u003e\u003cp\u003eThe fine-tuning and self-evaluation mechanisms employed in this study are dependent on the quality and structure of the training data. Any inherent biases or errors present in the dataset, such as inconsistencies in documentation or variations in medical terminology, could lead to biased model outputs.\u003c/p\u003e\u003c/p\u003e\u003cp\u003eFurthermore, the iterative self-evaluation mechanism, while improving text quality, may introduce compounding errors if the model\u0026rsquo;s initial outputs contain inaccuracies that are reinforced during the refinement process. These limitations underline the importance of rigorous data preprocessing and error mitigation strategies.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eConsiderations of computational resources and time costs\u003c/strong\u003e\u003cp\u003eThe DoRA fine-tuning method and the self-evaluation mechanism, while efficient, still require considerable computational resources, particularly in a hospital environment where high-end hardware, such as GPUs, may not be readily available. The iterative nature of the self-evaluation process increases time costs, which may limit the real-time applicability of the approach in settings where rapid document generation is needed. Balancing the trade-off between improved accuracy and resource constraints remains a challenge, particularly for smaller healthcare institutions with limited computational infrastructure.\u003c/p\u003e\u003c/p\u003e\n\u003ch3\u003eIV.D. Future Research Directions\u003c/h3\u003e\n\u003cp\u003eBuilding upon the findings and limitations of this study, several future research directions could further enhance the field of medical text generation: Extension to other types of medical texts: Future research could explore the application of the DoRA fine-tuning method and self-evaluation mechanism to generate other types of medical documents, such as clinical progress notes, radiology reports, or surgical summaries. Expanding the scope to include various forms of medical documentation could help verify the generalizability of the proposed approach and uncover additional domain- specific challenges.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eExploration of multimodal inputs\u003c/strong\u003e\u003cp\u003eIncorporating multimodal data, such as medical images, audio recordings, or lab results, alongside textual inputs could provide a richer context for generating more accurate and comprehensive summaries. Research into how large language models can integrate and process multimodal information effectively would represent a significant advancement in automated medical documentation.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eFurther optimization of the self-evaluation mechanism\u003c/strong\u003e\u003cp\u003eThe current self-evaluation process could be optimized to reduce computational costs and time requirements. Future work could focus on developing more efficient evaluation strategies that maintain high levels of accuracy while improving the speed of iteration. Additionally, integrating external validation sources, such as medical guidelines or clinical databases, could help the model better assess its outputs.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eStudy of long-term impacts on medical practice\u003c/strong\u003e\u003cp\u003eLongitudinal studies could examine the practical implications of integrating AI-generated discharge summaries into real -world clinical workflows. Research could focus on how such systems affect clinician workload, patient care quality, and documentation accuracy over time, as well as the potential legal and ethical considerations of relying on AI-generated medical texts.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eDevelopment of adaptive fine-tuning methods\u003c/strong\u003e\u003cp\u003eInvestigating new fine-tuning techniques that allow models to adapt dynamically to evolving medical knowledge and practices could further improve the utility of LLMs in healthcare. Adaptive methods could help maintain model relevance in the face of new medical guidelines or treatments without requiring full retraining.\u003c/p\u003e\u003c/p\u003e"},{"header":"V. Conclusions","content":"\u003cp\u003eThis study demonstrates significant advancements in the generation of medical discharge summaries by integrating large language models with the Decomposed Low-Rank Adaptation fine-tuning method. Compared to traditional techniques like LoRA and QLoRA, DoRA excels in handling the complexities of medical terminology and context, achieving superior performance in terms of accuracy, coherence, and semantic consistency across various evaluation metrics.\u003c/p\u003e\u003cp\u003eMoreover, the introduction of the self-evaluation mechanism allows the model to iteratively assess and refine its outputs, minimizing omissions and improving the completeness and accuracy of the generated texts. This iterative process ensures that the final discharge summaries are more aligned with the original input data and meet higher standards of clinical relevance and utility.\u003c/p\u003e\u003cp\u003eThe findings of this study suggest that AI-driven approaches can significantly reduce the time required for generating high-quality medical documentation, while also enhancing the consistency and reliability of these documents. This research paves the way for future exploration in automating various types of healthcare documentation, improving both efficiency and quality in clinical workflows.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eConflict of Interest Statement\u003c/p\u003e\n\u003cp\u003eThe authors have no relevant conflicts of interest to disclose.\u003c/p\u003e\n\u003cp\u003eFunding\u0026nbsp;Declaration\u003c/p\u003e\n\u003cp\u003eThe authors did not receive support from any organization for the submitted work.\u003c/p\u003e\n\u003cp\u003eData Availability Declaration\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eDue to ethical restrictions, the raw data cannot be made publicly available. However, de-identified data may be obtained from the first author upon reasonable request.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eWenbin Li conceptualized the research idea and wrote the main manuscript text. Hui Feng and Minpeng Xu conducted the literature review and contributed to drafting and editing portions of the manuscript. Chao Hu and Longlong Cheng compiled and analyzed the latest developments in machine learning techniques, wrote the technical methods section, and assisted in final manuscript revisions. All authors reviewed and approved the final version of the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eWilson, S., Ruscoe, W., Chapman, M., Miller, R. \u0026amp; Chen General practitioner\u0026ndash;hospital communications: a review of discharge summaries, Journal of quality in clinical practice 21, 104\u0026ndash;108 W.-K. Linear Networks and Systems. Belmont, CA: Wadsworth, 1993, pp. 123\u0026ndash;135. (2001).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003evan Walraven, C. \u0026amp; Rokosh, E. What is necessary for high-quality discharge summaries? \u003cem\u003eAm. J. Med. Qual.\u003c/em\u003e \u003cb\u003e14\u003c/b\u003e, 160\u0026ndash;169 (1999).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNeira, R. A. Q., de Vries, G. J., Caffarel, J. \u0026amp; Stretton, E. \u003cem\u003eExtraction of data from a hospital information system to perform process mining, in MEDINFO 2017: Precision Healthcare through Informatics\u003c/em\u003e pages 554\u0026ndash;558 (IOS, 2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eThirunavukarasu, A. J. et al. Large language models in medicine. \u003cem\u003eNat. Med.\u003c/em\u003e \u003cb\u003e29\u003c/b\u003e, 1930\u0026ndash;1940 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBrown, T. B. Language models are few-shot learners, arXiv preprint arXiv:2005.14165(2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFloridi, L., Chiriatti, M., GPT-3. \u0026amp; Its nature, scope, limits, and consequences. \u003cem\u003eMind. Mach.\u003c/em\u003e \u003cb\u003e30\u003c/b\u003e, 681\u0026ndash;694 (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhou, R., Chen, L. \u0026amp; Yu, K. Is LLM a Reliable Reviewer? A Comprehensive Evaluation of LLM on Automatic Paper Reviewing Tasks, in Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation 430 (LREC-COLING 2024), pages 9340\u0026ndash;9351, (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHe, Z. et al. Quality of Answers of Generative Large Language Models Versus Peer Users for Interpreting Laboratory Test Results for Lay Patients: Evaluation Study. \u003cem\u003eJ. Med. Internet. Res.\u003c/em\u003e \u003cb\u003e26\u003c/b\u003e, e56655 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWiest, I. C. et al. Anonymizing medical documents with local, privacy preserving large language models: The LLM-Anonymizer, medRxiv, 2024\u0026ndash;06 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHu, E. J. et al. Lora: Low-rank adaptation of large language models, arXiv preprint arXiv:2106.09685(2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu, S. Y. et al. Dora: Weight-decomposed low-rank adaptation, arXiv preprint arXiv:2402.09353(2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJiang, A. Q. et al. Mistral 7B, arXiv preprint arXiv:2310.06825 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJung, H. et al. Enhancing Clinical Efficiency through LLM: Discharge Note Generation for Cardiac Patients, arXiv preprint arXiv:2404.05144 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDubey, A. et al. The llama 3 herd of models, arXiv preprint arXiv:2407.21783 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBai, J. et al. Qwen technical report, arXiv preprint arXiv:2309.16609 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYang, A. et al. Qwen2 technical report, arXiv preprint arXiv:2407.10671 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWei, J. et al. Chain-of-thought prompting elicits reasoning in large language models. \u003cem\u003eAdv. Neural. Inf. Process. Syst.\u003c/em\u003e \u003cb\u003e35\u003c/b\u003e, 24824\u0026ndash;24837 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTang, A. Q., Zhang, X. \u0026amp; Dinh, M. N. Ignition Innovators at Discharge Me! Chain-of-Thought Instruction Finetuning Large Language Models for Discharge Summaries, arXiv preprint arXiv:2407.17636 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBrown, H., Lin, L., Kawaguchi, K. \u0026amp; Shieh, M. Self-Evaluation as a Defense Against Adversarial Attacks on LLMs, arXiv preprint arXiv:2407.03234 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMcAleese, N. et al. Llm critics help catch llm bugs, arXiv preprint arXiv:2407.00215 (2024).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Natural Language Processing, Large Language Models, Discharge Summaries","lastPublishedDoi":"10.21203/rs.3.rs-7814494/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7814494/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eThe generation of discharge summaries in healthcare is crucial but labor-intensive.\u003c/p\u003e\u003ch2\u003ePurpose\u003c/h2\u003e\u003cp\u003eThis study introduces a novel framework that enhances the automation of this process using advanced natural language processing (NLP) techniques.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eWe combine the Decomposed Low Rank Adaptation (DoRA) fine-tuning method with a unique self-evaluation mechanism to improve large language models (LLMs) tailored for medical text generation. Our methodology leverages DoRA to efficiently adapt pre-trained LLMs to the specialized medical domain, demonstrating superior performance over traditional methods like LoRA and QLoRA across multiple metrics. We further incorporate a self-evaluation mechanism inspired by cognitive psychology principles, enabling iterative refinement of model outputs to enhance accuracy and completeness.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eThis approach is rigorously compared against popular few-shot prompting and Chain of Thought (CoT) methods. Extensive experiments yield significant improvements in both quantitative (ROUGE, BLEU, BERT score) and qualitative (accuracy, completeness, relevance, consistency, and utility) assessments compared to baseline techniques.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eOur results demonstrate substantial enhancements in the quality and consistency of generated discharge summaries while markedly reducing the time required for their creation. This research underscores the potential of AI-driven tools in healthcare documentation, significantly reducing the time required for generating discharge summaries while improving their quality and consistency. The findings indicate promising prospects for automating medical documentation that adheres to high standards of accuracy and relevance.\u003c/p\u003e","manuscriptTitle":"Optimizing discharge summary generation: fine-tuning LLMs by DoRA and iterative self-evaluation for enhanced medical text generation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-04 12:20:48","doi":"10.21203/rs.3.rs-7814494/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-11-06T14:50:27+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-10-31T22:43:33+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-10-24T09:24:19+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"84058213127929099510472603588733082176","date":"2025-10-24T08:39:15+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-10-23T21:29:14+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"252546221748709460109205418468574935721","date":"2025-10-23T19:33:38+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"304335326773156208965532268149708862193","date":"2025-10-23T16:41:13+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-10-23T10:42:08+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-10-23T10:40:25+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-10-14T13:51:43+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-10-11T09:32:18+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-10-11T09:28:53+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"7c5d2f1f-f9da-4c31-89cd-41aa823e1781","owner":[],"postedDate":"November 4th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":57289470,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":57289471,"name":"Health sciences/Health care"},{"id":57289472,"name":"Physical sciences/Mathematics and computing"}],"tags":[],"updatedAt":"2026-01-19T16:50:47+00:00","versionOfRecord":{"articleIdentity":"rs-7814494","link":"https://doi.org/10.1038/s41598-026-35552-z","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2026-01-17 16:31:12","publishedOnDateReadable":"January 17th, 2026"},"versionCreatedAt":"2025-11-04 12:20:48","video":"","vorDoi":"10.1038/s41598-026-35552-z","vorDoiUrl":"https://doi.org/10.1038/s41598-026-35552-z","workflowStages":[]},"version":"v1","identity":"rs-7814494","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7814494","identity":"rs-7814494","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}