Clinical Text Summarization: Adapting Large Language Models Can Outperform Human Experts

“Large language models (LLMs) have demonstrated astounding performance, propelled by both the transformer architecture [73] and increasing scales of data and compute, resulting in widespread adoption of models such as ChatGPT [6].”

“We estimate intra-reader correlation based on a mean-rating, fixed agreement, two-may mixed effects model [42] using the Pingouin package [80].”

“Large language models (LLMs) have demonstrated astounding performance, propelled by both the transformer architecture [73] and increasing scales of data and compute, resulting in widespread adoption of models such as ChatGPT [6].”

“BLEU [52], the simplest metric, calculates the degree of overlap between the reference and generated texts by considering 1- to 4-gram sequences.”

“Low-rank adaptation (LoRA) [31] has emerged as an effective, lightweight approach for fine-tuning LLMs by altering a small subset of model weights—often < 0.”

“The original T5 “text-to-text transfer transformer” model [55] demonstrated excellent performance in transfer learning using the seq2seq architecture.”

“Beyond the scope of our work, there’s further potential to reduce hallucinations through incorporating checks by a human, checks by another LLM, or using a model ensemble to create a “committee of experts” [11, 39].”

“Despite the general nature of training, LLMs were able to answer clinical questions, solve diagnostic problems, and analyze medical texts, even if the model was not specifically trained on them [14, 15].”

“Large language models (LLMs) have demonstrated significant advances in medical applications [11,12], including information extraction from electronic health records [13], clinical text summarization [14], and document classification [15].”

“In the medical field, LLMs were applied to tasks such as clinical note summarization[6, 23] and question-answering [2, 4].”

“The tasks include medical text summarization [19], [20], medical text generation [21], and text simplification [22].”

“14,15,22,54,55 This study uses a language model fine-tuned on consensus data, annotated by clinical experts, to accelerate ontology engineering.”

“LLMs have shown remarkable capabilities across various tasks, including chatbot interactions [15, 3, 5], text summarization [13, 4, 14], and content creation [2, 7, 8], demonstrating their potential to vastly improve programming and documentation practices.”

“We used Gemini 1.0 Pro as the teacher model in this study. achieved using GPT-4 along with few-shot in-context learning (Giorgi et al., 2023; Van Veen et al., 2023).”

“Along with these findings, a large number of evidence is accumulating that LLMs can perform significantly well on clinical tasks other than solving structured clinical questions, such as clinical text summarization, when appropriate techniques are used for aligning the LLMs [24].”

“Continued domain-specific pretraining [105, 106, 107, 108, 109] has demonstrated improvements in clinical reasoning and documentation tasks.”

“With the rise of LLMs, Van Veen et al. [21], [22] demonstrated that adapted LLMs can even outperform medical experts on summarizing clinical notes while achieving higher relevance and accuracy scores than human-written summaries.”

“Research has shown that LLMs fine-tuned in the medical domain can generate summaries that are as accurate as, or even more accurate than, those written by human doctors (87% of clinicians rated LLM-generated summaries as equivalent or superior to human-written ones) [16] [15].”

“Artificial intelligence has been widely applied in the medical domain for tasks such as medical image captioning [6], clinical text summarization [26], discharge note generation [10], and medical question answering [28].”

“Overall, LLMs show promising potential to revolutionize research synthesis in healthcare, demanding further research to optimize their impact on systematic review quality [14].”

“Veen et al. [11] further showcased the capability of large language models in summarizing complex texts and conversations, aligning with our approach of mapping developer-tool interactions to specific labels.”

“Large GPU becomes a challenge for non-commercial purposes due to the lack of high resources in order to process big datasets [33-34].”

“Supervised ML approaches face difficulties in domain adaptation, and fine-tuning does not fully address these issues [1][3][5].”

“Each section was generated independently by LLM [30,31] using structured, section-specific prompts.”

“Continued domain-specific pretraining and multimodal extensions further integrate structured data and temporal signals into these generative frameworks [50, 51, 52, 53].”

“26 31 Moreover, advanced prompting techniques, such as chain-of-thought reasoning and stepwise decomposition, allow LLMs to articulate their decision-making process transparently, improving the interpretability and confidence.”

“Large language models have recently been applied to clinical summarization tasks, including discharge summaries, radiology reports, and longitudinal EHR compression (Van Veen et al., 2023; 2024).”

“drafting notes, extracting variables, and assisting with triage [1–3].”

“19 Another study evaluated LLMs in clinical note generation across 12,999 clinician annotated sentences and demonstrated 1.47% hallucination rate and 3.45% omission rate and with further refining prompts and workflows they achieve errors below previously human generated notes and it highlights the…”

“Moreover, the study relies on independent scoring of the model’s responses by two experts, which introduces subjectivity, and the lack of inter-rater reliability reporting further undermines the objectivity of the evaluation results [2].”

“Due to the complexity, length, and time-intensive nature (17,18).”

“Adapted LLMs can outperform human experts in summarizing clinical texts [163], improving documentation efficiency.”

“For example, [Van Veen et al ., 2023] demonstrated that adapted LLMs can outperform medical experts in clinical text summarization tasks, suggesting their potential in alleviating documentation burdens in healthcare settings.”

“This finding is in line with the studies of Van Veen et al. [9] [10], who evaluated the adaptation of LLMs for clinical text summarization and showed that models such as GPT-4 and Llama-2 outperformed human experts in completeness and correctness.”

“41 Veen et al. 42 applied QLoRA for summarizing radiology reports and progress notes, noting its potential to reduce documentation burden.”

“LLMs have been shown to extract information from many kinds of documents, such as full-length novels (Chang et al. 2023), elec-a tronic health records (Van Veen et al. 2023), and financial documents (Kim, Muhn, and Nikolaev 2024).”

“In particular, they have also proven to outperform previous approaches in the above tasks by far (Van Veen et al., 2023; Al-Thanyyan and Azmi, 2021; Kocmi and Federmann, 2023).”

“While fine-tuned models generally outperform their counterparts in few-shot setting(Van Veen et al., 2023b,a; Lewis et al., 2020), there is a growing interest in exploring domain adaptation techniques, mainly in-context learning, which do not require additional training (Agrawal et al., 2022;…”

“According to Van Veen et al. (2023), sifting through vast textual data and summarizing key information from electronic health records (EHR) imposes a substantial burden on clinicians' time.”

“Authors in [7] evaluated LLMs on clinical note generation tasks and found hallucination rates ranging from 15–35% depending on model and task complexity.”

“Their functional role depends on their position within the pipeline and the degree of responsibility assigned for signal interpretation, content planning, and clinical reasoning 202 .”

“LLMs have demonstrated strong performance across clinical tasks including clinical summarization [33, 77], diagnostic reasoning [30, 43], and question answering [31, 49, 67, 68].”

“Web-scraping tools can now continuously gather relevant news articles as they emerge, for instance, while large language models (LLMs) are increasingly capable of generating summaries and synthesizing content [20, 46].”

“…is expressed via abbreviations, paraphrases, or normalization (e.g., “acute renal failure” → “AKI”), a limitation typical of extraction/matching-based concept metrics [18]; however, they remain appropriate for relative comparison across ablation variants under identical extraction assumptions.”

“Meanwhile, large language models have demonstrated impressive summarization capabilities but also raise concerns related to hallucination and omission, which are unacceptable in safety-critical clinical contexts [11], [12], [13], [14].”

“LargeLanguage Models (LLMs) have demonstrated potential in medical Question Answering (QA) [6, 17], clinical summarization [2, 29], extracting medical information [7, 8, 33].”

“They have also been applied to clinical summarization, some-times outperforming clinicians (Fraile Navarro et al., 2025; Van Veen et al., 2024).”

“Van Veen et al. [9] conducted an extensive evaluation showing that large language models adapted to clinical text can match or exceed the performance of human medical experts on summarization tasks.”

“In an era where ED overcrowding and staff burnout are pervasive, these efficiency improvements are essential for the sustainability of emergency care delivery [7].”

“Large language models have recently been applied to a wide range of healthcare tasks, including clinical text understanding, medical question answering, and patient-level prediction from unstructured notes [53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63].”

“Veen et al. evaluated several open-source models (FLAN-T5, LLaMA-2, Vicuna, and Med-Alpaca) for clinical text summarization and found that adapted models using in-context learning (ICL) or QLoRA find-tuning outperformed both humans and zero-shot LLMs [18].”

“We randomly select 100 out of 126 samples containing an explicit “assessment and plan” section in the test dataset for our analysis [6].”

“Van Veen et al. (2023) [22] adapted LLMs for clinical text summarization using in-context learning (ICL) [23] and QLoRA [24] fine-tuning, achieving state-of-the-art results in four clinical tasks.”

“LLMs such as the GPT family [3] are now being explored for their potential in radiology to assist in generating summaries [4], extracting relevant clinical insights, and identifying patterns across large datasets [5-7].”

“Continued research and development in this field will be essential to fully realize the potential of LLMs in transforming healthcare data management [17][20].”

“Artificial intelligence has been widely applied in the medical domain for tasks such as medical image captioning [ Cheng et al. , 2025 ] , clinical text summarization [ Van Veen et al. , 2023 ] , discharge note generation [ Jung et al. , 2024 ] , and medical question answering [ Yan et al. , 2024 ] .”

“This straightforward method improves LLM performance for a specific job or topic by giving the model context [13, 14].”

“7.”

“…from clinical tasks, such as diagnostic assistance, patient triage, and automatic treatment plan generation, to documentation tasks, such as summarization, de-identification, coding, and automatic structuring of clinical narratives, and finally in medical research and education (8, 10, 159, 160).”

“Recent advances in large language models (LLMs) [3], [4] have significantly improved performance across a wide range of single-turn natural language processing (NLP) tasks, such as text classification [5], [6] and summarization [7], [8].”

“These datasets are not suitable for training and assessing large-scale deep neural-based MDS models due to their relatively limited scale [18].”

“In their study, Zhang et al. [7] revealed that domain-specific LLMs, including Clinical-T5, LLaMA2, and GPT-4, have the capability to produce summaries that have been rated favorably by clinicians in terms of completeness, correctness, and conciseness compared to reference summaries.”

“By integrating vast biomedical knowledge with context-sensitive reasoning, these models have begun to complement clinical expertise, offering timely insights that enhance decision-making in uncertain or data-intensive scenarios [7], [8].”

“In the clinical setting, Van Veen et al. [62] evaluated LLMs for radiology reports, clinical notes, and doctor–patient dialogues, applying prompt engineering, in-context learning (ICL), and QLoRA fine-tuning to improve domain alignment.”