语言模型驱动的乳腺癌病历增强工具用户界面测试
A mixed-methods evaluation of a language model-powered breast cancer note augmentation tool for patients
研究团队开发了一款使用语言模型(GPT-4-031)生成病历增强工具,包括简化文本、定义医学术语、常见问题解答等。通过200名健康女性和8份病历的测试,使用增强工具的参与者自报信心更高,理解能力更强。临床医生的错误分析显示大多数错误较小。该研究强调了在临床部署前需进行患者中心的混合方法测试和临床错误分析。
Abstract
Large language models (LLMs) have the potential to make clinical notes more understandable to patients, but tools must be tested with patients pre-deployment. We developed a patient-facing tool that uses an LLM (GPT-4-031) to generate “augmentations” to notes: simplified note text, defined medical terminology, frequently asked questions, extracted key information, and to-do items. We evaluated the tool with 200 healthy female participants and 8 notes. Compared to control, participants with access to augmentations had greater self-reported confidence (2.83 ± 0.05 augmentations, 2.18 ± 0.07 control, p<0.001) and scored higher on comprehension questions (0.63 ± 0.02 augmentations, 0.54 ± 0.04 control, p<0.001). We conducted interviews with 7 participants with a history of breast cancer and their own notes who described ease of use and perceived improvements in note understanding, but also reported concerns around LLMs. Three clinicians conducted an error analysis and found most errors to be minor and related to assumed context.
Introduction
Patients with cancer often receive large amounts of information about complex, interdisciplinary care. Coupled with anxiety about their diagnosis, prognosis, and treatment, it is common for oncology patients to struggle recalling information delivered to them during a clinic visit1. Electronic access to their clinical notes allows patients to revisit details of their visit at their own pace and improve understanding of their care2. Oncologists, however, continue to be divided on whether access to “open notes” is beneficial to patients, reporting concerns that documentation will take longer and that reading notes may cause confusion or worry among their patients3. A study analyzing over 100,000 notes from 60 oncologists at one institution showed that notes—particularly the “assessment and plan” sections—became longer and easier to read after the implementation of open notes, suggesting that providers may indeed be adjusting their writing for patients4.
Large language models (LLMs) may be able to combat these concerns by automatically simplifying notes, acting as an extension of the provider towards their patients. LLMs can simplify medical text while retaining accurate grammar and meaning5, translate medical text into patient-friendly language6, and match or even outperform clinicians in generating clinical summaries7. Researchers have suggested leveraging these applications in the field of oncology, where an LLM could translate clinical notes into simpler language, summarize recommendations from different specialists, and generate a list of next steps8. However, these augmentations need to be tested not only with oncology providers but also with their patients before an LLM-powered tool is deployed in the clinic.
In this study, we seek to move this work forward by designing and piloting a patient-facing user interface that allows patients to read clinical notes with different augmentations generated by an LLM on the backend. The augmentations include definitions, which recognizes and defines medical terms in a note, simplifications, which translates the original clinical note into layperson language, FAQ, which generates common questions and answers based on the note, key information, which extracts information such as appointments from the note, and to-do list, which extracts action items such as making follow-up appointments. We propose two versions of the user interface to support patients: Select, including definitions and simplifications augmentations, and All augmentations, which includes all those mentioned above. We focus on the area of breast cancer, the most common cancer among women worldwide.
We evaluated our user interfaces and augmentations via a survey study with 200 healthy participants to determine if LLM-augmented notes can improve comprehension. We then conducted in-depth interviews with patients with a history of breast cancer to better understand the acceptability of an LLM-powered user interface and how the augmentations can affect care. Finally, three clinicians conducted an error analysis of the augmentations, describing error types. We repeated this error analysis with GPT-5.5-generated augmentations and found fewer errors. Each of these evaluations (overview in Fig. 1) provide insights into how augmentations can be improved in the future.
Overview of study design. Synthetic and real notes (left), and two sets of augmentations (middle), were studied. Augmentation evaluation was conducted using mixed methods. Researchers on the team analyzed the results of the first several methods, and clinical research partners focused on clinical review (person with + icon). Readability was analyzed automatically. We used methods including both surveys (pen icon) and interviews (chat bubble icon), as well as clinical review (person with + icon) and automated analyses (computer monitor icon). Several of these evaluations focused on identifying errors (last column).
This combined error analysis and mixed methods evaluation with patients, supported by a patient-centered user interface, is a unique analysis of LLM-based simplification of medical text compared to related work, which typically focuses primarily on clinical review, automated metrics, LLM-as-a-judge evaluation, or some combination of these, but lack patient evaluations6,9,10,11,12,13,14, as observed in a recent systematic literature review15. Some include both patients and clinicians, but lack detailed error analyses16,17,18. Our findings from this multi-faceted evaluation suggest promise in LLM-based augmentations of medical text, but caution against deployment without patient-centered mixed methods evaluations and clinician-centered error analyses due to patient preferences and the presence of errors.
Methods
We will now describe our methods for collecting and augmenting clinical notes, creating the user interface, and conducting the mixed methods study.
Clinical notes augmentation types
We propose five types of augmentations to clinical notes to support patients. Each of the augmentations was obtained by passing a clinical note with a natural language instruction tailored for each augmentation as an input prompt to the LLM GPT-4 (specifically version GPT-4–0314)19. We employed prompt engineering strategies specific to each augmentation, primarily focusing on zero-shot prompting and prompt chaining20. We present the TRIPOD-LLM21 checklist in Appendix H.
-
Definitions: recognizing and defining medical terms in clinical notes. The prompt was to “Make a list of the medical words that appear in this clinical note. Write each word as it appears in the note without modification. Define each of these terms. Definitions should be 1–2 sentences in simple English.”
-
Simplification: translating the original clinical note into layperson language. The prompt was to “Rewrite this clinical note in simple English. Do not remove names of medications or diagnoses.”
-
FAQ: generating common questions and answers based on the clinical note. The prompt was to “Make a list of 7 common questions a patient might ask their doctor after reading this clinical note. Answer the questions in 2–3 sentences using simple English.”
-
Key Information: extracting key information from the clinical note. The prompt was to “Make a list of the medications mentioned in this clinical note. Make a list of the upcoming appointments explicitly mentioned in this clinical note. If the note does not mention appointments, say ’No Appointments’ ”
-
To-do List: extracting any action items that the patient needs to remember. The prompt was to “Make a todo list of tasks the patient is responsible for before their next appointment based on the clinical note. Tasks need to be specific and actionable. Tasks should not be recurring events. If there are no such tasks mentioned in the note, say ’No Tasks’ ”
Once we generated each of these augmentations, we integrated them into a user interface for patients to interact with to improve their experience reading clinical notes.
User interface
We developed a user interface, implemented as a web app, to integrate the augmentations presented in Section "Clinical Notes Augmentation Types" (Fig. 2). In the first panel, the reader has access to a tab containing the original clinical note and a second tab containing the generated simplified note (Fig. 1). On the original note, each medical term is highlighted in bold green and underlined to indicate that the user can click on it. When a user clicks on a medical term, the page will automatically scroll to the corresponding definition which is available in the third panel. The second panel consists of three tabs for the augmentations FAQ, Key Information, and the To-Do List. For FAQ, the reader can expand on each question to see its answer. In Fig. 3, we show explicit examples of the augmentations generated for one of the clinical notes used in our work.
Our user interface to assist in the comprehension of clinical notes, containing (from left to right) a toggle between the original clinical note and simplified augmentation, an optional pane with the FAQ, key information, and to-do list augmentations, and an optional definitions augmentation pane.
Examples of the augmentations considered in our interface using a sample excerpt from a synthetic clinical note.
Clinical notes collection process
To evaluate our interface, we needed a set of clinical notes for experimentation. We relied on two sources of clinical notes: synthetic notes written by a clinical research partner and real notes donated by patients with a history of breast cancer. The clinical research partner created four synthetic notes that mirror real clinical notes. These notes were inspired by the clinician research partner’s professional experience, but do not contain information about any of their patients. These synthetic notes were written at an intermediate difficulty, in the style of a clinician who is not writing for the purpose of patient comprehension.
We obtained eight real notes that were de-identified and donated by interview study participants, with note dates ranging from 2007 to 2023 (recruitment details can be found in Section "Interview Study Design"). The participants were self-identified patients with a history of breast cancer. Participants were told to look for clinical notes containing free text describing an appointment during their breast cancer treatment. Suitable notes might contain sections labeled “Impression”, “Assessment & Plan” or “Plan”. We received written informed consent from participants before they donated their clinical notes and additional written informed consent before participating in the video interview. All participants were given a verbal overview of the purpose and risks listed in the consent forms, and were given an opportunity to ask questions to the research team before signing. This data collection process was approved by the MIT Committee on the Use of Humans as Experimental Subjects (COUHES), as detailed in Appendix G. All research was performed in accordance with relevant guidelines/regulations.
In total, our analysis drew from 12 clinical notes, each containing an average of 322 words. We focused on this limited set of notes because the clinician research partner determined the corpus to represent a wide variety of clinical scenarios and note-writing styles, and because of the intensive time required to curate synthetic notes and collect and de-identify donated notes. Despite the variety in this dataset, additional notes should be explored in future work.
Survey study design
To test these augmentations’ effects on patient understanding, we first conducted a randomized web-based survey with healthy participants. We recruited participants using the Prolific crowdsourcing platform with the following inclusion criteria: female-identifying, over 40 years of age, fluent in English. Participants first completed a survey about their demographics and a health literacy test (Confident with Forms22), then viewed three notes. Study participants were randomized to either observe synthetic or real (donated, de-identified) notes. Notes were randomly selected from a pool of four synthetic and four real notes. See Appendix I.
Participants read the note with a given level of LLM augmentations, answered four multiple-choice action-oriented content questions with objective correct or incorrect answers, and then self-reported their confidence and comprehension of their reading. Each note was accompanied by one of three levels of augmentations: Control (no augmentations), Select (only simplified and definitions augmentations), and All (all simplified, definitions, FAQ, key information, and to-do list augmentations). All participants saw all three levels of augmentations in a randomized order.
Per note, augmentation, and participant, we obtained an action understanding score in the range [0, 1], a self-reported confidence score about understanding the care of the patient in [0, 3] (self-comprehension), and a self-reported confidence score about understanding the next steps of the patient in [0, 3] (self-confidence). At the conclusion of the study, we requested a list of augmentations they would like to use again, and participants additionally described what they used to answer the multiple-choice questions in a free-form text response. Information identified by our clinician research partner as misleading or false was explicitly listed and corrected at the end of the study to avoid misleading participants. The study took approximately 20 minutes and participants were compensated with $3.75, calculated based on Prolific’s fair payment principles. Full details of the survey study can be found in Appendix F. This study was determined exempt from approval by COUHES, with details in Appendix G. All research was performed in accordance with relevant guidelines/regulations.
Interview study design
To evaluate the impact of our proposed clinical note augmentations and supplement the survey with more detailed feedback, we also conducted in-depth interviews to explore patient perspectives. Our COREQ23 checklist is in Appendix H. We conducted purposive recruitment of participants who self-identified as patients with a history of breast cancer by advertising via the crowdsourcing platform Prolific and a clinical center at our institution. Details about participants’ backgrounds is reported in Table 1. Interview studies were conducted virtually with two study team members present, including one clinical partner, following a semi-structured interview protocol consisting of exploring our proposed augmentations interface on their own donated notes and answering a series of questions about the augmentations, including a Likert rating for each augmentation. Participants in the interviews received compensation in the form of a $30 gift card for their approximately 45- to 60-minute interaction. All sessions were audio- and video-recorded via a teleconference platform, then transcribed manually.
One researcher conducted interviews virtually from a private workplace, with participants joining from anywhere they preferred. A clinical research partner was present for all interviews to answer general safety-related medical questions as needed. Both researchers introduced themselves at the start of the interview and reviewed the study goals. Participants were asked to answer demographic questions. A semi-structured interview protocol was developed based on the initial survey findings. Participants were asked about their prior experience reading clinical notes, and then were given a customized link to the user interface where one of their donated clinical notes and its augmentations had been preloaded. They were asked to think aloud as they used the tool, and specifically were asked what they liked and disliked, whether anything about the augmentation worried them and if they identified any errors, and finally to rate on a Likert scale from 1−5 how much the augmentation assisted their reading experience. Full details of the interview protocol and questions are in Appendix B. This study was approved by COUHES (Appendix G). All research was performed in accordance with relevant guidelines/regulations.
Two researchers (BDL, EBK) used DeDoose software and Braun and Clarke’s methods for thematic analysis for qualitative analysis of the transcribed interviews24. One researcher first analyzed the interviews to derive major themes for an initial codebook, which was reviewed by the second researcher. Both researchers independently coded all of the interviews and iteratively updated the codebook and reconciled discrepant codes until consensus was reached. Further details are in Appendix C.
Error analysis of augmentations
All augmentations underwent manual review by three clinician research partners to identify misleading or false information. One clinician research partner conducted an initial examination of all notes and augmentations (BDL), documenting any errors found along with their rationale for categorizing them as errors. The other two clinicians (TU, CER) then validated these findings, to ensure consistency in our findings. Once evaluated by all three clinician research partners, they were sorted into error categories (“Omission”, “Not Simplified”, “Incorrect Reasoning”, “Incorrect Fact”, “Misleading Phrasing”, “Awkward Phrasing”, “Hallucination”, “Timing”, “Brand Name”, “Unnecessary”, “Clinician Preference”, “Assumed Context”) by multiple coauthors, including the clinician research partners and non-clinical research partners. We worked together to reach consensus assigning categories to each error found. The full categories are described in Appendix E.
In addition to manual review, we also conducted an automated analysis of the augmentations, acknowledging that automated analyses do not always correlate with human assessment of readability25. We reported the Flesch-Kincaid Grade26,27, which showed most correlation in28, and the percentage of words in a dictionary (Carnegie Mellon Pronouncing Dictionary Corpus Reader (https://www.nltk.org/_modules/nltk/corpus/reader/cmudict.html.)), to compare the acronyms and jargon in the original versus augmented notes.
Statistical analysis
To obtain 95% confidence intervals for mean estimates of outcomes, we used a method to correct for inter-subject variability29. To compare outcome means between two conditions, we used the pairwise Wilcoxon Signed Rank test to obtain p-values. We corrected for multiple hypothesis testing with the Benjamin-Hochberg adjustment procedure with \(\alpha =0.05\). We also conducted subgroup analyses, such as for those who never accessed a health portal, cancer patients and health professionals, those with an associate degree or less, and those who were not White race, to see whether overall trends held. All analyses were performed in R version 4.1.1 with the stats package.
Results
Study population
For our survey study, we recruited 200 participants through Prolific, many of whom had accessed a health portal, had adequate health literacy, and agreed that AI can have a beneficial impact in medicine. Most participants in both components of the study identified as female, and the majority of participants identified as white and non-Hispanic, Latino, or Spanish origin. The full demographic breakdown of the participants is in Table 1.
For our interview study, all seven participants had accessed a portal for patients previously, demonstrated adequate health literacy, and all but one participant agreed with the statement that AI can have a beneficial impact in medicine. We had strict inclusion criteria for the interview study; participants were required to have a history of breast cancer and be willing to donate their real clinical notes for evaluation.
Quantitative analysis
Results from our web-based survey for participants’ action understanding score for the Control, All, and Select conditions. We show scores for all notes (a), synthetic and real notes (b), and scores across multiple subgroups (c). Generally, Select leads to the best action understanding score.
In the survey study (N=200), we found that participants reading notes with either Select or All augmentations had statistically significantly higher action understanding scores, self-reported comprehension scores, and self-reported confidence scores when considering both real and synthetic note conditions in aggregate. Notably, with All augmentations compared to Control (no augmentations), action understanding scores increased from 0.537 to 0.605 (p=0.02), self-comprehension scores increased from 1.45 to 2.25 (p< 0.001), and self-confidence scores increased from 2.17 to 2.95 (p< 0.001) (Fig. 4 (a)). We found similar observations when comparing Select to Control.
We found that the trend of Select and All augmentations increasing self-reported and action understanding scores held when looking at specific subgroups of the study population, which we selected prior to analysis of the results (Fig. 4 (c)). We note that Select and All augmentations had higher action understanding, self-comprehension, and self-confidence scores on average in each of the four subgroups, though we did not evaluate statistical significance. We also looked at other possible subgroup categorizations post-hoc and observed similar trends.
When we split the participants based on the condition assigned of observing either synthetic notes or real notes, we found that Select and All augmentations statistically significantly improved action understanding scores only for the real notes and not for the synthetic notes (Fig. 4 (b)). For participants in the real notes condition, action understanding scores increased from 0.451 in Control to 0.541 in All augmentations (p=0.04); for participants in the synthetic notes condition, scores changed from 0.622 in Control to 0.669 in All (p=0.30).
The top two augmentations that participants indicated they would use if they were to read another clinical note were simplifications chosen by 83.5% (95% CI 77.7–88.0) of participants and definitions chosen by 72.5% (95% CI 65.9–78.2) of participants. The remaining preferred augmentations were To-Do-List 42.0% (95% CI 35.4–48.9), FAQ 41.0% (95% CI 34.4–47.9), and Key Information 35.5% (95% CI 29.2–42.3). In responses to the question “What did you use to answer these questions? Was it easy or hard?”, the most common word was “simplified” (occurred 97 times), followed by “notes” (71), “easy” (58), “original” (54), and “definitions” (54).
Qualitative analysis
Interview participants (N=7) described three major themes: (1) Current practices and attitudes towards note-reading, (2) Benefits from the augmentations, and (3) Concerns with the augmentations (Table 2). Within theme one, current practices around note-reading involved gathering information, checking the narrative, and understanding medical terminology: Participants shared that they read notes for facts such as lab results and future appointments, and to help their own memory and ensure they understood the visit correctly. Participants reported that they read notes in order to assess the narrative, for example ensuring that the clinician understood them and was aligned in their priorities. Many also described notes as a useful window into the clinician’s thinking: “You kind of want to have that insight into what the doctor’s take was on it” [P2]. All participants reported that medical terminology makes notes hard to understand, which can lead to time-consuming research and fear of the unknown.
All participants rated the definitions augmentation highest in the Likert ratings in terms of greatest benefit to note-reading. Participants described two main benefits with augmentations: (1) improved understanding of the note and (2) ease of use (Table 2). Participants reported that augmentations made notes easier to read, less stressful, more personable, and in some cases prompted new questions and ideas. A few participants commented that the tool performed intuitively, highlighting jargon they would have struggled with and emphasizing takeaways they also found to be important. Many reported that it was convenient to have the tool integrated with the note and that the tool kept them focused: “When I didn’t have this, I would go off on websites that probably didn’t even apply to my type of cancer and they were doom and gloom, I would literally sit at my desktop and cry. It was so scary” [P4]. Participants also liked that they were able to toggle between the original note and the augmented note in order to check for accuracy.
Two major themes emerged around concerns: (1) That the augmentations may miss something or mislead patients and (2) more information could be confusing (Table 2). One participant said, “I feel like there’s a tiny bit of anxiety being like, ’I’m having to trust that the AI got it right and that this is correct”’ [P2]. Others reported that augmentations may lose nuance in their oversimplification, misrepresent the gravity of the situation, or cause patients to think they understand more than they do. Participants also pointed out perceived discrepancies, such as a medication in the generated to-do list that they were no longer taking. We present our full codebook for the interviews in Appendix D.
Error analysis
Error analysis of GPT-4 augmentations. (a) (b) and (c) show the number of errors in different categories (type, note, and augmentation, respectively), as determined by clinical reviewers. (d) and (e) are the result of automated analysis of GPT-4 augmentations, with error bars representing confidence intervals computed using t-scores, excluding empty note augmentations.
We report the results of the error analysis in Fig. 5. We found that GPT-4 can make misleading errors in the context of breast cancer clinical note augmentations, and identified 82 total errors. In total, 28 (34%) errors were considered to be potentially harmful and misleading (Fig. 5 (a)). However, minor errors were most common, especially assumed context (Fig. 5 (a)). For example, a definition augmentation specifying that lower outer quadrant referred specifically to the breast, though the term can refer to a specific part of other tissues as well. Many of these errors appeared in the definitions augmentation, and errors were especially common in one note’s augmentations (Fig. 5 (b-c)). In particular, most errors occurred in augmentations to real notes, while only one error appeared in augmentations to the synthetic notes (Fig. 5 (b)).
In the automated analysis, we found that the percentage of words in the dictionary increased over control in all augmentations. Notably, when comparing control versus the simplified augmentation, the percentage increased for all notes, and the results of a paired-sample t-test are (\(t=-4.61\), \(p=0.0007\)) (Fig. 5 (d)). Moreover, the Flesch-Kincaid grade level decreased compared to control for the simplified and to-do list augmentations, when comparing control versus the simplified augmentation, the results of a paired-sample t-test are (\(t=2.24\), \(p=0.047\)). However, augmentations like definitions increased in grade level (Fig. 5 (e)), which could be due to additional length and medical terminology in each of the definitions. This was also observed by some of the qualitative interview participants. Future work could explore further simplification of the definitions and evaluation on a larger number of notes.
Limited GPT-5.5 error analysis
While we conducted our full mixed method evaluation with GPT-4-based augmentations, we conducted a limited error analysis and automated evaluation of GPT-5.5 results (gpt-5.5–2026-04–23), with details in Appendix J. These initial results show improvements in terms of automated readability metrics, number and severity of errors, and ability to find errors automatically via LLM-as-a-judge30. For example, there were approximately four errors (unnecessary, omission, omission, misleading phrasing) identified by our clinician research partner across augmentations and notes, which expanded to ten errors (adding an incorrect medical fact, two omissions, two misleading phrasings, and a hallucination) after our clinician research partner reviewed outputs from an LLM-as-a-judge evaluation. The LLM-as-a-judge independently identified 97 errors, including 33 errors it categorized as harmful, six of which our clinical collaborator agreed with and missed on an initial pass. The clinician found most other differences related to stylistic considerations, such as whether abbreviations should be expanded in the simplified augmentation or whether generic medication names should be provided alongside brand names when the original note included only the brand name. The clinician did note that several LLM-identified errors were not actually errors, and the LLM-as-a-judge procedure did not find those errors identified by our clinician research partner, indicating that clinical review is still necessary even in recent versions of GPT. Finally, we compared to the error analysis for GPT-4 to the error analysis for GPT-5.5. Errors from GPT-4 were resolved in GPT-5.5, but new errors were introduced in GPT-5.5 augmentations, indicating a need to continue reviewing outputs after versions are updated.
Discussion
In this mixed methods study, including manual and automated error analysis, quantitative surveys, and qualitative interviews, we showed promising preliminary results that an LLM can be used to improve understanding of clinical notes. Participants performed better on action-oriented content questions and reported self-perceived improvements in understanding and confidence when reading notes with LLM-generated augmentations.
Our exploratory subgroup analysis of vulnerable populations-less educated and nonwhite participants-showed that the benefits persist. These groups have previously reported the most benefit from reading clinical notes31, and leveraging LLMs to improve note readability may be an important tool for organizations’ diversity, equity, and inclusion efforts. While the participants who volunteered for our study might be more likely to view technology and AI favorably, approximately 33% of our survey respondents reported that they did not access their patient portal to read notes and 20% were undecided or disagreed that AI can have a beneficial impact in medicine, demonstrating heterogeneity in technology access and preferences. In our qualitative evaluation, patients reported that different people prefer different amounts of information about their health, and those information needs can vary over time.
Through clinical review of several LLM augmentations, we found and characterized error modes, making it clearer what work is needed to improve the accuracy of LLMs before they are deployed as standalone patient-facing tools. First, errors were more common in augmentations for real donated notes compared to augmentations for the carefully-written synthetic notes, suggesting that a real-world deployment would still be impacted by the quality of the original note, which can also be prone to errors32. Second, with our set of notes and corresponding GPT-4 augmentations, we found a high number of errors, including potentially harmful errors. We found fewer errors in GPT-5 augmentations for our notes, but errors, including potentially harmful ones, persisted. The LLM-as-a-judge approach offers a method for surfacing potential issues, which is important because constantly evolving models can make it difficult to reliably design, evaluate, and monitor these workflows in a real-world setting. Yet, final determinations continue to require clinician judgment. Furthermore, while hallucination errors were more rare, patients’ trust in qualitative interviews was still impacted by the errors we identified, even for those who were enthusiastic about the use of AI in medicine. Future work could explore patient evaluation of errors32,33,34. We note that in our web-based survey, we did not find an impact in action understanding score due to errors, likely because the identified errors only impacted the ability to answer the action-oriented questions in one note, and participants could still view the original note. Especially in the presence of errors, carefully presenting LLM-generated advice is extremely important to study for future deployment35. In this study, we began to consider this by presenting both the original and augmented notes for patients to reference.
Relatedly, note-based augmentations can only be as accurate as the underlying note, and LLMs may amplify existing documentation errors and biases32,36. In our study, patients preferred augmentations that provided medical definitions and simplified existing documentation. In the interviews, several participants also described the value of viewing the original note and emphasized the importance of reading notes not only for factual information but also to understand the clinician’s perspective and check for alignment. In the survey, we similarly hypothesize that Select augmentations slightly outperformed All augmentations because the additional augmentations in All may have opened up more potential misunderstandings.
Additionally, data security and patient privacy are critical considerations when applying LLMs in clinical settings. Clinical notes contain sensitive information, and protected health information (PHI) can only be used with commercial models under appropriate safeguards such as a Business Associate Agreement (BAA). Not all institutions have access to PHI-compliant models, and the development of note-reading tools requires careful planning in collaboration with institutional information technology infrastructure and security experts. Autonomy and AI overreliance (e.g37.,) are other important ethical considerations to consider before deployment. As more providers use ambient scribes–many of which rely on LLMs to generate the clinical note–it will also become more challenging to discern what the source of truth is from the visit. Regardless, patients are already interacting with LLMs and inputting medical notes and other data to prepare for clinical encounters38,39. Our study provides important insights into patient preferences and the shortcomings of these interactions, both of which can inform future co-development.
There are limitations to this study. We used a limited number of real notes, and in order to slightly expand our pool of available notes for augmentation, the synthetic notes we added were likely less complex and variable than the real donated notes. In fact, with these initial results, we found that GPT-4 augmentations led to improved scores on the real notes only, and most of the identified errors were found in the real notes. Future experiments may seek to evaluate the patient experience prospectively on real notes only. While our participants reflected the median age (63) and race (non-Hispanic white women) of those with the highest incidence of breast cancer in the United States40,41, evaluating a larger, more diverse cohort of patients at a medical center may also help to minimize the selection bias present in our study; participants who volunteered for our study may not represent all potential users. In addition, we did not compare to human-generated augmentations as they are likely infeasible due to challenges like physician burnout4,42,43, but this may be valuable in future work. In addition, further research is needed to improve LLMs for clinical use, particularly for tools that are used by patients. While we believe from our limited analysis of GPT-5.5 that performance has improved with more recent versions, another important direction to explore is improved interface design to make it easier for patients to notice potentially misleading augmentations in-line. Additional prompt engineering or agentic systems may help create custom augmentations that are tailored to patient readiness, their education level, or even their preferred language44,45. Further research is needed to understand how clinicians’ and patients’ voices can be preserved while alleviating documentation challenges and improving communication and education.
Data availability
The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request.
Code availability
Code is available from the corresponding author upon reasonable request.
References
Gabrijel, S. et al. Receiving the diagnosis of lung cancer: Patient recall of information and satisfaction with physician communication. J. Clin. Oncol. 26, 297–302 (2008).
Blease, C., Salmi, L. & DesRoches, C. M. Open notes in cancer care: coming soon to patients. The Lancet Oncology 21, 1136–1138 (2020).
Moll, J. & Cajander, Å. Oncology health-care professionals’ perceived effects of patient accessible electronic health records 6 years after launch: A survey study at a major university hospital in Sweden. Health Informatics J. 26, 1392–1403 (2020).
Rahimian, M. et al. Open notes sounds great, but will a provider’s documentation change? An exploratory study of the effect of open notes on oncology documentation. JAMIA Open 4, ooab051 (2021).
Moramarco, F. et al. Towards more patient friendly clinical notes through language models and ontologies. In AMIA Annual Symposium Proceedings 2021, 881 (2022).
Zaretsky, J. et al. Generative artificial intelligence to transform inpatient discharge summaries to patient-friendly language and format. JAMA network open 7, e240357–e240357 (2024).
Van Veen, D. et al. Adapted large language models can outperform medical experts in clinical text summarization. Nature medicine 30, 1134–1142 (2024).
Bitterman, D. S., Downing, A., Maués, J. & Lustberg, M. Promise and perils of large language models for cancer survivorship and supportive care. J. Clin. Oncol. 42, 1607–1611 (2024).
Asgari, E. et al. A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation. NPJ Digit. Med. 8274. (2025).
Salvi, R. C. et al. Percs: Persona-guided controllable biomedical summarization dataset. arXiv preprint. arXiv:2512.03340 (2025).
Naghdi, M. et al. Artificial intelligence-simplified information to advance reproductive genetic literacy and health equity. Human Reproduction 40, 1681–1688 (2025).
Joseph, S. et al. Factpico: Factuality evaluation for plain language summarization of medical evidence. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 8437–8464 (Association for Computational Linguistics, Bangkok, Thailand, 2024).
Dehkordi, M. K. H. Enhancing patient comprehension: An effective sequential prompting approach to simplifying ehrs using llms. In 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) 6370–6377 IEEE, (2024).
Dehkordi, M. K. H., Lu, J., Perl, Y. & Deek, F. P. Enhancing patient comprehension of discharge notes with a retrieval-augmented llm approach. In 2025 IEEE International conference on bioinformatics and biomedicine (BIBM), 6966–6973 IEEE, (2025).
Krenn, C. et al. Automated approaches of text simplification of patient education materials: scoping review. Journal of Medical Internet Research 28, e88365 (2026).
Li, J. et al. ’Honestly, i could lose a few hours on pubmed quite easily’: Breaking down the reading barriers of medical literature with an LLM-powered browser extension. Behav. Inf. Technol. 0, 1–24 (2026).
August, T. et al. Paper plain: Making medical research papers approachable to healthcare consumers with natural language processing. ACM Trans. Comput.-Hum. Interact. https://doi.org/10.1145/3589955 (2023).
Kim, S.-I. et al. Enhancing patient participation in emergency department through patient-friendly clinical notes generated by large language models. Sci. Rep. https://doi.org/10.1038/s41598-025-31113-y (2025).
OpenAI. Gpt-4 technical report. arxiv:2303.08774 (2023).
Wu, T., Terry, M. & Cai, C. J. Ai chains: Transparent and controllable human-ai interaction by chaining large language model prompts. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, CHI ’22 Association for Computing Machinery, New York, NY, USA. https://doi.org/10.1145/3491102.3517582 (2022).
Gallifant, J. et al. The tripod-llm reporting guideline for studies using large language models. Nat. Med. https://doi.org/10.1038/s41591-024-03425-5 (2025).
Chew, L. D. et al. Validation of screening questions for limited health literacy in a large va outpatient population. Journal of general internal medicine 23, 561–566 (2008).
Tong, A., Sainsbury, P. & Craig, J. Consolidated criteria for reporting qualitative research (coreq): a 32-item checklist for interviews and focus groups. International journal for quality in health care 19, 349–357 (2007).
Braun, V. & Clarke, V. Using thematic analysis in psychology. Qual. Res. Psychol. 3, 77–101 (2006).
Zheng, J. & Yu, H. Readability formulas and user perceptions of electronic health records difficulty: a corpus study. Journal of medical Internet research 19, e59 (2017).
Flesch, R. A new readability yardstick. J. Appl. Psychol. https://doi.org/10.1037/h0057532 (1948).
Kincaid, J. P., Fishburne, R. P. Jr., Rogers, R. L. & Chissom, B. S. Derivation of new readability formulas (automated readability index, fog count and flesch reading ease formula) for navy enlisted personnel (University of Central Florida, Institute for Simulation and Training, 1975).
Zheng, M., Marsh, J. & Nickerson, J. How causal information affects decisions. Cogn. Res. Princ. Implic. https://doi.org/10.1186/s41235-020-0206-z (2020).
Morey, R. D. et al. Confidence intervals from normalized data: A correction to cousineau (2005). Tutorials in Quantitative Methods for Psychology 4, 61–64 (2008).
Gu, J. et al. A survey on llm-as-a-judge. In The Innovation (2024).
Gerard, M. et al. The importance of visit notes on patient portals for engaging less educated or nonwhite patients: survey study. Journal of medical Internet research 20, e191 (2018).
Bell, S. K. et al. Frequency and types of patient-reported errors in electronic health record ambulatory care notes. JAMA Netw. Open 3e205867. (2020).
Zheng, E. L., Jin, W., Hamarneh, G. & Lee, S.S.-J. From human-in-the-loop to human-in-power. The American Journal of Bioethics 24, 84–86 (2024).
Walker, J. et al. Opennotes after 7 years: Patient experiences with ongoing access to their clinicians’ outpatient visit notes. J. Med. Internet Res. 21, e13876 (2019).
Ghassemi, M. Presentation matters for ai-generated clinical advice. Nature Human Behaviour 1–3, (2023).
Adam, H. et al. Write it like you see it: Detectable differences in clinical notes by race lead to differential model recommendations. In Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society 7–21 (2022).
Gaube, S. et al. Do as AI say: susceptibility in deployment of clinical decision-aids. NPJ digital medicine 4, 1–8 (2021).
Paruchuri, A. “what’s up, doc?”: Analyzing how users seek health information in large-scale conversational AI datasets. In Findings of the Association for Computational Linguistics: EMNLP 2025 (eds Christodoulopoulos, C. et al.) 2312–2336 (Association for Computational Linguistics, 2025).
Ayo-Ajibola, O., Davis, R. J., Lin, M. E., Riddell, J. & Kravitz, R. L. Characterizing the adoption and experiences of users of artificial intelligence-generated health information in the united states: cross-sectional questionnaire study. Journal of medical Internet research 26, e55138 (2024).
Race and Ethnicity. https://www.komen.org/breast-cancer/risk-factor/race-ethnicity/.
Breast cancer risk factors: Age. https://www.komen.org/breast-cancer/risk-factor/age/.
Blease, C. et al. Patient online record access in english primary care: qualitative survey study of general practitioners’ views. Journal of Medical Internet Research 25, e43496 (2023).
DesRoches, C. M. et al. The views and experiences of clinicians sharing medical record notes with patients. JAMA Network Open 3, e201753 (2020).
Zhu, W. Multilingual machine translation with large language models: Empirical results and analysis (2023) arXiv preprint. arXiv:2304.04675
Dumenci, L. et al. Measurement of cancer health literacy and identification of patients with limited cancer health literacy. Journal of health communication 19, 205–224 (2014).
Acknowledgements
We would like to thank Vaibhav Balloli, Hunter Lang, Shannon Shen, Daniel Ajayi, and Katie Liu for their helpful discussions throughout the project. We are grateful to the MIT Center for Clinical and Translational Research for their generous help with our human participant experiment. We also thank the reviewers at all stages of our paper development for the constructive feedback, which improved our work.
Funding
EBK was funded by the MIT CSAIL METEOR Fellowship. NM and DS were partially funded by the Mark Foundation for Cancer Research. M Asiedu was supported by the Schmidt Science Postdoctoral Fellowship. M Agrawal was supported by the MIT Deshpande Center and the MLA@CSAIL initiative. This project was also supported by the National Center for Advancing Translational Sciences, National Institutes of Health, Award Number UL1TR002544. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.
Author information
Author notes
These authors contributed equally: Niklas Mannhardt and Elizabeth Bondi-Kelly.
Authors and Affiliations
Massachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge, 02139, MA, USA
Niklas Mannhardt, Elizabeth Bondi-Kelly, Hussein Mozannar, Mercy Asiedu, Tatiana Urman, Catherine E. Ricciardi, Monica Agrawal, Marzyeh Ghassemi & David Sontag
University of Michigan, 2260 Hayward St, Ann Arbor, 48109, MI, USA
Elizabeth Bondi-Kelly
Division of Hematology and Oncology, Department of Medicine, University of Washington, 825 Eastlake Avenue East, Seattle, 98109, WA, USA
Barbara D. Lam
Microsoft Research, 14820 NE 36th St, Redmond, 98052, WA, USA
Hussein Mozannar
Massachusetts General Hospital, 55 Fruit St, Boston, 02114, MA, USA
Chloe O. Stanwyck
School of Medicine, Stanford University, 291 Campus Drive, Stanford, 94305, CA, USA
Chloe O. Stanwyck & Alejandro Buendia
Department of Oncology, Mayo Clinic, 5777 E. Mayo Blvd., Phoenix, 85054, AZ, USA
Irbaz Bin Riaz
Duke University, 2424 Erwin Rd, Durham, 27708, NC, USA
Monica Agrawal
Authors
- Niklas Mannhardt
- Elizabeth Bondi-Kelly
- Barbara D. Lam
- Hussein Mozannar
- Chloe O. Stanwyck
- Mercy Asiedu
- Alejandro Buendia
- Tatiana Urman
- Irbaz Bin Riaz
- Catherine E. Ricciardi
- Monica Agrawal
- Marzyeh Ghassemi
- David Sontag
Contributions
All authors contributed to writing and conceptualizing the study. In addition, authors made the following contributions. NM generated augmentations, developed the user interface and survey, and conducted qualitative interviews. EBK guided study design and conducted analyses of augmentations, survey results, and qualitative interview results. BDL contributed to study design, wrote synthetic clinical notes, conducted clinical error analysis, and coded qualitative interview results. HM contributed to survey study design and conducted results analysis. CS contributed to study design and qualitative interviews. M Asiedu contributed to error analysis and project conceptualization, including project scoping, augmentation selection, and tool design. AB contributed to project conceptualization and augmentation generation. TU and CER contributed to study design and preparation for institutional ethical review, as well as clinical error analysis. IBR contributed to project conceptualization, including augmentation selection. M Agrawal contributed to project conceptualization and error analysis. MG contributed to study design and interpreting results. DS guided the full project, including study design and analyses.
Corresponding author
Correspondence to Elizabeth Bondi-Kelly.
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
About this article
Cite this article
Mannhardt, N., Bondi-Kelly, E., Lam, B.D. et al. A mixed-methods evaluation of a language model-powered breast cancer note augmentation tool for patients. Sci Rep 16, 37054 (2026). https://doi.org/10.1038/s41598-026-59948-z
Received: 03 April 2026
Accepted: 24 June 2026
Published: 05 November 2026
Version of record: 05 November 2026
DOI: https://doi.org/10.1038/s41598-026-59948-z
来源:Nature 健康护理主题 · nature.com