Unlocking the Ability of Generative Artificial Intelligence to Improve Health Communication and Health Education

Kanhai S. Amin, BS^1,2; David L. Skaggs, MD³; Rushabh H. Doshi, MD, MSc, MPH⁴; Pavan Khosla, BA¹; Sanjiv Amin, MBBS, MS⁵; Linda Mayes, MD² (View author affiliations)

Suggested citation for this article: Amin KS, Skaggs DL, Doshi RH, Khosla P, Amin S, Mayes L. Unlocking the Ability of Generative Artificial Intelligence to Improve Health Communication and Health Education. Prev Chronic Dis 2026;23:260490. DOI: http://dx.doi.org/10.5888/pcd23.260490.

PEER REVIEWED

Abstract

Introduction

While large language models (LLMs) can simplify information for patients, they are often ineffective at reaching specific reading levels. Here, we demonstrate a novel approach that may help improve readability and aid in the dissemination of health information to children and adolescents.

Methods

We selected for analysis 200 common and diverse pediatric conditions and symptoms. We used ChatGPT to generate 36 explanations for each condition and symptom. For each explanation, we calculated reading grade levels (RGLs) and word counts. From the 36 outputs, we then selected single outputs for 5th-6th grade (level 1), 7th-8th grade (level 2), 9th-10th grade (level 3), and 11th-12th grade (level 4) by selecting outputs within 0.25 RGLs of the median of the grade range (ie, 5.25–5.75 for level 1). Three full professors then reviewed outputs related to their pediatric specialties for accuracy and completeness; they modified all incorrect information and added relevant missing information. We then tabulated RGLs, word counts, and the number of modifications for the edited outputs.

Results

Prompt architectures varied in how well they initially resulted in a desired RGL, and few prompts resulted in requested RGLs at baseline. After output selection and accuracy edits, median RGLs were 5.8 (5.4–6.4) for level 1, 7.5 (7.4–7.7) for level 2, 9.5 (9.4–9.7) for level 3, and 11.5 (11.3–11.7) for level 4. Accuracy edits were minimal, with 93.5%, 95.5%, 95.5%, and 95.0% of outputs requiring no word additions or deletions for levels 1 through 4, respectively.

Conclusion

Our approach offers health care providers a way to generate health information at a chosen reading grade level throughout the care continuum.

Top

Introduction

Health literacy is essential to ensuring safe, high-quality health care delivery, as emphasized by organizations including the Joint Commission, the American Academy of Pediatrics, the Centers for Disease Control and Prevention, and the US Department of Health and Human Services (1,2). Since the term “health literacy” was first introduced in 1974, the link between low health literacy and adverse health outcomes has become clear (3,4). For example, greater health literacy has been shown to improve medication adherence and reduce hospital admissions among adults (5–7).

Emerging research suggests that improved health literacy may also benefit children. Up to 85% of children and adolescents demonstrate low health literacy (1,8,9). Better health literacy in this population has been linked to smoother transitions to adult care, better self-management, and improved communication with health care providers, especially for chronic conditions such as congenital heart disease, spina bifida, cancer, and rheumatic diseases (10–14). Addressing health literacy among children and adolescents with chronic conditions, the prevalence of which is estimated as 8% to 25% among those aged 11 to 15 years (approximately grades 5–10), is particularly important, because at this age, many begin taking on self-care responsibilities (15,16).

Various strategies have been proposed to enhance health literacy, including patient portals, visual aids, and “show-me” or “teach-back” methods (17). Writing health information at lower reading grade levels (RGLs) — or levels tailored to specific patient needs — may also improve patient comprehension (18,19). However, many of these health interventions are time-consuming and expensive, particularly given that effective health literacy interventions often need to be tailored to the individual needs of patients (20).

Recently, large language models (LLMs) and generative artificial intelligence (AI) technologies have gained attention as a way to improve health literacy by assisting with health communication and education (21–23). Despite suggestions that LLMs can tailor explanations to individual reading levels, research indicates limitations in this capability (21–23). Namely, LLMs cannot generate information at requested RGLs, and information accuracy requires further analysis before dissemination to patients (21–23). The objective of this study was to demonstrate a novel approach for generating medical information at specific RGLs that pediatricians, parents, counselors, and others can use to help explain information to children and adolescents. Although we tailored our approach for children and adolescents, similar methods can be used across the care continuum to better and readily explain information written at specific reading grade levels.

Top

Methods

We selected 200 common yet diverse pediatric conditions and symptoms across a range of medical specialties, including cardiology and vascular diseases; dental and oral health; dermatologic and craniofacial conditions, as well as plastic surgery; endocrinology; otorhinolaryngology; gastroenterology and hepatology; general pediatrics; genetics; hematology and oncology; immunology, allergy, and infectious diseases; neurology and psychiatry; obstetrics and gynecology; orthopedics and musculoskeletal conditions; pulmonology; and urology and nephrology (Appendix). We chose these conditions and symptoms while balancing prevalence, severity, and inclusion of difficult-to-explain conditions, drawing insights from the conditions and symptoms listed on the websites of Seattle Children’s Hospital and Johns Hopkins Children’s Center (24,25).

We used OpenAI’s ChatGPT application programming interface. We wrote and used 36 prompts to generate 36 outputs for each condition using zero-shot prompting (ie, providing no examples or finetuning). We used 3 general prompt architectures based on previous insights and OpenAI’s recommendations (22,23,26–28): Prompt A: “I am a []-grader and I read at the []-grade reading level. I need help understanding []”; Prompt B: “I am a physician and I need help explaining [] to a []-grader. Please provide output at the []-grade reading level”; and Prompt C: “You are a health literacy tool. Explain [] to a []-grader. Please provide output at the []-grade reading level.” We added “Please provide output at the [] reading level by using basic words and keeping sentences short” to Prompt B and Prompt C for the first- to fifth-grade levels given the external “I am a physician” and “You are a health literacy tool” nature of the prompt and previous findings (23).

For each of these 36 outputs per condition, we calculated RGLs and word counts. The RGL was calculated by averaging the Automated Readability Index, the Coleman–Liau Index, the Gunning Fog Index, and the Flesch–Kincaid Grade Level following previously established methods (22,23,26). These readability scales calculate RGLs using a combination of sentence length, word length, and syllables per word.

For each condition, we selected 1 output for each of the following complexity levels: level 1, fifth to sixth grade; level 2, seventh to eighth grade; level 3, ninth to 10th grade; and level 4, 11th to 12th grade. Outputs were chosen if they fell within 0.25 RGLs of the median for each complexity level (eg, 5.25–5.75 for level 1). If multiple outputs met the criteria, we selected the output with the highest word count. If no output matched, we selected the output closest to the target RGL. After selection, we removed all extraneous information such as, “Sure I can help with that!”, “I hope that helps!”, and “Sure, I can explain at the []-grade level.”

After removal of extraneous information, 3 full professors (D.L.S., S.A., L.M.) then reviewed outputs related to their pediatric specialties, correcting inaccuracies and adding missing information with recognition that outputs would be made available online at www.pediatricliteracy.org (28). Each of the 800 outputs was reviewed by at least 1 reviewer. We tabulated the number of words deleted and added to each output; we counted additions and deletions counted separately (eg, a single output could have 15 words added and 8 words deleted). After review, we recalculated RGLs and word counts.

In accordance with the Common Rule, this study did not require institutional review board approval because it did not involve human participants. We used Python version 3.11 (Python Software Foundation) to generate output, readability scores, and word counts, and to select outputs. We used R version 2023.12.1 (R Foundation for Statistical Computing) to perform Wilcoxon signed-rank tests; significance was set at an α of .05. We also used R to create graphs. All code is accessible online with instructions (29).

Top

Results

Prompt architectures varied in their ability to reach a specified RGL, and few outputs reached the RGL requested in the prompt (Table). The lowest median RGL of any prompt, 6.6, came from prompt 25, above the desired fifth- to sixth-grade target for complexity level 1. However, by generating 36 outputs for each condition, target RGLs, including for level 1, were achieved in most cases. The median (IQR) RGLs were 5.8 (5.4–6.4) for complexity level 1, 7.5 (7.4–7.7) for complexity level 2, 9.5 (9.4–9.7) for complexity level 3, and 11.5 (11.3–11.7) for complexity level 4, with significant differences between complexity levels (Figure 1). Word counts for selected outputs were 97.0 (75.0–166.3) for complexity level 1, 258.0 (177.0–337.3) for complexity level 2, 345.0 (286.8–395.5) for complexity level 3, and 357.0 (318.8–401.0) for complexity level 4, with significant differences between complexity levels (Figure 2).

Figure 1.
Reading grade levels for each complexity level. Boxes indicate the 25th percentile, median, and 75th percentile; dots indicate outliers; whiskers indicate 1.5 IQR. [A text version of this figure is available.]

Figure 2.
Word counts of optimized outputs for each complexity level. Boxes indicate the 25th percentile, median, and 75th percentile; dots indicate outliers; whiskers indicate 1.5 IQR. [A text version of this figure is available.]

Accuracy edits for LLM outputs were minimal, with 93.5%, 95.5%, 95.5%, and 95.0% of outputs for complexity levels 1 through 4, respectively, requiring no word additions or deletions (Figure 3). For outputs that required edits, median (IQR) word additions were 6 (2.8–11.0) for complexity level 1, 4 (1.5–5.5) for complexity level 2, 5 (1.0–9.0) for complexity level 3, and 3 (2.0–6.0) for complexity level 4. Word deletions were 8 (5.0–13.5) for complexity level 1, 10 (5.3–11.8) for complexity level 2, 4.5 (1.8–22.3) for complexity level 3, and 5 (12.5–20.8) for complexity level 4. Many changes were minor. For example, “Remember, our body is amazing and strong, and with the help of doctors, we can overcome any challenges we face, just like superheroes!” was revised to “Remember, our body is amazing and strong, and with the help of doctors, we can overcome these challenges!”

Figure 3.
Jitter plot of the number of modifications made during expert review of medical information generated by prompts submitted to ChatGPT. Dots indicate the number of words reviewers deleted from and added to selected outputs. The short, thick lines at zero indicate the outputs that required no edits. [A text version of this figure is available.]

Top

Discussion

Consistent with prior findings, we found that the LLM with zero-shot prompting (ie, providing no examples or finetuning) struggled to produce outputs at the requested RGLs, especially for content below the seventh-grade level (22,23,26). This shortcoming likely stemmed from the LLM’s training data, which may have been biased toward higher RGLs. Other models, such as Google Gemini and Microsoft Bing, often produce output at the 10th- and 11th-grade levels at baseline, respectively, while ChatGPT frequently produces outputs at the college level (22,23). Despite this, we selected ChatGPT for its superior ability to vary output RGLs (22,23).

We anticipated that our model would have difficulty generating content at the requested RGLs, particularly at lower levels. However, by leveraging both the natural variation in LLM outputs and multiple prompt architectures, we were able to extract appropriate content for each reading level. Interestingly, framing the prompt with the persona of “you are a health literacy tool” and explicitly stating “by using basic words and keeping sentences short” improved the model’s ability to generate outputs down to the fifth-to-sixth–grade level (30). However, due to difficulties in generating lower reading-level output, complexity level 1 still had the largest deviation away from the target RGL of 5.5 and had an IQR of 5.4 to 6.4. For all other complexity levels, the IQRs were 0.2 RGLs away from the target complexity level, aligning with our output selection criteria (within 0.25 RGLs from the median). In future studies and applications, generating more than 36 outputs via an application programming interface or further prompt engineering could enhance the extraction of content at lower RGLs. Although zero-shot prompting is commonplace for LLMs, our findings suggest that generating multiple outputs followed by automated selection is the most likely to reach desired RGLs.

Because target reading levels for children are poorly understood (31), we used 4 complexity levels spanning a wide range, fifth grade to 12th grade. For adults, guidelines recommend that all adult-facing health information be written below the eighth-grade level, but these guidelines and readability scales often fail to consider prior knowledge, motivation, and context (22). Although RGLs are objective, they have limitations. Hence, we grouped RGLs (eg, fifth and sixth grades) into a single complexity level and generated outputs at 4 complexity levels, allowing users to choose output aligned with their needs.

Instead of using a Likert scale or other verification scales, we assessed accuracy by having expert reviewers ensure that the outputs had no incorrect or critical missing information before publication online (29). Although our process for assessing accuracy reflects some subjectivity, we believe that entrusting experts with ensuring accuracy and completeness is adequate and more relatable to the real world than other methods. For AI applications in health care, the process of engaging experts serves as the “human-in-the-loop” (32,33).

We found that our models accurately generated information, a finding that differs from those of several studies that examined requests for diagnostic decisions (21,34,35). For the limited outputs that required changes, the changes were minor. Our findings on accuracy may differ from findings on diagnostic accuracy because LLMs are generative AI technologies that predict the next word, which may increase their ability to provide useful definitions, explanations, or simplifications of medical information — particularly for terms that are likely to be in their training data.

Regardless of prompt, the expert reviewers indicated that most outputs were accurate. This finding is reassuring because children and adolescents increasingly turn to the internet for medical advice (36). Three-quarters of adolescents report using the internet as their most recent source of medical information (37,38). However, for clinical use, experts verifying LLM outputs before sharing them with patients is essential. The ideal workflow for generating health information for children and adolescents would include pediatricians, school nurses, or other health care professionals, as well as verification of accuracy and completeness of outputs. Although we did not measure verification or review time, the minimal changes required in our study suggest that reviewing outputs is not time-intensive, especially given the typical reading speed of college-educated adults (200–400 words per minute) (39).

Although our study focused on pediatric health applications, the potential of our approach of readily generating multiple outputs and computationally selecting the desired complexity level extends to many segments of the health care continuum and to other fields. From generating patient-friendly imaging reports, discharge summaries, medication guidelines, surgical guidelines, and more, this approach may help realize the health communication goals of patient-centered medicine (21,26,28). These goals are particularly relevant in the era of the 21st Century Cures Act (40) which mandates immediate release of many types of health data (eg, imaging reports) to patients even though the jargon and complex technological terms contained in these data releases often cause patient confusion and anxiety (21,26,41). For example, if LLMs were used to simplify radiology reports for patients, they could generate reports tailored to the reading level of individual patients. However, the attitude of health care providers toward AI-generated simplifications requires further assessment (42).

Ultimately, we demonstrated that generating multiple outputs and computationally selecting ideal outputs corrected by experts in a field can improve the readability of health information for patients. Through this approach, we can empower patients with health information at the appropriate reading level, improving both patient engagement and health outcomes.

Top

Acknowledgments

The authors received no external financial support for the research, authorship, or publication of this article. No copyrighted material, surveys, instruments, or tools were used in the research described in this article. The authors declared no conflicts of interest with respect to the research, authorship, or publication of this article. K.A. reports payments from Aledade unrelated to submitted work. D.L.S. reports research grant support from the Pediatric Spine Study Group and the Safety in Spine Surgery Project. D.L.S. is a consultant for Highridge, Globus Medical, and Orthobullets. D.L.S. holds stock options in Orthobullets and has a personal investment in Green Sun Medical. D.L.S. receives royalties from Wolters Kluwer Health (textbooks), Medtronic, Highridge, and Globus Medical. D.L.S. has patents with Medtronic and Highridge and holds personal patents all unrelated to the submitted work.

Top

Author Information

Corresponding Author: Kanhai Amin, BS, Yale School of Medicine, 333 Ceder St, New Haven, CT (kanhai.amin@yale.edu).

Author Affiliations: ¹Yale School of Medicine, New Haven, Connecticut. ²Yale Child Study Center, Yale School of Medicine, New Haven, Connecticut. ³Department of Orthopedic Surgery, Cedars-Sinai Medical Center, Los Angeles, California. ⁴Department of Internal Medicine, University of Southern California, Los Angeles, California. ⁵Department of Pediatrics, Central Michigan University, Mount Pleasant, Michigan.

Top

References

1. Morrison AK, Glick A, Yin HS. Health literacy: implications for child health. Pediatr Rev. 2019;40(6):263–277. PubMed doi:10.1542/pir.2018-0027
2. US Department of Health and Human Services, Office of Disease Prevention and Health Promotion. The National Action Plan to Improve Health Literacy. 2010. Accessed September 21, 2026. https://odphp.health.gov/sites/default/files/2019-09/Health_Literacy_Action_Plan.pdf
3. Nutbeam D. The evolving concept of health literacy. Soc Sci Med. 2008;67(12):2072–2078. PubMed doi:10.1016/j.socscimed.2008.09.050
4. Ratzan SC. Health literacy: communication for the public good. Health Promot Int. 2001;16(2):207–214. PubMed doi:10.1093/heapro/16.2.207
5. Baker DW, Parker RM, Williams MV, Clark WS. Health literacy and the risk of hospital admission. J Gen Intern Med. 1998;13(12):791–798. PubMed doi:10.1046/j.1525-1497.1998.00242.x
6. Omachi TA, Sarkar U, Yelin EH, Blanc PD, Katz PP. Lower health literacy is associated with poorer health status and outcomes in chronic obstructive pulmonary disease. J Gen Intern Med. 2013;28(1):74–81. PubMed doi:10.1007/s11606-012-2177-3
7. Baker DW, Parker RM, Williams MV, Clark WS, Nurss J. The relationship of patient reading ability to self-reported health and use of health services. Am J Public Health. 1997;87(6):1027–1030. PubMed doi:10.2105/AJPH.87.6.1027
8. Lam LT. Mental health literacy and mental health status in adolescents: a population-based survey. Child Adolesc Psychiatry Ment Health. 2014;8(1):26. PubMed doi:10.1186/1753-2000-8-26
9. Vongxay V, Albers F, Thongmixay S, Thongsombath M, Broerse JEW, Sychareun V, et al. . Sexual and reproductive health literacy of school adolescents in Lao PDR. PLoS One. 2019;14(1):e0209675. PubMed doi:10.1371/journal.pone.0209675
10. Otth M, Denzler S, Koenig C, Koehler H, Scheinemann K. Transition from pediatric to adult follow-up care in childhood cancer survivors — a systematic review. J Cancer Surviv. 2021;15(1):151–162. PubMed doi:10.1007/s11764-020-00920-9
11. Rague JT, Kim S, Hirsch JA, Meyer T, Rosoklija I, Larson JE, et al. . Assessment of health literacy and self-reported readiness for transition to adult care among adolescents and young adults with spina bifida. JAMA Netw Open. 2021;4(9):e2127034. PubMed doi:10.1001/jamanetworkopen.2021.27034
12. Zhong Y, Patel N, Ferris M, Rak E. Health literacy, nutrition knowledge, and health care transition readiness in youth with chronic kidney disease or hypertension: a cross-sectional study. J Child Health Care. 2020;24(2):246–259. PubMed doi:10.1177/1367493519831493
13. Bitencourt N, Lawson E, Bridges J, Carandang K, Chintagunta E, Chiraseveenuprapund P, et al. . Pediatric to adult transition literature: scoping review and rheumatology research prioritization survey results. J Rheumatol. 2022;49(11):1201–1213. PubMed doi:10.3899/jrheum.220262
14. Amin K, Patel K. Role of psychologists in pediatric congenital heart disease. Pediatr Clin North Am. 2022;69(5):865–878. PubMed doi:10.1016/j.pcl.2022.05.002
15. Riemann L, Lubasch JS, Heep A, Ansmann L. The role of health literacy in health behavior, health service use, health outcomes, and empowerment in pediatric patients with chronic disease: a systematic review. Int J Environ Res Public Health. 2021;18(23):12464. PubMed doi:10.3390/ijerph182312464
16. Van Cleave J, Gortmaker SL, Perrin JM. Dynamics of obesity and chronic health conditions among children and youth. JAMA. 2010;303(7):623–630. PubMed doi:10.1001/jama.2010.104
17. Kountz DS. Strategies for improving low health literacy. Postgrad Med. 2009;121(5):171–177. PubMed doi:10.3810/pgm.2009.09.2065
18. Eaton ML, Holloway RL. Patient comprehension of written drug information. Am J Hosp Pharm. 1980;37(2):240–243. PubMed
19. Meade CD, Byrd JC, Lee M. Improving patient comprehension of literature on smoking. Am J Public Health. 1989;79(10):1411–1412. PubMed doi:10.2105/AJPH.79.10.1411
20. Visscher BB, Steunenberg B, Heijmans M, Hofstede JM, Devillé W, van der Heide I, et al. . Evidence on the effectiveness of health literacy interventions in the EU: a systematic review. BMC Public Health. 2018;18(1):1414. PubMed doi:10.1186/s12889-018-6331-7
21. Amin KS, Davis MA, Doshi R, Haims AH, Khosla P, Forman HP. Accuracy of ChatGPT, Google Bard, and Microsoft Bing for simplifying radiology reports. Radiology. 2023;309(2):e232561. PubMed doi:10.1148/radiol.232561
22. Amin K, Doshi R, Forman HP. Large language models as a source of health information: are they patient-centered? A longitudinal analysis. Healthcare (Amst). 2024;12(1):100731. PubMed doi:10.1016/j.hjdsi.2023.100731
23. Amin KS, Mayes LC, Khosla P, Doshi RH. Assessing the efficacy of large language models in health literacy: a comprehensive cross-sectional study. Yale J Biol Med. 2024;97(1):17–27. PubMed doi:10.59249/ZTOZ1966
24. Seattle Children’s Hospital. All conditions (A to Z). Accessed May 15, 2023. https://www.seattlechildrens.org/conditions/a-z
25. Johns Hopkins University. Pediatric conditions. Accessed May 15, 2023. https://www.hopkinsmedicine.org/johns-hopkins-childrens-center/what-we-treat/conditions
26. Doshi R, Amin KS, Khosla P, Bajaj SS, Chheang S, Forman HP. Quantitative evaluation of large language models to streamline radiology report impressions: a multimodal retrospective analysis. Radiology. 2024;310(3):e231593. PubMed doi:10.1148/radiol.231593
27. Amin KS, Forman HP, Davis MA. Even with ChatGPT, race matters. Clin Imaging. 2024;109:110113. PubMed doi:10.1016/j.clinimag.2024.110113
28. Doshi RH, Amin KS, Kapadia S, McKenzie P, Preda-Naumescu A, Tkachenko E, et al. . Characteristics of information on inflammatory skin diseases produced by four large language models. Int J Dermatol. 2025;64(4):773–775. PubMed doi:10.1111/ijd.17560
29. Pediatric Literacy.org. How to use. Accessed October 11, 2026. https://www.pediatricliteracy.org
30. OpenAI Developers. Prompt engineering. Accessed October 11, 2026. https://platform.openai.com/docs/guides/prompt-engineering
31. Sanders LM, Federico S, Klass P, Abrams MA, Dreyer B. Literacy and child health: a systematic review. Arch Pediatr Adolesc Med. 2009;163(2):131–140. PubMed doi:10.1001/archpediatrics.2008.539
32. Bakken S. AI in health: keeping the human in the loop. Oxford University Press; 2023:1225–1226.
33. Meng X-L. Data Science and Engineering with Human in the Loop, Behind the Loop, and Above the Loop. Harvard Data Sci Rev. 2023;5(2). doi:10.1162/99608f92.68a012eb
34. Barile J, Margolis A, Cason G, Kim R, Kalash S, Tchaconas A, et al. . Diagnostic accuracy of a large language model in pediatric case studies. JAMA Pediatr. 2024;178(3):313–315. PubMed doi:10.1001/jamapediatrics.2023.5750
35. Johnson D, Goodman R, Patrinely J, Stone C, Zimmerman E, Donald R, et al. . Assessing the accuracy and reliability of AI-generated medical responses: an evaluation of the Chat-GPT model. Preprint. Posted online February 28, 2023. Res Sq.
36. Pratt N, Madhavan R, Weleff J. Digital dialogue — how youth are interacting with chatbots. JAMA Pediatr. 2024;178(5):429–430. PubMed doi:10.1001/jamapediatrics.2024.0084
37. Manganello JA, Clayman ML. The association of understanding of medical statistics with health information seeking and health provider interaction in a national sample of young adults. J Health Commun. 2011;16(Suppl 3):163–176. PubMed doi:10.1080/10810730.2011.604704
38. Freeman JL, Caldwell PHY, Bennett PA, Scott KM. How adolescents search for and appraise online health information: a systematic review. J Pediatr. 2018;195:244–255.e1. PubMed doi:10.1016/j.jpeds.2017.11.031
39. Rayner K, Schotter ER, Masson ME, Potter MC, Treiman R. So much to read, so little time: how do we read, and can speed reading help? Psychol Sci Public Interest. 2016;17(1):4–34. PubMed doi:10.1177/1529100615623267
40. 21st Century Cures Act, Pub L No. 114-255, 130 Stat 1033 (2016). https://www.govinfo.gov/app/details/PLAW-114publ255
41. Amin KS, Davis MA, Naderi A, Forman HP. Increasing patient viewership of complex imaging reports: the paradox of the Cures Act. Clin Imaging. 2025;119:110398. PubMed doi:10.1016/j.clinimag.2024.110398
42. Amin KS, Davis MA, Naderi A, Forman HP. Release of complex imaging reports to patients, do radiologists trust AI to help? Curr Probl Diagn Radiol. 2025;54(2):147–150. PubMed doi:10.1067/j.cpradiol.2024.12.008

Top

Table

Abbreviation: RGL, reading grade level.
^a Calculated by averaging the Automated Readability Index, the Coleman-Liau Index, the Gunning Fog Index, and the Flesch–Kincaid Grade Level following previously established methods (22,23,26).
^b For each condition, we selected 1 output for each of the following complexity levels: level 1, fifth to sixth grade; level 2, seventh to eighth grade; level 3, ninth to 10th grade; and level 4, 11th to 12th grade. Table denotes number of times the prompt’s output was used when selecting 1 output per complexity level and condition/symptom.

Top

Appendix. Categories for 200 Conditions and Symptoms Commonly Found in Pediatric Populations in the US^a

^a Conditions and symptoms were selected by balancing prevalence, severity, and inclusion of difficult-to-explain conditions and by referring to the websites of Seattle Children’s Hospital and Johns Hopkins Children’s Center (24,25).

Top