Effective communication in healthcare hinges not only on clinical accuracy but also on the ability to convey information in a manner aligned with patients’ literacy levels and comprehension needs^5,6. Our results demonstrate that LLAMA3-MedReadCtrl meaningfully advances this goal by enabling large language models to dynamically control text readability while preserving content quality and relevance—a capability critical for patient education, informed consent, and shared decision-making^2,3. Compared to leading proprietary models such as GPT-4 and Claude-3, as well as the baseline LLAMA3 model, MedReadCtrl achieves superior readability control across both medical and general domains (Figure 1), with particularly strong performance on clinical datasets where simplification must be balanced with medical precision. Automatic metrics (Table 1) further support this: MedReadCtrl consistently yields higher ROUGE, BLEU, and SARI scores across diverse tasks and unseen datasets, demonstrating both robustness and generalizability.

Human evaluations add an important layer of interpretability. At lower grade levels—especially Grades 2 and 5, where health literacy challenges are most acute—MedReadCtrl outperforms GPT-4 in clarity, fluency, and clinical accuracy (Table 2, Figure 2). These improvements are more than stylistic; they translate directly into better comprehension for patients who may otherwise misinterpret discharge instructions, medication labels, or follow-up plans²³. Notably, MedReadCtrl maintains semantic fidelity even when simplifying technical content, avoiding jargon while preserving intent—an essential property for equitable, patient-centered care delivery⁹.

At lower readability levels, the ability to convey biomedical information in accessible and age-appropriate language is essential for reducing comprehension disparities among low-literacy populations, including children, older adults, and non-native speakers^21,24,25. These groups face heightened risk of misunderstanding key health information, which can contribute to poor treatment adherence, medication errors, or delayed care-seeking behavior. In this context, LLAMA3-MedReadCtrl demonstrates strong alignment with health literacy guidelines by consistently avoiding complex or technical terminology and producing intuitive, patient-friendly explanations. For instance, in the Grade 2 example describing an X-ray procedure (Table 7), GPT-4 generates: “Taking an X-ray picture of the spinal cord after putting special dye into a space around it.” While structurally simplified, it retains specialized vocabulary such as “X-ray” and “dye,” which may be unfamiliar to early readers. In contrast, MedReadCtrl rewrites this sentence as: “A special picture of the spine that helps doctors see inside,” eliminating jargon while preserving the medical intent. This form of semantic simplification aligns with plain language principles and supports comprehension among young patients or caregivers with limited background knowledge¹². A similar pattern appears in the case describing a cough with sputum and blood streaks. GPT-4’s Grade 2 output—”She has a cough with some phlegm and a little bit of blood sometimes, but no big blood”—includes opaque terms like “phlegm” and uses awkward quantifiers that disrupt fluency. MedReadCtrl instead produces: “She has a cough and sometimes brings up yucky stuff from her lungs. Sometimes it might have a little bit of red in it, but it’s not too much,” employing concrete, relatable language that mirrors health educator strategies in pediatric and low-literacy settings. Another compelling example involves anatomical description at Grade 5. GPT-4’s sentence—”She does have some soreness in her groin on both sides”—retains the anatomically correct but potentially confusing term “groin.” MedReadCtrl transforms this into: “She does have some pain in the area where her legs and hips meet, on both sides,” offering a precise yet accessible explanation that would better support health comprehension in community education portals, post-visit instructions, or telehealth consultations⁹. These adaptations reflect the model’s capacity to restructure content in a way that prioritizes understanding over literal term retention—mirroring strategies used by professional health educators and offering critical value for patient-centered communication in post-visit instructions, pediatric care, and health literacy tools²⁶.

At higher readability levels (Grades 8 and 11), MedReadCtrl continues to offer value through more context-sensitive and humanized phrasing. For example, in the Grade 8 simplification of “depression treatment drug” (Table 8), GPT-4 generates the generic phrase “a medicine for treating depression,” which is technically accurate but semantically flat. In contrast, MedReadCtrl rewrites this as: “A medicine that helps people feel better when they are sad,” offering a more empathetic and humanized description. Similarly, in a Grade 11 paraphrase of erectile dysfunction, GPT-4 states: “The inability of a man to get an erection because of mental issues or problems with his body.” MedReadCtrl’s version—”A condition in men where they are unable to get or maintain an erection, often due to a physical or mental health issue”—uses more clinical but naturalistic phrasing and avoids stigmatizing language, which may be especially important for sensitive topics. This level of semantic elaboration can support communication not only with patients but also with caregivers, educators, and multidisciplinary teams involved in long-term condition management^27,28. A particularly illustrative demonstration of MedReadCtrl’s readability control occurs in the context of follow-up care planning. GPT-4 generates the same sentence—”I will check back in three months”—across all readability levels, showing no sensitivity to audience comprehension needs. In contrast, MedReadCtrl scales the expression appropriately: for Grade 2, it outputs “I will see you again in three months,” using simple, conversational phrasing. By Grade 11, the model elaborates to: “I will schedule a follow-up appointment with you in approximately three months to assess the status of your condition and determine any necessary adjustments to your treatment plan.” This progression illustrates the model’s ability to integrate both linguistic complexity and communicative intent—an essential capability for generating documentation templates and educational materials tailored to adolescents, caregivers, or patients managing long-term conditions⁹.

Taken together, these examples underscore MedReadCtrl’s strength in generating audience-appropriate, semantically faithful text across a range of literacy levels. The model goes beyond surface simplification by restructuring sentence syntax, reframing technical terms, and adjusting tone—all while maintaining clinical accuracy. These capabilities are essential for patient-facing technologies, such as discharge instruction portals, caregiver education tools, and AI-powered chatbots, which must adapt content to diverse comprehension needs in real-world settings^5,6,9.

Despite these strengths, several limitations remain. First, like other large language models, MedReadCtrl is susceptible to factual hallucinations. As shown in Table 9 (Case 1), the model occasionally introduces subtle inaccuracies that may mislead clinical interpretation, such as incorrectly attributing throat dryness to a symptom rather than to a condition. In another example, the phrase “increasing chest pain episodes” is paraphrased as “he gets hurt in his chest a lot,” which incorrectly frames angina-like symptoms as physical injury. While seemingly minor, such distortions can erode trust or lead to serious miscommunication in diagnostic explanations and treatment decisions, especially in patient self-management or telemedicine settings. Integration with retrieval-augmented generation (RAG) frameworks²⁹ or structured biomedical knowledge sources (e.g., medical dictionaries or biomedical knowledge graphs) may help reduce the risk of hallucination and improve factual grounding. Additionally, incorporating fact-checking frameworks³⁰ to validate outputs against trusted references before generation may improve accuracy in sensitive clinical scenarios.

Second, we observe that readability control sometimes becomes conflated with verbosity, particularly at higher readability levels. In Case 2, the model’s Grade 11 explanation of Sjögren’s syndrome becomes overly technical, introducing complex constructs such as “autoimmune destruction of salivary glands,” along with redundant phrases like “the second form is secondary.” These choices can overwhelm target readers, including adolescents or adult learners. Future iterations could incorporate multidimensional control signals—encoding not only readability grade but also length, tone, or reader type—within the prompt structure³¹. Reinforcement learning with human feedback (RLHF)³², especially when combined with preference models that penalize redundant or overly medicalized outputs, may further refine generation. Incorporating examples with graded prompts (“explain to a high schooler in 2 sentences”) during fine-tuning may improve precision in stylistic control. In future iterations, human-annotated readability and stylistic preference healthcare data can inform reward modeling under RLHF frameworks, enabling more precise alignment with user expectations and literacy needs.

A third challenge concerns the occasional misinterpretation of instruction. As illustrated in Case 3, when presented with the input “Description: A sample note on Rheumatoid Arthritis,” the model incorrectly generates a definition rather than simplifying the content. This indicates a failure to differentiate between content-bearing inputs and meta-instructions, a weakness exacerbated in zero- or low-shot generalization scenarios. Mitigating this issue may require explicit instruction boundary encoding or task-specific tokens that delimit transformations versus descriptions. Fine-tuning on synthetic instruction-following data with varied structure, along with task-grounded prompt engineering strategies^33,34, may also improve adherence.

Additionally, discrepancies between expert judgment and automated LLM-as-Judge assessments raise concerns about current evaluation methods. Although MedReadCtrl received a 71.6% preference score from human clinicians, it received only a 60.0% preference score from LLM-based judges. This divergence suggests that existing automatic evaluators may undercapture clinical relevance, semantic fidelity, and subtle differences in readability optimization. For instance, in a Grade 5 simplification task, clinicians preferred MedReadCtrl’s empathetic phrasing—”pain in the area where the legs and hips meet”—as it demonstrated improved contextual sensitivity and accessibility for lay audiences. In contrast, the LLM-as-Judge assigned a higher score to GPT-4’s output—”groin pain”—which, while medically accurate, retained potentially unfamiliar terminology. This illustrates the current evaluators’ blind spots when judging language adapted for low-literacy or sensitive contexts. Particularly in domains such as pediatric or mental health communication, black-box metrics may be insufficient. Moreover, limitations such as evaluation bias, domain contamination, and interpretability issues have been documented in prior work^35–37. Future benchmarks should consider hybrid evaluation pipelines that combine expert-derived annotations (e.g., health literacy compliance checklists or comprehension probes) with LLM-based assessments, enabling more nuanced and trustworthy evaluations.

In sum, MedReadCtrl demonstrates that fine-grained readability control is both technically feasible and clinically valuable. By enabling personalized content generation tailored to varying health literacy levels, the system addresses longstanding gaps in patient-centered communication and equitable access to medical information⁹. Realizing its full potential will require continued attention to hallucination mitigation, stylistic calibration, task clarity, and robust evaluation frameworks. As large language models are increasingly deployed in clinical education tools, caregiver interfaces, and remote health portals, systems like MedReadCtrl can serve as foundational infrastructure for safe, comprehensible, and inclusive medical AI communication³⁸. To ensure trustworthy and scalable deployment in real-world settings, future work should also consider model transparency, regulatory compliance, and human-in-the-loop oversight strategies^39,40. Such efforts can not only improve system reliability but also inform the development of policy and governance frameworks for responsible LLM integration into patient-facing applications⁴¹.

While our findings establish MedReadCtrl’s effectiveness in producing semantically faithful, audience-appropriate outputs across diverse readability levels, several limitations warrant consideration. First, all experiments were conducted on English-language datasets, limiting the generalizability of our findings to non-English settings. Extending this framework to support multilingual or culturally adapted readability control remains an important direction for future research. Second, our study focused on three NLG tasks—text simplification, paraphrase generation, and semantic entailment generation—which, while representative, do not encompass the full range of document styles and discourse types seen in clinical communication (e.g., patient-provider dialogues⁴²). Further work is needed to explore how readability control operates across broader clinical tasks and note types. Additionally, our use of traditional readability metrics, while widely adopted in the literature, may be insufficient to assess fluency and understandability in fragmented or telegraphic clinical texts, such as EHRs. Alternative approaches—such as domain-adapted, machine learning–based readability scoring—may provide more robust assessment. Lastly, while our human evaluation included expert scoring across multiple grade levels, it did not incorporate real users (e.g., patients or caregivers) from the target literacy groups. Future work should include end-user evaluation to assess practical effectiveness and refine feedback loops in human-centered settings. These limitations notwithstanding, our work provides a scalable and adaptable foundation for personalized medical text generation, helping bridge the persistent health literacy gap in clinical care delivery.