Introduction
Hospital discharge letters are a cornerstone of patient safety, serving as the primary source of instructions for post-hospital care. However, these documents are typically written by clinicians for other clinicians, resulting in a "paternalistic" style filled with dense medical jargon and abbreviations. This communication barrier is a significant factor in patient non-adherence; nearly 50% of patients with chronic diseases fail to follow their prescribed therapies, often because they do not understand the "why" or "how" of their care. While patient-centered letters are known to improve outcomes, the time required to draft them manually is a major hurdle for busy medical staff. This study explores whether Large Language Models (LLMs), specifically GPT-4, can automate the transformation of clinical summaries into accessible, actionable patient guides.
Methods
Researchers conducted a prospective, exploratory study using a "gold standard" approach to evaluate GPT-4's performance. They developed three high-fidelity synthetic discharge letters covering common conditions: arterial hypertension (AHT), type 2 diabetes (DM), and diabetic kidney disease (DKD).
To measure the effectiveness of the transformation, the team defined 72 specific "learning objectives" (LOs)—safety-relevant pieces of information essential for patient care. These LOs were categorized using Bloom’s Revised Taxonomy into two cognitive levels: Remember (factual recall, such as medication names) and Understand (comprehending the underlying reasoning, such as why a medication is necessary for kidney protection).
The transformation process utilized a multi-stage prompt chaining pipeline rather than a single prompt. This included data extraction into structured JSON, general summary generation, action point extraction, and adversarial prompting specifically designed to simplify remaining jargon to an 8th-grade reading level. Finally, two experienced clinicians reviewed 860 generated sentences for medical accuracy and patient-centricity.
Key Results
GPT-4 demonstrated a strong ability to improve the accessibility of medical information, though some safety gaps remained:
- Information Transfer: The AI successfully captured 78% (56/72) of the safety-relevant learning objectives. However, 15% (11/72) were missed in the majority of generated letters.
- The Cognitive Gap: Analysis revealed that GPT-4 was significantly better at transferring "Remember" level objectives than "Understand" level ones. Most omissions occurred in the "prevention of complications" field, where the AI struggled with complex, multi-dependency logical structures.
- Medical Accuracy: 88% of generated sentences were medically correct. Interestingly, 71% of the errors identified were "hallucinations"—information the AI added itself—rather than misinterpretations of the source text.
- Readability and Jargon: The transformation resulted in a massive leap in readability. Flesch Reading Ease scores moved from "difficult" (~40) to "standard" (~60). Medical terms were reduced from 60% to 15% of the text, and 92% of the remaining terms were accompanied by explanations.
Discussion and Conclusion
GPT-4 shows immense potential for transforming clinical communication into patient-centered narratives. It excels at "soft" communication—addressing the patient directly and simplifying complex terminology. However, the 15% omission rate for safety-critical information indicates that the technology is not yet ready for unsupervised clinical use.
Future improvements may lie in advanced prompt engineering, such as "Chain-of-Thought" prompting to help the AI better link actions to medical reasoning. For now, the researchers suggest a "human-in-the-loop" model where clinicians use AI-generated drafts as a starting point, performing a final safety check before the letters reach the patient. This approach could significantly reduce the administrative burden on doctors while substantially improving patient understanding and safety.
Summary updated