Patient-centric radiology: Utilising large language models (LLMs) to improve patient communication and education

This randomized controlled trial demonstrates that using large language models to simplify radiology reports significantly improves patient readability and comprehension with minimal hallucination rates, thereby enhancing patient-clinician communication and meeting strong patient demand for accessible medical information.

The Gist

Imagine you go to the doctor, get an X-ray or a scan, and then you log into your online health portal to see the results. You open the file, but instead of a clear explanation, you are hit with a wall of confusing medical jargon, complex sentences, and scary-sounding words. It's like trying to read a manual for a spaceship when you just wanted to know how to fix your toaster.

This is the problem a new study from Australia set out to solve. The researchers asked a simple question: Can Artificial Intelligence (AI) act as a translator to turn those scary, complex medical reports into plain English that anyone can understand?

Here is the story of their experiment, explained simply.

The Problem: The "Doctor-Speak" Barrier

Radiology reports are written for doctors, not patients. They are packed with technical terms that require a university-level education to decipher.

• The Reality: Most people read at a middle-school level.
• The Report: Most radiology reports are written at a university level.
• The Result: Patients get anxious, confused, or misinterpret their own health news because they can't understand what the report actually says.

The Solution: The "AI Translator"

The researchers used a smart AI tool (called a Large Language Model, or LLM, similar to the technology behind advanced chatbots) to act as a translator. They took the original, complex reports and asked the AI to rewrite them in simple, everyday language—like explaining a medical issue to a smart 10-year-old.

They ran a test with 120 patients:

1. Group A got the original, complex report.
2. Group B got the original report plus the new, AI-simplified version.

What Happened? (The Results)

The results were like night and day.

1. The "Reading Level" Test
Think of the original report as a book written for a PhD student. The AI version was rewritten for a 5th or 6th grader.

• Before: Patients struggled to read the original.
• After: 97% of patients said the AI version was "easy" or "very easy" to read.

2. The "Understanding" Test
This was the big win.

• Without the AI: Only about half the patients felt they understood their report.
• With the AI: 95% of patients said they understood more than half of their report.
• Analogy: It's the difference between being handed a map written in a foreign language versus a map with clear, simple arrows pointing to your destination.

3. The "Safety Check" (Did the AI lie?)
The biggest fear with AI is "hallucination"—when the AI makes up facts or gets things wrong. The researchers had two expert radiologists check the AI's work.

• Accuracy: The AI was almost perfect.
• Hallucinations: They found very few errors (only about 1.7%), and crucially, none of the errors were dangerous. The AI didn't tell anyone they were sick when they were healthy, or vice versa. It was safe to use.

4. What Patients Wanted
Almost everyone (98%) said, "Yes, please give us the simple version next time!" They felt empowered and less anxious when they could actually read what was happening to their bodies.

Why This Matters

We are moving toward a future where patients can see their medical records instantly online. But if we just dump complex, scary reports on people without explaining them, it causes more stress than help.

This study shows that AI can be the bridge between the doctor's technical world and the patient's everyday life. It doesn't replace the doctor; it just translates the message so the patient can hear it clearly.

The Bottom Line

Imagine if every time you got a medical report, you also got a "Plain English" summary written by a super-smart, careful translator. That is what this study suggests is possible. It makes healthcare less scary, less confusing, and much more human.

In short: AI can turn a "medical mystery novel" into a clear, easy-to-read story, helping patients understand their health without needing a medical degree.

Technical Summary

1. Problem Statement

The study addresses the critical gap between the complexity of standard radiology reports and the health literacy levels of the general public.

• The Barrier: Radiology reports are written for clinicians, utilizing complex medical jargon and structured at a reading level equivalent to the 12th–13th grade. Conversely, the average reading level of the public is significantly lower (approx. 8th grade in the US, 10th grade in Australia).
• The Consequence: This discrepancy leads to patient confusion, increased anxiety, misinterpretation of findings, and unnecessary follow-up inquiries.
• The Context: Recent legislative changes in Australia (the Modernising My Health Record – Sharing by Default Bill 2024) mandate that radiology reports be shared immediately with patients via "My Health Record" (MHR). This increases the urgency for making these documents accessible without compromising diagnostic accuracy.
• Current Limitations: Existing strategies (e.g., structured reporting, summaries) are often impractical in high-volume clinical settings due to time constraints and the risk of radiologist burnout.

2. Methodology

The researchers conducted a single-centre, randomised controlled trial (RCT) at Barwon Health, Victoria, Australia, between May and June 2025.

• Participants: 120 outpatient patients (aged 18+) undergoing X-ray, ultrasound, or CT scans.
  • Exclusion Criteria: Inpatients, those with medical training, non-English speakers, new diagnoses of sinister pathology (e.g., cancer), and obstetric ultrasounds.
• Study Arms (1:1 Randomisation):
  1. Control Group: Received the formal radiology report only.
  2. Intervention Group: Received the formal report plus an LLM-simplified version.
• Intervention (LLM Implementation):
  • Model: Microsoft 365 CoPilot (based on GPT-4o architecture).
  • Prompting: A single zero-shot prompt was used to simplify de-identified reports to a 6th-grade reading level.
  • Security: The LLM was hosted locally within the institutional IT infrastructure to ensure data privacy; no patient data left the secure network.
• Data Collection & Metrics:
  • Subjective Metrics: Patient surveys using 5-point Likert scales to assess perceived readability ("easy" to "very difficult") and comprehension levels (percentage of report understood).
  • Objective Metrics: Readability scores calculated via four standard indices:
    • Simple Measure of Gobbledygook (SMOG)
    • Automated Readability Index (ARI)
    • Flesch Reading Ease
    • Flesch-Kincaid Grade Level
  • Accuracy Assessment: Two consultant radiologists independently reviewed the LLM outputs for factual accuracy (5-point Likert scale) and hallucination (binary scale: present/absent, and potential for harm).
• Statistical Analysis: Mann-Whitney U tests for ordinal data, paired t-tests for readability indices, and Chi-squared ($\chi^2$) tests for categorical accuracy data. Significance was set at $p < 0.05$.

3. Key Contributions

• Real-World Validation: Unlike previous studies using fictitious reports, this study evaluated LLM simplification on actual, individualized patient reports in a prospective clinical setting.
• Security-First AI Integration: Demonstrated a workflow where LLMs are hosted locally to maintain strict data privacy compliance, addressing a major barrier to clinical AI adoption.
• Zero-Shot Efficiency: Showed that a single, detailed zero-shot prompt (without few-shot examples) is sufficient to achieve high accuracy and low hallucination rates, suggesting a low-barrier implementation strategy for clinical workflows.
• Balancing Nuance and Simplicity: Provided evidence that reports can be simplified to a patient-friendly level without losing essential clinical nuance or diagnostic accuracy.

4. Key Results

• Readability Improvement:
  • Objective: Formal reports required an 11th–13th grade reading level. LLM-simplified reports were reduced to a 5th–9th grade level ($p < 0.001$).
  • Subjective: 97% of patients found the LLM reports "easy" or "very easy" to read, compared to only 28% for formal reports.
• Comprehension:
  • 95% of patients in the intervention group reported understanding >50% of their report, compared to only 46% in the control group ($p < 0.001$).
• Accuracy and Safety:
  • Factual Accuracy: 97–100% of simplified reports were rated as "accurate" or "very accurate" by radiologists.
  • Hallucinations: The hallucination rate was minimal at 1.7% (2 instances total across the cohort).
  • Harm Assessment: Crucially, zero hallucinations were deemed potentially harmful to the patient (no false reassurance or induced anxiety regarding sinister pathology).
• Patient Preference: 98.3% of participants expressed a desire to receive simplified reports alongside formal ones in future practice.

5. Significance and Conclusion

The study provides robust evidence that Large Language Models (LLMs) can effectively bridge the health literacy gap in radiology.

• Clinical Impact: LLMs offer a scalable solution to improve patient understanding and reduce anxiety without significantly increasing radiologist workload.
• Policy Relevance: The findings support the implementation of simplified reports in the context of Australia's new "sharing by default" legislation, ensuring patients can access their records without being overwhelmed by technical jargon.
• Safety: While hallucinations remain an inherent risk of generative AI, the study demonstrates that with local hosting and human oversight, the risk of patient harm is negligible.
• Future Directions: The authors recommend integrating these tools into patient portals (like MHR) while continuing to refine privacy frameworks and establish standardized thresholds for acceptable hallucination rates in clinical settings.

Limitations: The study was single-centre, limited to X-ray/US/CT (excluding MRI/Nuclear Medicine), and relied on self-reported patient data which may be subject to response bias.