AI improves patient understanding of complex CT reports

Technical medical reports often contain terminology that is difficult for many patients to interpret. A team of researchers at the Technical University of Munich has now investigated how artificial intelligence (AI) can help make CT reports easier to understand. Their results show that AI-generated simplifications improve readability and increase patient satisfaction.

To simplify original CT documents, the team used an open-source large language model that ran exclusively on the TUM University Hospital’s internal systems in compliance with data protection regulations. One example of the transformation is:
“The cardiomediastinal silhouette is midline. [...] A small pericardial effusion is noted”
which the AI simplified to:
“Heart: The report notes a small amount of fluid around your heart. This is a common finding, and your doctor will determine if it needs any attention.”

This approach aims to help patients better understand their diagnostic information during cancer treatment.

For the researchers, clear communication is essential. Felix Busch states: “Ensuring that patients understand their reports, examinations, and treatments is a central pillar of modern medicine. This is the only way to guarantee informed consent and strengthen health literacy.”

To evaluate the real-world impact, 200 patients undergoing CT imaging for cancer were included. One half received the original reports, while the other received AI-simplified versions.

The outcome was clear: reading time for original reports averaged seven minutes, while simplified texts required only two. Patients with simplified versions found them easier to read (81%) and easier to understand (80%). They also rated the simplified texts as more helpful (82%) and more informative (82%).
“Various objective measurements also confirmed the improved readability of the simplified reports,” adds Felix Busch.

Despite the advantages, the team cautions against using general chatbots to interpret medical documents. Dr. Philipp Prucker notes: “Aside from data protection concerns, language models always carry the risk of factual errors.” In the study, 6% of AI-generated results contained inaccuracies, 7% omitted information, and 3% added details not in the original.

All simplified texts were reviewed and corrected before being given to patients. Prucker concludes: “Language models are useful tools, but they are no substitute for medical staff. Without trained specialists verifying the findings, patients may, in the worst case, receive incorrect information about their illness.”

MEDICA-tradefair.com; Source: Technical University of Munich