From complexity to clarity: GPT-4 for patient-centered prostate cancer pathology reports

To evaluate whether GPT-4-turbo can generate accurate, patient-centered prostate cancer pathology reports at or below a 6th-grade reading level using a validated reporting template.

We retrieved 44 prostate cancer pathology reports from The Cancer Genome Atlas database. We used twenty reports to iteratively refine our prompt refinement and tested the final prompts on 24 unseen reports. GPT-4-turbo generated two patient-centered versions for each report, one below and one above the 6th-grade reading level. We assessed readability using the TextEvaluator tool, which measures eight educational text-complexity dimensions aligned with Common Core standards. We performed paired t-tests to compare the original reports to both simplified versions.

GPT-4 significantly reduced the overall text complexity of the pathology reports (p < 0.001). The below 6th-grade versions showed the most improvement. These versions reduced academic vocabulary (mean 68.47-23.16), simplified syntactic structures (p = 0.0018), and used more concrete language (36.65-44.12). Our iterative prompt engineering eliminated hallucinations and ensured clinical accuracy.

GPT-4-turbo, when guided by a well-designed prompt and validated template, can produce accurate, patient-accessible prostate cancer pathology summaries.

This approach could improve health communication, particularly for patients with limited health literacy, and offers a low-cost, scalable solution for integrating PCPRs into clinical workflows with minimal burden on clinicians. This workflow may improve patient comprehension of cancer diagnoses, enhance shared decision-making, and promote more equitable access to understandable medical information without substantial additional resource demands.