Large language model-based simplification of breast imaging reports: a prospective multicentre study with 2965 patient-based evaluations.
Authors
Affiliations (8)
Affiliations (8)
- Postgraduate School in Radiodiagnostics, Università degli Studi di Milano, Milan, Italy.
- Radiology and Interventional Radiology, Fondazione IRCCS Ca' Granda Ospedale Maggiore Policlinico, Milan, Italy.
- Institute of Integrated Diagnostics of Southern Switzerland (IDISI), Ente Ospedaliero Cantonale (EOC), Lugano, Switzerland.
- Department of Biomedical, Surgical and Dental Sciences, Università degli Studi di Milano, Milan, Italy. [email protected].
- Breast Imaging Division, Radiology Department, IEO European Institute of Oncology IRCCS, Milan, Italy.
- UOSD Senologia Ospedaliera, Ospedale A. Perrino, Brindisi, Italy.
- Radiology Department, Ospedale di Stato-Istituto per la Sicurezza Sociale, Cailungo, San Marino.
- Department of Oncology and Haemato-Oncology, Università degli Studi di Milano, Milan, Italy.
Abstract
To compare patient perceptions of radiologist-written and artificial intelligence (AI)-generated simplifications of breast imaging reports in terms of simplicity, comprehensibility, empathy, and overall preference. In this prospective multicentre survey-based study, 10 anonymized breast imaging reports (5 ultrasound, 5 mammography; one BI-RADS category 1-5 per modality) were simplified by ChatGPT-4 within a 50-word limit. After assessment by a panel of 20 physicians, paired radiologist-written and AI-generated reports were evaluated by 314 patients (2965 evaluations) who rated simplicity, comprehensibility, and empathy on a 5-point Likert scale and indicated their preferred report. The origin of each report was concealed, and each pair was presented in randomized order. Univariable and multivariable mixed-effects logistic regression analyses were performed to identify factors associated with preferring AI-generated reports. AI-generated reports were preferred in 63.3% (1876/2965) of evaluations. High simplicity and comprehensibility scores (4-5) were more frequently assigned to AI reports than to radiologist-written reports (70.0% vs 43.4% and 68.1% vs 40.4%, respectively). In multivariable analysis, higher comprehensibility was the strongest predictor in both directions: higher comprehensibility of AI-generated reports increased the odds of preferring AI (score 4: adjusted OR 1.87, 95% CI 1.08-3.23, p = 0.025), while higher comprehensibility of radiologist-written reports reduced them (score 4: adjusted OR 0.45, 95% CI 0.29-0.69, p < 0.001). Preference for AI-generated reports showed a non-significant trend toward increase with educational level. AI-generated breast imaging reports were preferred by most patients and perceived as clearer, supporting a complementary and supervised use of AI-generated summaries. Question Patients increasingly access their own radiology reports, yet standard radiologist-written reports are often difficult to understand, highlighting the need for patient-friendly alternatives. Findings In this study, AI-generated breast imaging reports were preferred by patients in 63.3% of evaluations, with comprehensibility emerging as the strongest predictor of preference. AI-generated breast imaging reports were preferred by most patients and received higher ratings for simplicity and comprehensibility than radiologist-written reports, supporting a supervised and complementary use of AI-generated summaries in clinical breast imaging.