Large generative models for radiology report interpretation: assessing feasibility and patient safety

Abstract

With the digitalization of healthcare, patients have gained access to their own medical records. However, poor clarity of medical texts often prevents them from interpreting it correctly. Large generative models (LGMs) have the potential to become a tool for adapting medical texts, but their use is currently fraught with risks. The aim of the study was to evaluate the safety of using LGMs to interpret radiology protocols for patients. Seven models performed a simplified text interpretation using eight computed tomography protocols as input. The generated interpretations were submitted to physicians and respondents without medical training for evaluation. The resulting scores were analyzed to determine the safety of implementing this technology and its feasibility.All models generated text that met the main quality criteria. However, consistent ethical and safety violations were observed. A comparative analysis failed to identify a model that was superior across all criteria. The study also identified criteria for which assessments by physicians and respondents without medical training differed significantly.It was demonstrated that, although large-scale generative models are formally successful in simplified interpretation of medical protocols, their direct application without a control system in clinical practice is extremely unsafe. The main problem is the distortion of the original information—the inclusion of additional recommendations, diagnoses, and prognoses, which contravenes patient communication standards. It was shown that despite the technology's potential within the field, safe implementation requires the preliminary development of a quality control system for large-scale generative models, a questionnaire that takes into account the competencies of both experts and non-experts, and clear threshold criteria. This work represents the first step toward creating such a system.

Ähnliche Arbeiten