Impact of Documentation Aids on the Inclusion of Salient Left-Ventricular Late Gadolinium Enhancement Characteristics in Cardiac Magnetic Resonance Imaging Impressions

Abstract Late gadolinium enhancement (LGE) on cardiac magnetic resonance (CMR) imaging is a key diagnostic and prognostic biomarker in ischemic cardiomyopathy (ICM) and nonischemic cardiomyopathy (NICM). However, the completeness of LGE reporting in routine CMR impressions remains incompletely characterized. We aimed to evaluate the impact of large language model (LLM)-assisted reporting on the consistency of LGE pattern, distribution, and severity documentation in CMR impressions for ICM and NICM. We analyzed 828 unstructured (1990–2023) and 400 structured (2024–2025) CMR reports to identify documentation of LGE pattern, distribution, and severity. A Llama 3 language model fine-tuned on 27,000 structured CMR reports was used to generate 9650 CMR impressions; LGE attributes were extracted using an LLM judge. LLM reporting behavior was further assessed using quantized low-rank adaptation adapters trained on structured and unstructured impressions. Structured reporting captured LGE pattern more frequently than unstructured reporting overall (85.5% vs. 61.0%, p = 9.60 × 10 −7 ) and for ICM (90.0% vs. 69.4%, p = 0.005). LLM-generated impressions demonstrated more frequent reporting of LGE distribution (75.3%) and severity (70.9%) versus structured and unstructured reports ( p < 0.001). Fine-tuning with an adapter trained on unstructured impressions significantly improved reporting of all three LGE attributes. Structured reporting improved LGE pattern documentation, but gaps in LGE distribution and severity documentation persisted across human report formats. LLM-generated impressions achieved broader attribute coverage for LGE distribution and severity. Fine-tuning on unstructured data improved model performance, suggesting that passive augmentation with data capturing the documentation styles of multiple providers may help to address reporting gaps.