Abstract
AbstractBackgroundThe increasing need for diagnostic echocardiography (echo) tests presents challenges in preserving the quality and promptness of reports. While Large Language Models (LLMs) have proven effective in summarizing clinical texts, their application in echo remains underexplored. To address this, we proposed EchoGPT, a dedicated, domain-specific LLM focused on echo report summarization.MethodsAdult echo studies conducted at the Mayo Clinic from January 1, 2017, to December 31, 2017, were collected and categorized into two groups: development (all Mayo locations except Arizona) and external validation (Mayo Arizona). We adapted open-source LLMs (Llama-2, MedAlpaca, Zephyr, and Flan-T5) using In-Context Learning (ICL) and Quantized Low-Rank Adaptation (QLoRA) fine-tuning for echo text summarization. The models’ performance was assessed both quantitatively with automatic metrics and qualitatively by cardiologists.ResultsThe development dataset included 97,506 reports from 71,717 unique patients, predominantly male (54.3%), with an average age of 64.1±16.1 years. The final split contains 95,506 for training, and 1,000 each for validation and testing. EchoGPT, a QLoRA fine-tuned Llama-2 model, outperformed other LLMs with about 90% win rates in various metrics (BLEU, METEOR, ROUGE-L, BERT Score, and RadGraph F1 Score), and produced reports comparable to cardiologists in 30 randomly selected cases for qualitative human review (significantly preferred in conciseness (p< 0.001), with no significant preference in completeness, correctness, and clinical utility). In the external validation set (n=1,000), EchoGPT consistently outperformed fine-tuned Zephyr model across the same automatic metrics (all p < 0.0001).ConclusionsCapable of generating echocardiography reports on par with human experts, EchoGPT could be used to generate draft reports for human review and approval, with significant workflow advantages.Clinical PerspectiveWhat is new?● This study is the first attempt to compare multiple open-source LLMs and different model adaptation methods in echocardiography report summarization.● The resulting system, EchoGPT, can generate echo reports comparable in quality to cardiologists.● Future metrics for echo report quality should emphasize factual correctness, especially on numerical measurements.What are the clinical implications?● EchoGPT system demonstrated the potential of introducing LLMs into echocardiography practice.● EchoGPT could be used as an AI co-pilot to generate echo reports.
Publisher
Cold Spring Harbor Laboratory