Fine-tuned large language models can generate expert-level echocardiography reports

Trans-thoracic echocardiograms (TTEs) are one of the most requested non-invasive cardiac imaging tests worldwide.1,2 It consists of 30–60 videos that physicians interpret and synthesize into comprehensive reports. These reports typically include detailed ‘Findings’ sections, which list all observations and measurements, and concise ‘Impressions’ sections that summarize the most clinically relevant information. This process is not only time-consuming but also requires significant expertise to ensure accuracy and clinical relevance.3 The increasing demand for TTEs coupled with long waiting times4 has created a pressing need for more efficient reporting methods. Artificial intelligence (AI), particularly large language models (LLMs), offers a promising solution to partially automate and streamline the trans-thoracic echocardiograms report creation, potentially enhancing both the speed and consistency of echocardiography reporting. In the present issue of the European Heart Journal – Digital Health, Chao et al.5 report their findings on the development and evaluation of EchoGPT, a LLM fine-tuned for echocardiography report summarization. They trained EchoGPT, based on the LLaMA-2-7 Billion model6 on a substantial dataset of 97 506 echocardiogram reports from Mayo Clinic to generate the ‘Impressions’ section, from the ‘Findings’ section. Due to patient privacy policy regulations, EchoGPT was compared exclusively against open-source baseline models. These included two general-purpose LLMs: Llama-2-7b-chat6 and Zephyr-7b,7 one medical-specific LLM: Med-Alpaca8 and a sequence-to-sequence model: Flan-T5. The study employed several common natural language processing (NLP) metrics, each providing a different perspective on the quality of the generated text. BLEU (Bilingual Evaluation Understudy) measures n-gram precision, METEOR (Metric for Evaluation of Translation with Explicit ORdering) considers synonyms and word order, ROUGE-L (Recall-Oriented Understudy for Gisting Evaluation—Longest Common Subsequence) focuses on the longest common subsequence, and BERT (Bidirectional encoder representation from transformers) Score evaluates semantic similarity using contextual embeddings. The RadGraph-F1 metric,9 a medical-specific measure for assessing the accuracy of structured data representation in medical text, was used as the primary evaluation criterion for ranking different models based on the factual correctness of the generated clinical content. These metrics range from 0 to 100, with higher scores indicating better performance. The results showed that EchoGPT achieved significant improvements over baseline models across all metrics: BLEU (45.9 ± 18.9, P < 0.0001) suggesting moderate similarity to reference texts, METEOR (62.4 ± 18.0, P < 0.0001) indicating good capture of meaning and structure, ROUGE-L (55.7 ± 17.8, P < 0.0001) demonstrating moderate to good summarization quality, BERT Score (91.6 ± 3.0, P < 0.0001) revealing high semantic alignment with human-written reports, and RadGraph-F19 (47.7 ± 14.9, P < 0.0001), a medical specific NLP metric, revealing moderate success in accurately extracting and identifying medical entities from the original reports. EchoGPT was compared against human expert evaluation. A panel of cardiologists evaluated the generated summaries of 30 randomly selected cases based on four criteria (4C): completeness, measuring whether all relevant details were included; conciseness, evaluating clarity and brevity; correctness, assessing accuracy; and clinical utility, determining usefulness for clinical decision-making. Scores range from 0 to 1, with higher scores indicating better performance. The results were as follows: correctness (0.17) suggesting slight agreement, conciseness (0.22), clinical utility (0.34) indicating fair agreement, and completeness (0.49) revealing moderate agreement. Additionally, cardiologist compared original and generated reports on the 4C metrics with scores ranging from −10 to 10, where positive values indicate EchoGPT outperformed human experts. EchoGPT significantly outperformed human experts on conciseness (1.67 ± 5.40, P < 0.001), and showed no significant differences in the other criteria. To assess the sensitivity of automatic metrics to clinically significant errors, the researchers replaced actual measurements in the reports with random numbers. Surprisingly, this manipulation, which rendered the reports medically meaningless, only slightly lowered the automatic metric scores (P < 0.0001). This demonstrates that common NLP metrics are inadequate for assessing the clinical relevance and potential impact of errors in medical reporting. Chao et al. have made a significant contribution to the field of AI in cardiology by developing and evaluating EchoGPT, an LLM fine-tuned for TTE report summarization. Their study demonstrates that EchoGPT not only outperforms other open-source LLMs in this task but also produces reports comparable to those written by cardiologists in terms of completeness, correctness, conciseness, and clinical utility. Notably, EchoGPT’s outputs were significantly preferred for their conciseness, suggesting potential improvements in reporting efficiency. Furthermore, the authors are to be commended for their commitment to open science and research reproducibility. While patient privacy concerns prevent the release of the primary Mayo Clinic dataset, the publicly available MIMIC-EchoNotes10 dataset provides a valuable resource for other researchers to conduct comparative studies and validate findings. Moreover, by releasing a checkpoint of their models, and statistical analysis code on GitHub (https://github.com/chiehjuchao/EchoGPT.git), the authors have set a commendable example of transparency in AI research. This approach not only enhances the credibility of their work but also facilitates further advancements in the field by enabling other researchers to build upon and potentially improve their model. Such openness is crucial for the responsible development and evaluation of AI in medical imaging, striking a balance between data privacy and scientific collaboration.11 Although EchoGPT has demonstrated promising results in summarizing echocardiography reports, this study highlights the need for more robust and scalable evaluation methods tailored specifically for medical AI applications. The medical-specific metric RadGraph-F19 proved more suitable than traditional NLP metrics as it was the only automatic metric that exhibited moderate correlation with human expert evaluation. Key areas for improvement include enhancing factual accuracy, minimizing hallucinations, improving generalizability across diverse healthcare settings, integrating it with current echocardiography reporting tools, and validating its impact. Additionally, integrating patient history and other relevant clinical data could result in more contextualized and clinically meaningful reports. These advancements, along with rigorous clinical validation and continuous human expert oversight, have the potential to revolutionize echocardiography reporting. However, it is crucial that these developments prioritize patient safety, clinical accuracy, and ethical considerations in AI-assisted medical decision-making. Robert Avram is supported by the ‘Fonds de la recherche en santé du Québec’ (Grant 312758), by the Canadian Institute for Advanced Research (CIFAR), the Montreal Heart Institute Research Centre, the Montreal Heart Institute Foundation and by the Des Groseillers-Bérard Research Chair.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.