{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,9,7]],"date-time":"2026-09-07T06:08:20Z","timestamp":1788761300966,"version":"build-2803163510"},"reference-count":28,"publisher":"Frontiers Media SA","license":[{"start":{"date-parts":[[2026,9,7]],"date-time":"2026-09-07T00:00:00Z","timestamp":1788739200000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"content-domain":{"domain":["frontiersin.org"],"crossmark-restriction":true},"short-container-title":["Front. Digit. Health"],"abstract":"<jats:sec>\n                    <jats:title>Introduction<\/jats:title>\n                    <jats:p>Radiologists in resource-limited settings often face high workloads, especially in chest X-ray interpretation. Manual annotation of large-scale imaging datasets remains costly and time-consuming. This study aims to explore the feasibility of using large language models (LLMs), specifically GPT-4o, to generate binary disease presence labels from free-text radiology reports, and to use these labels to train deep learning models for automated chest X-ray classification.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Methods<\/jats:title>\n                    <jats:p>A two-stage supervised learning pipeline was developed using the publicly available MIMIC-CXR v2.1.0 dataset. First, GPT-4o was prompted with a structured clinical protocol to classify each radiology report as either \u201cdiseased\u201d or \u201cno disease.\u201d Second, the generated labels were used to supervise the training of four convolutional neural networks: ResNet-18, DenseNet-121, EfficientNet-B1, and ConvNeXt-Tiny. A patient-level 70\/10\/20 split was employed to prevent data leakage across sets. Each model was trained across five random seeds (42\u201346), and 95% confidence intervals were computed using the t-distribution. Label quality was evaluated by comparing 210 generated labels against radiologist annotations from a board-certified radiologist.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Results<\/jats:title>\n                    <jats:p>\n                      GPT-4o achieved an overall accuracy of 92.9% with expert labels on the 210-report validation set. For the \u201cdiseased\u201d class, the precision was 97.4% and recall was 90.5%; for \u201cno disease,\u201d precision was 87.1% and recall was 96.4%. Among the CNN models evaluated on the held-out test set, ConvNeXt-Tiny achieved the highest area under the curve (\n                      <jats:inline-formula>\n                        <mml:math xmlns:mml=\"http:\/\/www.w3.org\/1998\/Math\/MathML\" id=\"IM1\">\n                          <mml:mrow>\n                            <mml:mi mathvariant=\"normal\">AUC<\/mml:mi>\n                          <\/mml:mrow>\n                          <mml:mo>=<\/mml:mo>\n                          <mml:mn>0.832<\/mml:mn>\n                        <\/mml:math>\n                      <\/jats:inline-formula>\n                      , 95% CI [0.801, 0.863]) and balanced accuracy (0.739), significantly outperforming EfficientNet-B1 (\n                      <jats:inline-formula>\n                        <mml:math xmlns:mml=\"http:\/\/www.w3.org\/1998\/Math\/MathML\" id=\"IM2\">\n                          <mml:mrow>\n                            <mml:mi mathvariant=\"normal\">AUC<\/mml:mi>\n                          <\/mml:mrow>\n                          <mml:mo>=<\/mml:mo>\n                          <mml:mn>0.797<\/mml:mn>\n                        <\/mml:math>\n                      <\/jats:inline-formula>\n                      ; paired\n                      <jats:inline-formula>\n                        <mml:math xmlns:mml=\"http:\/\/www.w3.org\/1998\/Math\/MathML\" id=\"IM3\">\n                          <mml:mi>t<\/mml:mi>\n                        <\/mml:math>\n                      <\/jats:inline-formula>\n                      -test,\n                      <jats:inline-formula>\n                        <mml:math xmlns:mml=\"http:\/\/www.w3.org\/1998\/Math\/MathML\" id=\"IM4\">\n                          <mml:mi>p<\/mml:mi>\n                          <mml:mo>=<\/mml:mo>\n                          <mml:mn>0.014<\/mml:mn>\n                        <\/mml:math>\n                      <\/jats:inline-formula>\n                      ). ResNet-18 (\n                      <jats:inline-formula>\n                        <mml:math xmlns:mml=\"http:\/\/www.w3.org\/1998\/Math\/MathML\" id=\"IM5\">\n                          <mml:mrow>\n                            <mml:mi mathvariant=\"normal\">AUC<\/mml:mi>\n                          <\/mml:mrow>\n                          <mml:mo>=<\/mml:mo>\n                          <mml:mn>0.822<\/mml:mn>\n                        <\/mml:math>\n                      <\/jats:inline-formula>\n                      ) and DenseNet-121 (\n                      <jats:inline-formula>\n                        <mml:math xmlns:mml=\"http:\/\/www.w3.org\/1998\/Math\/MathML\" id=\"IM6\">\n                          <mml:mrow>\n                            <mml:mi mathvariant=\"normal\">AUC<\/mml:mi>\n                          <\/mml:mrow>\n                          <mml:mo>=<\/mml:mo>\n                          <mml:mn>0.808<\/mml:mn>\n                        <\/mml:math>\n                      <\/jats:inline-formula>\n                      ) showed intermediate performance. All models demonstrated AUC values above 0.79, confirming the viability of LLM-derived weak supervision.\n                    <\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Discussion<\/jats:title>\n                    <jats:p>This study demonstrates that LLMs can be effectively employed to generate supervision labels for medical imaging tasks. The proposed approach offers a scalable and low-cost solution for preliminary disease screening, particularly in healthcare environments with limited expert availability. The multi-seed evaluation with confidence intervals provides a rigorous assessment of model stability. Further work is needed to improve label reliability and expand to multi-label classification.<\/jats:p>\n                  <\/jats:sec>","DOI":"10.3389\/fdgth.2026.1883337","type":"journal-article","created":{"date-parts":[[2026,9,7]],"date-time":"2026-09-07T05:34:16Z","timestamp":1788759256000},"update-policy":"https:\/\/doi.org\/10.3389\/crossmark-policy","source":"Crossref","is-referenced-by-count":0,"title":["Automated chest X-ray disease screening using large language models and deep convolutional neural networks on the MIMIC-CXR dataset"],"prefix":"10.3389","volume":"8","author":[{"given":"Qingyu","family":"Zhang","sequence":"first","affiliation":[{"name":"Institute of Health Informatics, University College London","place":["London, United Kingdom"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Pardeep","family":"Vasudev","sequence":"additional","affiliation":[{"name":"Institute of Health Informatics, University College London","place":["London, United Kingdom"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Kezhi","family":"Li","sequence":"additional","affiliation":[{"name":"Institute of Health Informatics, University College London","place":["London, United Kingdom"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Holger","family":"Kunz","sequence":"additional","affiliation":[{"name":"Institute of Health Informatics, University College London","place":["London, United Kingdom"]},{"name":"Department of Primary Care and Public Health, Imperial College London","place":["London, United Kingdom"]}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"1965","published-online":{"date-parts":[[2026,9,7]]},"reference":[{"key":"B1","doi-asserted-by":"publisher","first-page":"923","DOI":"10.1016\/j.aej.2022.10.053","article-title":"A deep learning architecture for multi-class lung diseases classification using chest X-ray (CXR) images","volume":"64","author":"Alshmrani","year":"2023","journal-title":"Alex Eng J"},{"key":"B2","author":"Sidume","year":""},{"key":"B3","doi-asserted-by":"publisher","first-page":"102125","DOI":"10.1016\/j.media.2021.102125","article-title":"Deep learning for chest X-ray analysis: A survey","volume":"72","author":"\u00c7all\u0131","year":"2021","journal-title":"Med Image Anal"},{"key":"B4","doi-asserted-by":"publisher","first-page":"101562","DOI":"10.1016\/j.media.2019.101562","article-title":"Pre and post-hoc diagnosis and interpretation of malignancy from breast DCE-MRI","volume":"58","author":"Maicas","year":"2019","journal-title":"Med Image Anal"},{"key":"B5","doi-asserted-by":"publisher","first-page":"12495","DOI":"10.1038\/s41598-019-48995-4","article-title":"Deep learning to improve breast cancer detection on screening mammography","volume":"9","author":"Shen","year":"2019","journal-title":"Sci Rep"},{"key":"B6","doi-asserted-by":"publisher","first-page":"H390","DOI":"10.1152\/ajpheart.00416.2020","article-title":"Deep learning-based automated left ventricular ejection fraction assessment using 2-D echocardiography","volume":"321","author":"Liu","year":"2021","journal-title":"Am J Physiol Heart Circ Physiol"},{"key":"B7","doi-asserted-by":"publisher","first-page":"17485","DOI":"10.1038\/s41598-021-96707-8","article-title":"AI outperformed every dermatologist in dermoscopic melanoma diagnosis, using an optimized deep-CNN architecture with custom mini-batch logic and loss function","volume":"11","author":"Pham","year":"2021","journal-title":"Sci Rep"},{"key":"B8","doi-asserted-by":"publisher","first-page":"1836","DOI":"10.1093\/annonc\/mdy166","article-title":"Man against machine: diagnostic performance of a deep learning convolutional neural network for dermoscopic melanoma recognition in comparison to 58 dermatologists","volume":"29","author":"Haenssle","year":"2018","journal-title":"Ann Oncol"},{"key":"B9","author":"Rajpurkar","year":""},{"key":"B10","doi-asserted-by":"publisher","first-page":"3320","DOI":"10.48550\/arXiv.1411.1792","article-title":"How transferable are features in deep neural networks?","volume":"27","author":"Yosinski","year":"2014","journal-title":"Adv. Neural Inf. Process. Syst"},{"key":"B11","author":"Baltruschat","year":""},{"key":"B12","doi-asserted-by":"publisher","first-page":"5998","DOI":"10.48550\/arXiv.1706.03762","article-title":"Attention is all you need","volume":"30","author":"Vaswani","year":"2017","journal-title":"Adv. Neural Inf. Process. Syst."},{"key":"B13","author":"Devlin","year":""},{"key":"B14","doi-asserted-by":"publisher","first-page":"1877","DOI":"10.48550\/arXiv.2005.14165","article-title":"Language models are few-shot learners","volume":"33","author":"Brown","year":"2020","journal-title":"Adv. Neural Inf. Process. Syst"},{"key":"B15","author":"Touvron","year":""},{"key":"B16","doi-asserted-by":"publisher","first-page":"255","DOI":"10.1002\/hcs2.61","article-title":"Large language models in health care: Development, applications, and challenges","volume":"2","author":"Yang","year":"2023","journal-title":"Health Care Sci"},{"key":"B17","doi-asserted-by":"publisher","first-page":"319","DOI":"10.1001\/jama.2024.21700","article-title":"Testing and evaluation of health care applications of large language models: a systematic review","volume":"333","author":"Bedi","year":"2025","journal-title":"JAMA"},{"key":"B18","doi-asserted-by":"publisher","first-page":"286","DOI":"10.3390\/diagnostics13020286","article-title":"Natural language processing applications for computer-aided diagnosis in oncology","volume":"13","author":"Li","year":"2023","journal-title":"Diagnostics"},{"key":"B19","doi-asserted-by":"publisher","first-page":"8839524","DOI":"10.1155\/2020\/8839524","article-title":"Text messaging-based medical diagnosis using natural language processing and fuzzy logic","volume":"2020","author":"Omoregbe","year":"2020","journal-title":"J Healthc Eng"},{"key":"B20","doi-asserted-by":"publisher","first-page":"e0307741","DOI":"10.1371\/journal.pone.0307741","article-title":"GPT-4 as an X data annotator: Unraveling its performance on a stance classification task","volume":"19","author":"Liyanage","year":"2024","journal-title":"PLoS ONE"},{"key":"B21","author":"Irvin","year":""},{"key":"B22","author":"Atf","year":""},{"key":"B23","author":"He","year":""},{"key":"B24","author":"Huang","year":""},{"key":"B25","author":"Liu","year":""},{"key":"B26","author":"Tan","year":""},{"key":"B27","doi-asserted-by":"publisher","first-page":"317","DOI":"10.1038\/s41597-019-0322-0","article-title":"MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports","volume":"6","author":"Johnson","year":"2019","journal-title":"Sci Data"},{"key":"B28","doi-asserted-by":"publisher","first-page":"e076865","DOI":"10.1136\/bmjopen-2023-076865","article-title":"Large language model-based information extraction from free-text radiology reports: a scoping review protocol","volume":"13","author":"Reichenpfader","year":"2023","journal-title":"BMJ Open"}],"container-title":["Frontiers in Digital Health"],"original-title":[],"link":[{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/fdgth.2026.1883337\/full","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,9,7]],"date-time":"2026-09-07T05:34:17Z","timestamp":1788759257000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/fdgth.2026.1883337\/full"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,9,7]]},"references-count":28,"alternative-id":["10.3389\/fdgth.2026.1883337"],"URL":"https:\/\/doi.org\/10.3389\/fdgth.2026.1883337","relation":{},"ISSN":["2673-253X"],"issn-type":[{"value":"2673-253X","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,9,7]]},"article-number":"1883337"}}