Topic Extraction of Research Seeds by Natural Language Processing From Programs of Annual Scientific Meetings of the Japanese Circulation Society Authors: Itaru Hosaka, Yukinori Akiyama, Marenao Tanaka, Tatsuya Sato, Masato Furuhashi Published: 2026 ABSTRACT BackgroundManually tracking research trends in extensive conference programs is challenging, so we used a natural language processing approach to automatically extract trending topics from PDF-formatted programs of the Japanese Circulation Society (JCS).Methods and ResultsPrograms from JCS2023 to JCS2026 were analyzed by GiNZA and Latent Dirichlet Allocation. Among the 42,400 extracted text blocks, there were 8 primary research themes, including sustained interests in coronary artery syndrome and heart failure, an increase in interprofessional collaboration and emerging clusters in arrhythmia and valvular intervention.ConclusionsThe automated workflow successfully visualized evolving academic trends, providing a robust tool for comprehensive research exploration. Central Figure Attending annual academic meetings is not only an opportunity to present one’s own research findings but also to explore potential future research topics and current trends. However, during a long-running meeting with numerous sessions, there is a limit to the number of presentations each individual can physically attend, with the attendant risk of missing topics of interest. Here, we report our approach to automatically extracting information on trending research topics from conference programs distributed in portable document format (PDF) using morphological analysis based on natural language processing (NLP). Methods Programs of consecutive annual meetings of the Japanese Circulation Society (JCS) between 2023 and 2026 were downloaded from websites and analyzed. The program of JCS2022 was excluded because of the different structure during the COVID-19 pandemic. The analysis workflow had 3 steps: (1) text extraction on a block-by-block basis from the PDF file of each year’s program; (2) tokenization and vectorization for nouns with morphological analysis by GiNZA [https://github.com/megagonlabs/ginza]; and (3) word analysis, including topic modeling and co-occurrence network, using Latent Dirichlet Allocation (LDA).1 In the text extraction phase, stop words were set to exclude various noisy terms, including “Alena”, “Messe”, “center” and “luncheon”. Japanese words in the presentation title were manually translated into English. All analysis and visualization were performed by Python 3 (ver. 3.10). Results A total of 42,400 text blocks were extracted from 4 PDF files: JCS2023, JCS2024, JCS2025 and JCS2026. After applying 139 stop words, the tokenizer found 142,298 words from the blocks. We set the number of topics at 8 for training the LDA model, and the top 5 representative words were highlighted (Figure A). The 8 topics were different research themes, comprising topic 0: acute coronary syndrome (ACS); topic 1: heart failure; topic 2: transcatheter aortic valve replacement (TAVR); topic 3: arterial fibrillation and ablation; topic 4: treatment of HF (written in Japanese); topic 5: interprofessional collaboration (written in Japanese); topic 6: ACS and valve disease; and topic 7: pulmonary hypertension (PH). Topics 4 and 5 involved medical staff sessions because topic terms were written in Japanese. Topic 6 mainly contained sessions of case reports. The main trends in recent years were research topics on ACS and heart failure, and during the same period, the topic of TAVR declined slightly, while studies focused on interprofessional collaboration increased. The number of research topics on PH increased in 2024 but decreased thereafter. Figure.Research trends in the Japanese Circulation Society (JCS) meetings. (A) Trend stream plot of research topics from 2023 to 2026. (B) Co-occurrence network plot of research keywords in JCS2026. The green nodes with the suffix “ja” represent Japanese words that were manually translated into English. Research trends in the Japanese Circulation Society (JCS) meetings. (A) Trend stream plot of research topics from 2023 to 2026. (B) Co-occurrence network plot of research keywords in JCS2026. The green nodes with the suffix “ja” represent Japanese words that were manually translated into English. The co-occurrence network unveiled various distinctive clusters of research themes in JCS2026 (Figure B). The size of each node represents the number of research, and the width of edge represents the strength of the relationship between nodes. Clusters located at the outer edges of the diagram indicate higher novelty compared to previous years. The largest bundle formed in the center of diagram, and was enriched with terms of heart failure related to ischemia, arrhythmia, myopathy and thromboembolism. Of note, it was remarkable that the size of arrhythmia-related words involving pulse field ablation and left appendage was comparable to that of words related to heart failure. In the lower left of the diagram, terms related to TAVR, mitral regurgitation (MR) and tricuspid regurgitation (TR) clustered together, suggesting the impact of new devices for MR and TR. Along the edges of the diagram, several interesting terms, including “adipose tissue”, “limb ischemia”, “pulmonary vein isolation” and “amyloidosis”, were sparsely scattered. Discussion Topic modeling and co-occurrence network were constructed by NLP using PDF files of annual meeting programs from the JCS. A key advantage of this method is the ability to extract distinctive topics for each year by processing multi-year programs simultaneously, leading to comprehensive exploration of trending research topics. IRB Information The present study was granted an exemption from requiring ethics approval by Sapporo Medical University because it did not involve human subjects. Disclosures M.F. is a member of Circulation Reports’ Editorial Team.