This study presents a comprehensive topic modeling analysis of scientific abstracts in the field of artificial intelligence (AI) applied to dentistry. We compiled and analyzed 3,170 peer-reviewed abstracts published between 2019 and 2025 from the Dimensions and Scopus databases. Three complementary approaches were compared: (1) Structural Topic Modeling (STM), a probabilistic framework incorporating publication year as a covariate to enable temporal trend analysis; (2) embedding-based clustering using the Leiden algorithm on OpenAI text embeddings; and (3) zero-shot GPT-based topic modeling, in which GPT-4o generated topics, descriptions, and keywords directly from batches of abstracts without model training. Topic quality was evaluated using compactness, distinctiveness, silhouette scores, and label redundancy. STM consistently produced the most compact and well-separated topics, embedding-based clustering excelled in identifying discrete semantic groupings, and GPT-based modeling provided interpretable, human-readable labels but exhibited greater thematic overlap. To ensure comparability across methods, we introduced a two-layer alignment framework that integrates topic-level similarity with document-level consensus, enabling robust cross-model comparison. Using this framework, we identified stable topics consistently recovered across methods (e.g., caries detection, radiographic AI diagnostics) as well as method-specific themes. Temporal trend analysis in this shared space revealed a clear shift from foundational AI methods (e.g., image segmentation, image enhancement) toward applied and integrative areas, including large language model applications, patient-facing tools, and AI in clinical education. Our results underscore the value of combining classical probabilistic models with modern large language model (LLM) tools for optimal topic modeling performance. While GPT-4o enhances interpretability, it should not be used in isolation for mapping thematic structures in scientific literature, at least not without pre-screening and prompt experimentation. Overall, our findings demonstrate the importance of hybrid topic modeling for mapping thematic structures in fast-evolving scientific domains, with dentistry serving as a case study.
This study presents a comprehensive topic modeling analysis of scientific abstracts in the field of artificial intelligence (AI) applied to dentistry. We compiled and analyzed 3,170 peer-reviewed abstracts published between 2019 and 2025 from the Dimensions and Scopus databases. Three complementary approaches were compared: (1) Structural Topic Modeling (STM), a probabilistic framework incorporating publication year as a covariate to enable temporal trend analysis; (2) embedding-based clustering using the Leiden algorithm on OpenAI text embeddings; and (3) zero-shot GPT-based topic modeling, in which GPT-4o generated topics, descriptions, and keywords directly from batches of abstracts without model training. Topic quality was evaluated using compactness, distinctiveness, silhouette scores, and label redundancy. STM consistently produced the most compact and well-separated topics, embedding-based clustering excelled in identifying discrete semantic groupings, and GPT-based modeling provided interpretable, human-readable labels but exhibited greater thematic overlap. To ensure comparability across methods, we introduced a two-layer alignment framework that integrates topic-level similarity with document-level consensus, enabling robust cross-model comparison. Using this framework, we identified stable topics consistently recovered across methods (e.g., caries detection, radiographic AI diagnostics) as well as method-specific themes. Temporal trend analysis in this shared space revealed a clear shift from foundational AI methods (e.g., image segmentation, image enhancement) toward applied and integrative areas, including large language model applications, patient-facing tools, and AI in clinical education. Our results underscore the value of combining classical probabilistic models with modern large language model (LLM) tools for optimal topic modeling performance. While GPT-4o enhances interpretability, it should not be used in isolation for mapping thematic structures in scientific literature, at least not without pre-screening and prompt experimentation. Overall, our findings demonstrate the importance of hybrid topic modeling for mapping thematic structures in fast-evolving scientific domains, with dentistry serving as a case study.
The study is supported by Russian Science Foundation, grant number 23-15-00060.