TLDR: BioASQ 2025, the thirteenth challenge, advanced large-scale biomedical semantic indexing and question answering. It featured updated versions of Task b and Synergy, plus four new tasks: multilingual clinical summarization (MultiClinSum), nested named entity linking in Russian and English (BioNNE-L), clinical coding in cardiology (ELCardioCC), and gut-brain interplay information extraction (GutBrainIE). With 83 teams and over 1000 submissions, the challenge showcased significant progress in applying AI, especially LLMs and BERT-based models, to complex biomedical information needs across multiple languages and document types.
The BioASQ challenge, a leading international initiative, recently concluded its thirteenth edition in 2025, continuing its mission to drive advancements in large-scale biomedical semantic indexing and question answering. This year’s challenge, held in the context of the Conference and Labs of the Evaluation Forum (CLEF) 2025, brought together 83 competing teams who submitted over 1000 distinct solutions across six diverse tasks. The event highlighted the continuous evolution of state-of-the-art techniques in the field, with many participating systems achieving competitive performance.
Established Tasks See Renewed Focus
The 2025 BioASQ challenge revisited two of its well-established tasks, Task b and Synergy, introducing new editions that reflected current biomedical information needs. Task 13b focused on biomedical semantic question answering, unfolding in three phases: retrieving relevant material (Phase A), providing ‘exact’ and ‘ideal’ answers (Phase A+), and generating new answers based on expert-selected material (Phase B). This year saw a significant increase in participation for Task 13b, with 46 teams and 734 submissions, underscoring its ongoing relevance. Systems leveraged a range of techniques, from traditional document retrieval methods like BM25 to advanced Retrieval-Augmented Generation (RAG) frameworks utilizing Large Language Models (LLMs) such as Llama, Gemma, GPT, Claude, and Mistral.
Task Synergy 13 continued its unique approach of fostering research in developing biomedical topics, such as infectious, rare, and genetic diseases, and women’s and reproductive health. Designed as an ongoing dialogue, experts posed open-ended questions, providing feedback on retrieved material and answers iteratively. Five teams participated, primarily employing LLMs like DeepSeek-R1 and Llama, enhanced with RAG frameworks, optimized prompting, and Named Entity Recognition (NER).
Four New Frontiers in Biomedical AI
A significant highlight of BioASQ 2025 was the introduction of four novel tasks, each addressing critical unmet needs in biomedical information processing:
MultiClinSum: Multilingual Clinical Summarization
This new task tackled the challenge of automatically summarizing lengthy clinical case reports written in English, Spanish, French, and Portuguese. With a corpus of over 1,280 manually selected full-text and summary pairs in English, and hundreds more in other languages (further augmented by neural machine translation), MultiClinSum aimed to evaluate methods for condensing clinical documents while retaining key insights. 11 teams submitted runs, with the English sub-track seeing the highest participation and best results, though other languages were also well-represented.
BioNNE-L: Nested Named Entity Linking in Russian and English
BioNNE-L focused on the complex task of mapping medical entities, including nested structures, to concepts from the UMLS metathesaurus. This challenge was unique in its focus on both Russian and English data, with subtasks for monolingual and bilingual tracks. Seven teams participated, with top-performing systems predominantly utilizing biomedical BERT-based retrieval and re-ranking architectures like SapBERT and BERGAMOT, highlighting the importance of domain-specific methods.
ELCardioCC: Clinical Coding in Cardiology
Addressing the critical need for automated clinical coding, ELCardioCC focused on assigning cardiology-related ICD-10 codes to discharge letters from Greek hospitals and extracting specific mentions of these codes. Five teams engaged in subtasks covering Named Entity Recognition (NER), Entity Linking (EL), and Multi-label Classification – Explainable AI (MLC-X). Most approaches leveraged transformer-based models, particularly BERT variants and multilingual LLMs, adapted for Greek medical texts.
GutBrainIE: Gut-Brain Interplay Information Extraction
This task aimed to foster the development of Information Extraction (IE) systems to automatically extract and link knowledge from biomedical abstracts concerning the gut-brain interplay and its role in mental health and neurological diseases. It comprised four subtasks of increasing difficulty: NER, Binary Tag-based Relation Extraction (BT-RE), Ternary Tag-based Relation Extraction (TT-RE), and Ternary Mention-based Relation Extraction (TM-RE). Sixteen teams participated in NER, with strong performance from supervised fine-tuning and transformer-based models. Relation extraction subtasks proved more challenging, demonstrating the complexity of simultaneously locating and labeling entities and identifying their relationships.
Also Read:
- Advancing Medical AI: A Survey of Reasoning Capabilities in Large Language Models
- AI-Generated Data Enhances Accuracy in Health Fact-Checking
Looking Ahead
The thirteenth BioASQ challenge showcased significant progress in biomedical AI, with increased participation and competitive performance across all tasks. The introduction of new tasks in multilingual summarization, nested entity linking, clinical coding, and gut-brain information extraction expanded the challenge’s scope to six languages (English, Spanish, French, Portuguese, Russian, and Greek) and diverse document types (biomedical articles, clinical case reports, and discharge letters). The findings underscore the growing capability of AI systems to address complex biomedical information needs, while also highlighting areas for continued research and development, particularly in specialized domains and cross-lingual applications. For more detailed information, you can refer to the Overview of BioASQ 2025 research paper.
