Lunit, a medical artificial intelligence (AI) company, announced on Wednesday that it has successfully completed the first phase of clinical validation for the AI-Specialized Foundation Model project, a national strategic initiative by the Ministry of Science and Information and Communications Technology (ICT). The company is now advancing to the second phase, which involves nationwide validation and model refinement.
Lunit is spearheading a consortium for the project titled, Development of a Multi-Scale Medical Science Specialized Foundation Model for Innovation from Molecules to Populations, which brings together 23 industry, academic, research, and medical institutions.
During this initial phase, the Lunit consortium developed a lightweight medical science specialized foundation model with 16 billion parameters entirely in-house. This model strikes a balance between practical applicability in real medical settings and cost efficiency.
The first phase of training leveraged an extensive array of specialized data encompassing all areas of healthcare, including research papers, clinical records, drug information, and treatment guidelines.
This effort resulted in the creation of a search-augmented generation (RAG) feature that automatically retrieves and presents relevant evidence, alongside an agent system that links multiple specialized models to tackle complex problems systematically.
Despite its lightweight design, the model exhibited world-class performance in the medical field. It met or exceeded target performance across five key evaluation metrics, including medical literature comprehension and evidence-based responses, significantly outperforming large language models from other companies.
The most significant achievement was demonstrating the model’s real-world applicability during clinical validation tests conducted at the National Health Insurance Ilsan Hospital.
The Lunit consortium implemented the initial clinical decision support system (CDSS) model in analyzing adverse drug reactions (ADR), classifying emergency room patients, and supporting initial diagnoses.
They confirmed that the Lunit foundation model can alleviate pharmacists’ workload during ADR analysis and improve drug safety management efficiency. Moreover, it proved capable of swiftly supporting initial assessments by medical staff in emergency rooms and facilitating rapid treatment decisions.
Looking ahead, the Lunit consortium plans to broaden the validation scope in phase two. This expansion will involve collaborations with nine medical institutions nationwide, including Yongin Severance Hospital, Korea University Medical Center, Konyang University Hospital, and Yangsan Busan National University Hospital, as well as two pharmaceutical companies.
Lunit Chief Executive Officer (CEO) Seo Beom-seok stated that the specialized foundation model developed by the Lunit consortium serves as a cornerstone infrastructure for medical science, with applications extending to clinical decision-making, drug discovery, and public health. Having validated the model’s real-world applicability in phase one, it is committed to driving tangible improvements in healthcare through further model enhancement and expanded validation in phase two.