Lunit and KAIST Specialized Foundation Models Pass Interim Evaluation
Full-Scale Real-World Testing Begins in Hospitals and Pharmaceutical Settings

Domestic medical and bio artificial intelligence (AI) is rapidly advancing to a level where it can be applied in real-world settings, from classifying and assisting in the diagnosis of emergency room patients to predicting new drug candidates.


It is expected that in hospitals, these technologies will help physicians make diagnostic and treatment decisions faster and more accurately, while in the pharmaceutical sector, they could significantly reduce the time required for new drug development. As a result, the pace of medical innovation felt by the public is also expected to accelerate.

Conceptual Diagram of Clinical Decision Support Agent System (CDSS). ① Construction of a medical science foundation model (FM) trained on core medical knowledge such as medical papers, clinical trials, drugs, and real clinical data ② Application of retrieval-augmented generation (RAG) function that searches and references medical guidelines, papers, and drug information in real time ③ Advancement into a clinical specialized model tailored to actual hospital treatment environments to support diagnosis and treatment decisions. Provided by the Ministry of Science and ICT

Conceptual Diagram of Clinical Decision Support Agent System (CDSS). ① Construction of a medical science foundation model (FM) trained on core medical knowledge such as medical papers, clinical trials, drugs, and real clinical data ② Application of retrieval-augmented generation (RAG) function that searches and references medical guidelines, papers, and drug information in real time ③ Advancement into a clinical specialized model tailored to actual hospital treatment environments to support diagnosis and treatment decisions. Provided by the Ministry of Science and ICT

View original image

The Ministry of Science and ICT announced on March 31 that, based on the interim evaluation results of the 'AI Specialized Foundation Model' project, both the Lunit consortium and the Korea Advanced Institute of Science and Technology (KAIST) consortium have met the criteria for the second phase of support. Accordingly, both teams will continue to receive support for 256 B200 graphic processing units (GPUs) until September, accelerating the advancement of AI models specialized in the medical and bio fields.


The most notable aspect of this achievement is its applicability in the medical field. The Lunit consortium has built a 'Clinical Decision Support System (CDSS)' based on a medical science foundation model trained on medical papers and clinical data.


This system demonstrated high accuracy in the five-level triage of emergency room patients using real medical data from the National Health Insurance Service Ilsan Hospital, and achieved a diagnosis match rate of approximately 94%. It also delivered excellent performance in identifying adverse drug reactions and generating reports, and was evaluated as being at a level that healthcare professionals can use in practice.


In particular, in time-critical environments such as emergency rooms, the system can rapidly present physicians with a simultaneous review of a patient's condition, the latest guidelines, drug safety, and clinical data, which is highly significant. For medical staff, this means less time spent repeatedly checking references and searching data; for patients, it leads to faster severity classification and clinical decision-making, resulting in a tangible improvement in experience.

New Possibility of K-Fold-Based Drug Development as a Service Presented. Provided by the Ministry of Science and ICT.

New Possibility of K-Fold-Based Drug Development as a Service Presented. Provided by the Ministry of Science and ICT.

View original image

Beyond Hospital Diagnosis, New Drug Development at '30x the Speed'


Major changes are also underway in the field of new drug development. The KAIST consortium's bio AI model 'K-Fold' demonstrated accuracy in predicting protein complex structures comparable to Google DeepMind's AlphaFold 3, while delivering results in less than one minute on average—a speed up to 30 times faster than the previous standard of 30 minutes. The model's improved performance in predicting rare new drug complexes is also expected to greatly reduce the time required to discover new drug candidates.


This technology is noteworthy not only for enhancing research efficiency, but also for reducing trial and error in the early stages of new drug candidate discovery, thereby lowering costs for pharmaceutical companies and research institutions. Previously, analyzing candidate protein structures required significant time, but now faster iterative validation is possible, meaning the overall pace of therapeutic development itself can be accelerated.


These AI models will be released as open source starting in April. Lunit plans to expand real-world testing at nine hospitals and SK Biopharmaceuticals, and will also pilot a customized medical chatbot that allows the general public to consult about their health status and medical questions.



The Ministry of Science and ICT views these achievements as a turning point, moving beyond the laboratory to real-world applications in diagnosis, treatment, and new drug development. Choi Dongwon, Director of AI Infrastructure Policy at the Ministry of Science and ICT, stated, "Even within a short period of about five months, we have secured a foundation of medical and bio-specialized AI technology that can compete in the global market," adding, "We will actively support commercialization and field dissemination so that the public can experience faster and more precise medical care and the benefits of new drugs."


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing