Deepnoid on Medical AI: 'An AI That Works'
Medical AI is evolving beyond the stage of simply finding lesions (abnormal areas of the body) into an 'agent' — an AI that makes its own judgments and handles hospital tasks directly. Deepnoid CEO Choi Woo-shik presented on the theme 'From Generative AI Medical Devices to AI Agents' at the 2026 Future Healthcare Trends Conference, held on September 3 at COEX in Seoul.
Deepnoid spent more than a decade developing AI that detects lesions in medical images, but since the arrival of ChatGPT, it has broadened its development into generative AI. Unlike conventional reading-assist AI, which could only indicate the presence or absence of predetermined conditions, Deepnoid's generative chest X-ray interpretation AI 'M4CXR' reads 41 types of abnormal findings and generates preliminary interpretation reports. In November 2025, the product became the first generative medical AI device to be designated an Innovative Medical Device, and in June this year it received Class 3 medical device approval from the Ministry of Food and Drug Safety, with clinical deployment beginning in July.
Even after interpretation is complete, hospitals still face considerable work: drafting discharge summaries and progress notes, auditing medical records, and compiling metrics. Rather than merely answering questions, an AI agent can query hospital systems, exercise judgment, and autonomously carry out tasks from writing documents to notifying the staff in charge. To this end, Deepnoid developed its own medical-specialized large language model (LLM), 'MedZero.'
MedZero is a model of roughly 27 billion parameters, pre-trained on 11 trillion tokens of data. Through three stages of training — Korean language, medical knowledge, and reasoning — it can comprehend lengthy medical documents, make tool calls to query hospital systems, and perform clinical reasoning. Whereas using external APIs risks patient data leaving the hospital, MedZero runs on the hospital's closed internal network, allowing full control across the entire pipeline — a key security advantage.
When asked to draft a discharge summary, it reorganizes scattered information — diagnoses, surgeries and procedures, and the hospitalization course — in chronological order. The task of checking incomplete medical records, previously done by transferring discharged patient lists into Excel, cross-referencing them, and contacting physicians individually, is now handled autonomously by MedZero, which notifies the physician in charge of each case. It can also categorize the reasons for readmissions within 30 days case by case and group recurring causes into improvement tasks.
MedZero is designed to operate within the hospital's closed internal network with no external transmission, incorporating medical data security by design. It is becoming the technological foundation underpinning medical AI's evolution into 'an AI that works.'
