Clinical AI Moves from Promise to Practice
Clinical AI creates lasting value when it solves a real medical problem, fits naturally into the work of clinicians, earns trust through evidence, and becomes sustainable through regulatory, reimbursement, and security pathways. At the Global Health and Purpose Summit, as part of People and Planet United, Junmyung Kwon, Founder and CEO of Medical AI, joins host Ivan Ruiz, Partner at FINN Partners, for a leadership conversation on “What It Takes to Bring Clinical AI into Real Practice.”
Ruiz frames the session around a central challenge for healthcare innovation. AI can move from technical capability into real clinical application only when adoption, usability, and impact are designed into the work from the beginning. Kwon answers that challenge through Medical AI’s experience bringing ECG-based clinical AI into hospitals, explaining how the company has built products, evidence, workflows, regulatory approvals, reimbursement pathways, and secure deployment models around the realities of care delivery.
The session is especially valuable because it distinguishes impressive technology from usable clinical infrastructure. Kwon presents Medical AI’s ETIA product family as a case study in how AI can use raw ECG data to uncover hidden cardiovascular signals while preserving the existing hospital workflow. The larger message is practical and important. Clinical AI succeeds when technology becomes easier for clinicians to use than to avoid, and when the benefit appears in the patient pathway rather than in the software demonstration.
The Gap Between Algorithms and Hospitals
Kwon begins by identifying the central question that defines the session. Many AI products look compelling in academic papers, technical presentations, or controlled demonstrations, but the real test is whether they work inside a hospital. That distinction changes how leaders should evaluate clinical AI. The relevant question is not whether a model can perform in isolation. The relevant question is whether a model can be validated, integrated, paid for, trusted, and used without adding friction to clinical practice.
Clinical AI is not just a software problem. It is a clinical problem, a data problem, and an engineering problem all at once.
— Junmyung Kwon, Medical AIMedical AI was created to close that gap between clinical reality and computer science. Kwon describes himself as both a medical doctor trained in emergency and critical care and an AI developer who writes code himself. He explains that the company was built by doctors, scientists, and AI engineers, with five full-time medical doctors and roughly 100 people overall, most of them engineers and researchers. The operating principle is straightforward. Doctors and engineers need to work at the same table because the clinical problem, the data architecture, and the engineering requirements are inseparable.
Solving a Real Cardiovascular Problem
The first requirement for clinical AI, Kwon argues, is solving a real and important problem. In this case, the problem is cardiovascular disease, which he identifies as the world’s leading killer, with 19.9 million deaths each year. Heart failure is a particular focus because many patients are diagnosed too late, and the tools currently used for screening create a difficult tradeoff. Echocardiography is the gold standard, but it is expensive, time-consuming, and dependent on specialist access. ECGs are fast, inexpensive, non-invasive, and widely available, but traditional ECG interpretation misses structural heart problems.
Medical AI’s flagship product, ETIA LVST, is designed around that gap. It uses a normal 12-lead ECG, the test that hospitals already use, and generates a risk score from zero to 100 with a clear high-risk or low-risk result. Kwon emphasizes that it requires no new machine, no extra test, and no major change in workflow. The result appears as a line on the existing ECG report. In his presentation, Kwon states that ETIA can analyze a simple ECG and detect conditions including heart failure, heart attack, and aortic stenosis, with more than 150,000 patients paying to use the software each month in real hospitals.
The approach rests on a fundamental difference between what humans see and what machines can analyze. Physicians usually read an ECG as a visual waveform, using established rules learned through years of training. Medical AI works with raw ECG data. A 10-second 12-lead ECG sampled 500 times per second contains 60,000 data points. Kwon’s argument is that those raw numbers include hidden signals that the human eye cannot see and that AI can translate into clinical insight.
Evidence Before Adoption
The second requirement for real-world clinical AI is evidence. Kwon stresses that clinical AI needs peer-reviewed science, external validation, regulatory confidence, and proof that performance carries across real populations. Medical AI reports more than 70 papers in international journals, including the European Heart Journal and The Lancet Digital Health. In the session, Kwon presents ETIA LVST as achieving an AUROC of 91 percent in subgroup analysis, compared with 70 percent for NT-proBNP, and a negative predictive value of 98 percent.
The validation examples are central to the article’s leadership significance. Kwon describes a Dutch study that invited four leading AI ECG models from the United States, Europe, Taiwan, and South Korea and used an external dataset of Dutch patients with both ECG and cardiac MRI. He says Medical AI’s model showed the highest performance, which he considers especially strong evidence because it used data that did not belong to Medical AI. He also cites validation in India, where the model achieved 97.4 percent AUROC in a single center and where many apparent false positives had early structural heart disease. In Kenya, he describes a prospective study across eight healthcare sites with nearly 6,000 patients, reporting sensitivity above 95 percent and negative predictive value above 99 percent.
These examples support a larger point about generalizability. Medical AI collected more than 20 million raw ECG datasets from more than seven countries, across age groups, ethnic groups, and clinical settings. By working from raw ECG signals rather than image files, Kwon argues, the model keeps more of the information needed to perform reliably across contexts.
The Power of Raw Signals and Foundation Models
Kwon presents Medical AI not only as a product company but as a foundation AI platform for bio-signals. He describes a strategic choice the company made four years earlier, when many AI companies were renting cloud GPUs. Medical AI bought its own NVIDIA supercomputer, with 160 GPUs and 100 petaFLOPS, and dedicated it to AI ECG development. This infrastructure, paired with the company’s raw ECG dataset, enabled the development of a foundation model trained to understand ECG signals at scale.
The platform model matters because much of healthcare AI still operates as one disease, one model. A separate disease often requires a separate AI product. Medical AI’s approach, as Kwon describes it, is to use a single massive foundation model trained on raw ECG data and then fine-tune it for different diseases, including heart failure, heart attack, aortic valve stenosis, and other conditions. The objective is higher accuracy, stronger robustness, and faster development of new disease-specific products.
Ivan Ruiz captures the clinical meaning of this approach in the discussion that follows the presentation. He notes that AI in this context functions as augmented intelligence for clinicians because it can detect patterns that are present in the signal but not visible to the naked eye. Kwon reinforces the point by explaining that ECG devices already produce numerical data, but physicians do not usually use that raw numerical value. When AI works directly with the raw signal, it can identify disease patterns that traditional ECG interpretation does not capture.
Workflow Is the Battleground
The third requirement is workflow integration. Kwon is clear that clinical AI must fit the physician’s workflow, not force physicians to work around the technology. In a normal hospital workflow, a physician orders an ECG, a technician performs the test, and the result returns to the physician’s screen. With Medical AI’s workflow, the physician orders an AI ECG, the same machine and same patient are used, and the same 10-second test produces a result that includes the risk score and risk level.
Kwon’s strongest example comes from Kangbuk Samsung Hospital. A man visited a checkup center without symptoms. A technician performed a normal ECG, and Medical AI’s software flagged the patient as high risk. Because the hospital had already built a clinical pathway, the patient was sent directly to the emergency room rather than going home. Within 15 minutes of arriving, his heart stopped. The hospital team performed CPR, ECMO, and ultimately heart transplantation. The patient survived and went home.
The example illustrates the distinction between AI as a detection tool and AI as a clinical system. The software identified risk, but the hospital’s pathway turned that signal into action. The physician, technician, and patient did not need to perform extra work. AI operated in the background, while the clinical pathway moved the patient to the right place at the right time.
Reimbursement, Regulation, and Sustainable Scale
Real practice also requires business-model realism. Kwon states plainly that even strong AI cannot survive if no one will pay for it. In South Korea, Medical AI has a real reimbursement code from the national government. Kwon explains that a normal ECG costs about five dollars, while Medical AI adds about seven dollars, bringing the total to roughly 12 dollars. He compares that with a 100-dollar NT-proBNP blood test and a hospitalization that can cost dramatically more.
The point is not only reimbursement. Kwon also describes value creation where a specific code is not yet available. If the software helps fewer patients be missed, moves high-risk patients to the front of the echocardiography queue, and reduces late hospitalizations and readmissions, then the economics can work at the system level. That argument is especially relevant in markets where hospitals face penalties when heart failure patients return too quickly.
Regulatory pathways are equally important. Kwon says Medical AI has approvals in Korea, Europe, the United Kingdom, India, Vietnam, Indonesia, Thailand, and Africa. He frames the path from lab to regulator, hospital, and patient as the difference between a research project and clinical AI in real practice.
Clinical validation comes first. AI must prove safety, accuracy, and clinical utility.
— Junmyung Kwon, Medical AISafe Integration and the Next Platform for Bio-Signal AI
Kwon closes with a principle that every healthcare technology leader understands. In healthcare, organizations cannot move fast and break things. They must integrate safely and securely. Medical AI offers on-premise deployment for hospitals that want patient data to remain inside the building and secure cloud deployment for smaller centers. Kwon says the system has passed security reviews in demanding hospital environments and that the company holds multiple international security certifications.
In health care, you cannot move fast and break things. You must integrate safely and securely.
— Junmyung Kwon, Medical AIThe future vision extends beyond the hospital. Kwon describes Medical AI’s work on single-lead ECG and wearable devices, including smartwatches, patches, home monitoring devices, and bedside monitors. The goal is to bring hospital-grade heart screening and monitoring closer to people wherever a one-lead ECG can be captured. He describes this as a future built with device makers, hospital networks, pharmaceutical companies, and digital healthcare platforms that share the same vision.
What It Takes to Bring Clinical AI into Real Practice offers a clear playbook for healthcare leaders evaluating the next generation of AI tools. Clinical AI needs more than algorithms. It needs clinical relevance, raw data quality, scientific validation, workflow integration, regulatory approval, reimbursement pathways, security, and a path to scale. When those conditions come together, AI can move beyond the demonstration stage and become part of everyday care.
Watch the Dedicated Session
Access the full dedicated session page for What It Takes to Bring Clinical AI into Real Practice, featuring Junmyung Kwon and host Ivan Ruiz.
View the Session Page
