Why a Seoul geriatrician seeing 100 patients a day decided the ward needed an algorithm that could explain itself
Dr Kwang Joon Kim sees close to 100 patients a day at Severance Hospital in Seoul, and he has never stopped. The geriatrician and endocrinologist runs a medical AI company alongside a full clinical load, which is unusual, and the reason is that the company exists because of what the clinical load kept showing him.
“I could not be at every bedside, and I often thought that with better data, faster insight, and smarter tools, many patient outcomes could have gone differently,” he said.
The wider pattern troubled him more than any single case. Korean patients concentrate in a handful of famous hospitals, which has hollowed out regional care and turned routine appointments into expeditions. “A patient in Busan might spend up to 13 hours travelling and waiting at a hospital in Seoul just for a consultation,” Kim said. “AI has to play a real role in closing those gaps.”
AITRICS was built on that conviction in 2016. Its flagship product, AITRICS-VC (VitalCare), integrates with electronic medical record systems and analyses routinely collected vital signs and laboratory results to generate real-time risk scores for clinicians. In general wards it predicts death, cardiac arrest or ICU transfer within six hours, sepsis within four hours, and cardiac arrest within 24 hours, while in intensive care units it predicts mortality risk within six hours. The software is live in more than 190 Korean hospitals and cleared as a medical device in South Korea, Hong Kong, Vietnam, Indonesia and Malaysia.
Three years of cleaning hospital data proved harder than building the model
The first obstacle had nothing to do with algorithmic sophistication. It was the state of the record itself.
“Real hospital data is full of noise that has nothing to do with a patient’s actual condition,” Kim said. “For instance, a temperature typed as 367 instead of 36.7, or a nurse writing 36.7, sweating a lot, into a field meant for numbers only.” The company spent close to three years on intensive data cleaning, and Kim describes it as still among the hardest and most important parts of the work.
A discipline emerged from that period which continues to shape the company’s research posture. “We were disciplined about never fixing biased data to get the result we wanted, which would have been unethical and would have quietly wrecked the AI model’s real-world validity,” he said. That refusal has a commercial cost in the short term and a defensive value in the long term, because models tuned to flatter their own validation sets tend to collapse the moment they meet a second hospital.
Regulators wanted one national threshold, and the company spent years arguing for configurable ones
Market authorisation took more than three years, and a substantial portion of that delay came from a genuine disagreement rather than administrative drag. AITRICS argued that alarm sensitivity and specificity thresholds needed to be adjustable by hospital and by ward.
“A large tertiary hospital wants to catch every possible deterioration even at the cost of false alarms, while a small hospital with limited staff wants alarms only when something is truly wrong,” Kim said. Regulators initially sought a single fixed national standard, and shifting that view required extended negotiation.
Healthcare does not run on startup timelines, and the company absorbed team attrition, cash shortages and regulatory delay across that stretch. “What got us through these difficult moments was the team,” Kim said. “We had and still have dedicated teams consisting of physicians, engineers, and data scientists approaching the same problem from completely different angles.”
AITRICS designed VitalCare for hospital Rapid Response Teams. Real usage told a different story, and the company followed it.
“We discovered that ward nurses, who need to know what is wrong with my patient, were actually the heavier users,” Kim said. “We rebuilt the interface and dashboards around that real usage pattern rather than defend our original assumption.”
The correction carried commercial weight beyond user experience. Rapid response coverage in Korea is thin, and research on the sector has noted that only two institutions operate 24-hour rapid response systems while most hospitals run partial cover or none. A product dependent on dedicated response teams would have addressed a fraction of the available beds. Kim recounts one overnight case in which an alarm triggered, the response team reviewed it and monitored the patient closely, detected a change in consciousness early, and carried out blood tests, medication and treatment in time.
The timing of Korea’s staffing collapse turned continuous ward monitoring into a procurement priority
Adoption accelerated against a backdrop of acute national strain. South Korea has one of the lowest doctor-to-patient ratios among OECD countries alongside the most rapidly ageing population, with a projected shortage of 15,000 healthcare professionals by 2035.
More than 90% of the country’s roughly 13,000 junior doctors resigned in February 2024 in protest at medical school quota expansion, and acute staffing shortages left some hospitals unable to admit new ambulances. Software that watches hundreds of ward patients continuously, and escalates only the ones deteriorating, became considerably easier to justify to hospital administrators across that period.
Combining AI with physiological data is well-trodden territory, and the Korean market alone holds several serious contenders. VUNO Med-DeepCARS focuses on cardiac arrest prediction using existing EMR data, thynC by Seers Technology collects data through wearable biosensors, and AITRICS-VC uses EHR integration to analyse changes in a patient’s overall health status. DeepCARS has been adopted across more than 48,000 hospital beds in South Korea, including 20 tertiary general hospitals, and received Breakthrough Device Designation from the US Food and Drug Administration in 2023.
Published performance separates the field. A multicentre prospective study of 55,083 general ward patients published in Critical Care found DeepCARS achieved an AUROC of 0.869 against 0.767 for the National Early Warning Score and 0.756 for the Modified Early Warning Score. AITRICS has pursued a comparable path with its VitalCare-Major Adverse Event Score and VitalCare-SEPsis Score models, built on a bidirectional long short-term memory architecture to predict clinical deterioration events within six hours and sepsis onset within four hours, with external validation in an independent cohort.
The cautionary case sits in the United States and has reset buyer expectations everywhere. The Epic Sepsis Model, implemented at hundreds of US hospitals, was externally validated across 38,455 hospitalisations at Michigan Medicine. It returned a sensitivity of 33%, a positive predictive value of 12% and an area under the curve of 0.63. Clinicians would have needed to review 109 flagged patients to find one requiring intervention to avoid sepsis.
Kim treats multi-institutional research as the price of entry. “In medicine, trust is only built on evidence, which is why we publish extensively and run multi-centre clinical studies to show that AITRICS-VC’s value is not only real but reproducible,” he said. “Every hospital has different patient demographics and workflows, so multi-institutional research is not optional.”
The commercial model reflects hospital budget reality rather than software convention. AITRICS charges on a per-bed, out-of-pocket basis, with the fee shared between hospital and company. “This model is favoured by hospitals as there is no initial investment from hospitals’ side in the product adoption and maintenance,” Kim said. In other markets, the company licenses on a yearly subscription tied to expected bed occupancy.
The addressable market is growing faster than the category containing it. The AI-powered clinical decision support market is estimated at $0.87 billion in 2025 and forecast to reach $1.79 billion by 2030 at a compound annual growth rate of 15.6%. Within that, predictive patient risk scoring holds roughly 14.7% share and is forecast to grow at approximately 26.4% annually through 2034, as value-based care contracts create direct financial incentives to identify high-risk patients before costly acute events occur.
Investors have backed the thesis five times, and the capital is now buying geography
Nine Korean and foreign venture capital firms and financial institutions invested a total of 35 billion won, roughly $23.7 million, completing the company’s Series C in December 2025. Premier Partners, Han River Partners, Mirae Asset Venture Investment and Shinhan Investment-BSK Investment returned as existing backers, while KB Securities-Solidus Investment, SV Investment and Mirae Asset Capital joined as new investors. Earlier rounds comprised 7.5 billion won across seed and Series A in 2017 and 2019, 3.5 billion won in a Pre-Series B in 2021, and 27.1 billion won in a Series B in 2024, taking the total raised to 73.1 billion won.
That capital is being spent on distribution as much as on research. The company established a Japanese subsidiary in March and a US subsidiary in December 2023, and has secured collaboration with the Mayo Clinic Platform covering model development, validation and global market strategy. The Mayo agreement includes future rights and distribution collaboration spanning R&D through commercialisation, and the company has separately worked with Cleveland Clinic to validate its flagship offering. MDSAP certification now supports multi-country entry, and full deployment is underway in Vietnam.
The most recent opening is in the Gulf. AITRICS has been selected for Cohort II of the Presight AI Accelerator, chosen from 376 applications across 62 countries, with the initial 33-company shortlist having collectively raised more than $341 Million in capital, generated over $28 Million in contracted annual recurring revenue and reached a combined valuation exceeding $2.1 Billion.
The programme offers direct connections to G42’s ecosystem of companies and government relationships, compressing a path from proof of concept to enterprise deployment that would typically take years to build independently. Cohort I participants generated more than 70 qualified commercial and strategic leads within 52 days of the programme starting, with several reaching advanced partnership and contract discussions with G42 Group companies, major enterprises and government entities. Thomas Pramotedham, chief executive of Presight, said that bringing leading AI innovators to Abu Dhabi strengthens the ecosystem with every cohort and reinforces the UAE as a home for applied and sovereign AI, where technologies are both developed and deployed at scale.
Kim treats the entry as a repeatable method rather than a single market win. “The UAE’s ambition to become a global hub for AI-driven healthcare, combined with Presight’s and G42’s network and infrastructure, makes this one of the most promising expansion opportunities we have outside South Korea,” he said. “We see it as a template for how AITRICS enters other markets going forward.”
The roadmap extends past the ward. The company intends to move beyond hospital walls into homes and long-term care using wearables and other data sources, and a second product line, the V.Doc Pro clinical co-pilot, is absorbing part of the Series C proceeds. “Our goal is AI that physicians trust and health systems adopt,” Kim said.
