Ukrainians are creating an AI model that has few analogues in the world

Cardio.AI is a Ukrainian AI platform capable of conducting mass screening for heart disease without hospitalization, wires, or queues. Instead of a bulky Holter monitor, it uses a lightweight disposable sensor. Powered by biosensors and NVIDIA H100 capabilities, the system performs diagnostics in minutes instead of hours.

How it works: we tell the story from the inside. Cardio.AI CEO Maxim Dyachenko shares in an interview with Liga.net the technology that could radically change the approach to modern diagnosis and monitoring of heart health.

The problem that has been solved

Today, diagnosing arrhythmias remains difficult and inconvenient for patients. The classic Holter monitor, which records cardiac activity, is worn by the patient for only 24 hours. This is usually not enough time, as most serious arrhythmias manifest themselves within 3-7 days or even two weeks. In addition, the device limits the person's mobility and comfort.

Cardio.AI CEO Maxim Dyachenko says: "We conducted our own randomized study: thanks to a grant, we provided biosensors free of charge to more than 400 random people. The duration of monitoring ranged from two to seven days. We wanted to find out whether it makes sense to conduct mass screening of the entire population, not just clinic patients. It turned out that the situation among ordinary Ukrainians is even worse than among patients in private clinics."

Cardio.AI allows for longer and more comfortable monitoring. As a result, doctors identify more risks and prevent serious complications.

“We want to make arrhythmia screening as routine as fluorography. So that everyone can undergo it once a year or two” explains Dyachenko.

How the sensor, cloud, and model work

The platform is integrated with disposable biosensors from the American company LifeSignals, with which Cardio.AI has an exclusive agreement for Ukraine. The sensor weighs only 25 grams and is attached to the patient's body, where it works continuously for seven days.

The CEO of Cardio.AI emphasizes: "This is convenient for clinics: if the sensor is lost, it's a loss of $35–40, not several thousand euros, as in the case of a classic Holter monitor. The patient does not need to come to the clinic several times, because everything is simplified as much as possible: the sensor is attached, the patient's data is entered into the system, and they go home. The data is automatically uploaded to the cloud and further processed. The patient can even return the sensor by mail in an envelope if it is more convenient for them."

The data is automatically uploaded to the cloud, where an AI model performs so-called labelling — marking each heartbeat. Previously, the doctor did this manually, but now they receive a pre-marked recording. The report is generated in 6–10 minutes instead of 30–40.

“From the very beginning, our ambition was for the system to be able to process millions of screenings per month within a single territory. This is a scale that no one else currently provides” the team says.

“AI does 99% of the work. The doctor checks the result and makes adjustments. One specialist can process data for up to 100 patients per day” emphasizes Dyachenko.

Cardio.AI does not operate on the principle of selling “boxed” software, but offers clinics a comprehensive end-to-end service for diagnosing arrhythmias. The team integrates the system into the workflows of medical institutions and takes care of the entire cycle: from connection to data processing. For doctors, this means a ready-to-use tool “from day one”; for patients, it means comfort; and for the company itself, it means participation at the beginning of the value chain, rather than only at its final stage, as is the case with equipment manufacturers.

For clinics, the integration process is as simple as possible and does not require significant costs: payment is made after receiving the result. This is significantly different from classic Holter monitors, where the cost can be thousands or even tens of thousands of dollars. Medical staff undergo a short onboarding process: how to properly attach the sensor, prepare the patient's skin, and enter their data into the system. No special training is required for further work.

The only challenge may be the quality of the internet connection: without connectivity, the doctor cannot quickly download the data. The new sensors have a streaming function via a smartphone or a special gateway. It allows you to track changes in real time, but reduces battery life. In rural areas, this can unfortunately be an additional limitation.

Technical basis: millions of data points and regulatory compliance

Development of our own AI model began back in 2017. The first stable results appeared in 2020, and since 2021, the system has been tested in clinics.

The model has been trained on 2 million hours of ECG recordings, and the database continues to grow: “The system is constantly learning, we mark new strips, correct errors, and add them to the training set. Our first breakthrough came three years after we started. And even with new frameworks, it is difficult to beat that model: the problem is very complex. It is not just pattern matching, but also analysis of the sequence of beats and their context. There are many types of arrhythmias in cardiology, and the model must know them all. Otherwise, it is not suitable for screening. So our approach is human in the loop. The doctor checks the results, as in the usual Holter software. The algorithm is just much stronger. But we are moving towards full screening, where the doctor only intervenes in cases that require it.”

The team explains that it is impossible to use ready-made AI solutions from open access because there are not enough high-quality ones among them: "There are over 50 types of arrhythmias, which are very similar to each other. To perform screening safely, the model must recognize all classes and types, otherwise the result is not suitable. That's why we spent so long building our own dataset and improving the algorithm."

Training takes place on the Ukrainian De Novo platform with NVIDIA H100 GPUs, and processing (inference) is carried out in AWS clouds. At the same time, an important rule is followed: patient data is stored only in the country where it was collected.

Training is done in parts, because a seven-day ECG recording cannot be loaded into a single graphics processor. Therefore, the signal is broken down into 30-minute strips. To avoid the algorithm getting used to a specific patient, no more than three strips from one person are included in the training.

“For American patients, the data is stored in the US, for European patients — in the EU, and for Ukrainian patients — in Ukraine. This allows us to scale the system without violating the law” notes the CEO.

Cardio.AI complies with international HIPAA, QMS, and MDR standards and is certified in the EU and the US.

Prospects and scaling

Currently, most premium clinics in Ukraine, from small offices to large chains, work with Cardio.AI. In addition, a pilot project has already been launched in Azerbaijan. In Europe, the team is looking at the markets in Germany, Latvia, Spain, and Italy, where waiting lists for cardiologists can be up to six months long, and for Holter monitors, up to two months.

So Cardio.AI is not just replacing the outdated Holter monitor — it is the first Ukrainian world-class technology that allows for large-scale and rapid diagnosis of heart disorders. And with its entry into the European and American markets, the country has a chance to become a leader in global cardiac diagnostics.

Where in Ukraine can you get computing power for AI, and which GPUs are in high demand?

Cardio.AI trains its AI model on NVIDIA H100, renting capacity from De Novo, a provider that currently has the largest fleet of GPUs in Ukraine, specialized for machine learning, generative AI, and video analytics tasks.

De Novo CEO Gennady Karpov said that since the launch of the ML-cloud pilot project in early 2024, Ukrainian companies have leased dozens of NVIDIA accelerators of various models in the cloud: "Most customers are in the real sector, where AI is integrated into manufacturing, agriculture, and logistics. NVIDIA H100/A100 are primarily used for such tasks. Simpler models (L4 and L40S) are used by those who are in the stage of testing technical hypotheses or building MVP. The most powerful of our accelerators is the H200NVL. This flagship GPU is suitable for working with the largest (hundreds of billions of parameters) generative models.

The key parameter for running large generative models is GPU memory capacity. We are talking about hundreds of gigabytes, and there are simply no individual accelerators with such characteristics. Therefore, De Novo uses several H200NVLs connected via NVlink (a high-speed connection between graphics cards).

"In fact, four accelerators function as one with a total memory of 564 gigabytes. This is enough to run the largest models, which are much ‘smarter’ than smaller models" commented De Novo, which has invested over UAH 80 million in the launch and development of Ukrainian ML Cloud, Ukraine's first cloud infrastructure built specifically for the needs of artificial intelligence.

“We are already in the second wave of investment in ML infrastructure. We have purchased additional GPUs and storage systems specifically for the needs of customers who are experimenting with AI. This is not much by global standards, but so far it is the largest investment in AI infrastructure in Ukraine” - CEO De Novo Hennadiy Karpov emphasizes.