AI in ECG Analysis: How Algorithms Detect VF/VT

[00:00:00] Lamborghini:
Could AI be used to say, “This is shockable with 80% accuracy,” or even more: “This is shockable because of this specific peak that looks abnormal on the ECG”?

[00:00:13] Voice-Over:
Welcome to MedTech Unfiltered, the Corscience podcast where we talk about the real challenges of developing and approving complex medical devices.

[00:00:27] Lamborghini:
Welcome to another episode of our podcast. This time we’ll be talking about ventricular fibrillation and ventricular tachycardia, and how AI is used to tackle these issues in AEDs, including some of the regulatory challenges. We have an expert with us today. Halil, would you like to introduce yourself?

[00:00:50] Halil:
Yes, I’m Halil. I’m an algorithm and software developer here at Corscience. My focus is on AI, specifically for our devices. Since we work with very critical devices, this comes with special considerations and constraints.

[00:01:14] Lamborghini:
Halil has been a developer here at Corscience for many years and is one of our go-to people whenever we have complicated issues to tackle. That’s why we have him here today. What you just mentioned brings us to our first question: how is AI used to analyze VF and VT in general, and how is it different from traditional methods, which rely on thresholds and fixed definitions? Can you explain that a bit?

[00:01:51] Halil:
Sure. Traditional methods are rule-based: the developer defines the features and characteristics to analyze, then sets thresholds that determine whether it’s VF, VT, or neither. An AI model learns these characteristics itself, through training with labeled data. So there’s always a training step before the model can be used for actual detection.

[00:02:32] Lamborghini:
So it’s the classical approach of training an AI model before deploying it. And as you said, traditional methods are rule-based: you define thresholds and the steps the device follows to decide whether something is VF, shockable, or not shockable, in the case of AEDs. Would you say there are advantages to using AI for this?

[00:02:59] Halil:
Yes, there are advantages. With rule-based systems, the developer has to figure out what to look for in the signal to identify VF or VT. That requires a lot of medical knowledge, and mistakes or wrong assumptions are possible. With AI, the model learns these characteristics itself from labeled data, so it knows whether a signal is shockable or not, and determines the result based on that. We assume this generally leads to higher detection rates than classical rule-based systems.

[00:03:59] Lamborghini:
So what I understand, from a non-developer’s perspective, is that it’s easier to train an AI system to detect VF and VT, because you don’t need all the domain knowledge a developer would need to set these thresholds, right?

[00:04:18] Halil:
Yes, that’s right. You take the labeled data, feed it to the model, and the model handles everything itself, setting its own internal thresholds and methods. With traditional methods, you first have to think through what VF or VT looks like, describe it mathematically, build a model from that, and then implement and test it.

[00:04:59] Lamborghini:
So that takes considerably more work. You mentioned data, which is a very important topic in AI, the classic “garbage in, garbage out.” We’re talking specifically about labeled data here. Can we talk more about that? How hard is it to get ECG data for training, and how is it used?

[00:05:32] Halil:
Data is a very important part of AI. As you said, garbage in, garbage out, so it has to be handled with a lot of responsibility. First, you have to obtain the data, which can be difficult, especially for VF and VT. There are several restrictions: you need access to the devices themselves, or you have to go through clinics, which also have to extract the data from their systems. Patient data privacy is also a major factor, so you can’t just use everything. This means the data needs careful preprocessing with respect to patient information. Another point is that the data has to be representative. You shouldn’t only use data from 80-year-old men. It has to cover all age ranges, both sexes, and different regions, not just white Europeans, but also data from Asia and Africa. That’s the ideal approach. So preparing data for training is a topic in itself.

[00:07:04] Lamborghini:
As you said, it’s practically a whole project just to get the data to train an AI. I agree. From what you mentioned, I identified a few challenges: getting the data directly from devices, which would require collaboration with companies that deploy AEDs, something like, “We need the ECGs from your past 200 missions,” because that data includes shockable and non-shockable cases. That would be your labeled data: this ECG was shockable, this one wasn’t.

[00:07:43] Halil:
The labeling itself is done by experts, not by us. We need at least two medical experts, doctors, to independently confirm whether a case is shockable or not. So again, this requires cooperation between medical professionals and us.

[00:07:59] Lamborghini:
That’s the first challenge. Then there’s the regulatory side. In the EU, data privacy is very strict. I recall there are seven, not articles, but labels that need to be removed from the data, like the patient’s name, to anonymize it properly.

[00:08:23] Halil:
You also have to check what’s allowed in each specific case. Age is usually fine, sex as well, in most cases. But anything that could identify the person has to be removed. For example, hair color isn’t allowed, because it could be combined with other information to identify the patient. Anything that could lead back to a person has to be deleted.

[00:09:01] Lamborghini:
From a technical standpoint, are there any data points, among the ones that are usually removed, that you’d actually like to keep? Maybe not hair color, but anything else that’s normally censored but could be useful for VF/VT detection?

[00:09:23] Halil:
The more information we have, generally speaking, the better. Information about medical treatment or other conditions can help, for example, if a patient had a previous heart attack and is now presenting with VF or VT. That context can help with categorization.

[00:09:48] Lamborghini:
I read about studies where researchers used a patient’s full clinical record, obviously not something you could analyze in real time during a mission while the device is reading the ECG, but this could be used afterward. I believe there was a study where they trained an AI that reached something like 70 percent accuracy in predicting that a person was at high risk of a heart attack within two weeks. Of course, it couldn’t say it would happen on a specific day and time, but it gave a useful window.

[00:10:27] Halil:
Exactly. If you have information beyond the ECG, you can feed it into the AI model as additional input and weight it accordingly, as a soft decision factor.

[00:10:42] Lamborghini:
How do you see AI being used directly in current workflows? AEDs are largely automatic, but there’s also the professional side, devices used by trained personnel, paramedics, or doctors. From my perspective, this is like a second opinion: ultimately, the doctor decides whether a shock should be delivered. Could AI be used to say, “This is shockable with 80 percent confidence,” or even, “This is shockable because of this specific peak that looks abnormal on the ECG”? How do you see this?

[00:11:28] Halil:
It’s very exciting. I think this represents the second generation of detection algorithms. AI can make fast decisions and also provide a confidence percentage for how reliable that decision is, giving direct feedback. I think this is one of the future directions for AI in medical devices, especially for VF/VT detection.

[00:11:59] Lamborghini:
So we’ve talked about how AI is trained, how it differs from traditional threshold- and rule-based methods, and how it fits into medical workflows for doctors and paramedics. We also touched on data and how it can be biased, and how large the datasets need to be. This is a question a lot of people have: how much training data do you actually need? Can you give us some examples or numbers for a company developing its own AI model?

[00:12:44] Halil:
You need data for training, validation, and testing. For the testing part specifically, there are clearly defined numbers from regulatory standards. For example, we need around 200 shockable VF/VT signals and a few hundred non-shockable signals, such as normal sinus rhythm and other rhythms. That’s just for testing. Training and validation require their own separate datasets.

[00:13:26] Lamborghini:
We briefly discussed this before the podcast, but what about synthetic data? Can it be used for training or validation? Why, or why not?

[00:13:38] Halil:
I think it can be used, especially for edge cases, like very high heart rates with a normal sinus rhythm, say 300 beats per minute, which we don’t really see in the real world but could still be useful as training data. You can also take real data and add synthetic elements, like noise, or shift it slightly. This is a common technique for generating additional training data, and I believe it’s allowed.

[00:14:14] Lamborghini:
Since you mentioned noise, I get the impression, and I’ve read this too, that AI tends to perform better under noisy conditions, which is actually the normal scenario for AEDs. These are devices meant to be used anywhere, a train station, a park, often by untrained people. That means situations where the pads aren’t attached correctly are common, which introduces noise. AI seems to handle those scenarios better, right?

[00:14:53] Halil:
Yes, better than a purely rule-based approach, because with rule-based systems you’d have to anticipate every possible type of noise in advance, and that’s not realistic. But if you train the AI system on data that includes different types of noise, it can learn to recognize and account for it, which leads to better results overall.

[00:15:32] Lamborghini:
I think that’s one of the amazing things about AI in general, not just for ECGs: it picks up on patterns that we might miss, processing large amounts of information and identifying what’s noise and what’s relevant. Earlier I mentioned a study on predicting high-risk patients, but I also recall researchers, I don’t remember which university, who used ECG data recorded after a cardiac event had already occurred, and were able to predict it about 30 seconds in advance. I found that impressive, because with traditional methods you’re essentially waiting for the ECG curve to show clear abnormalities before triggering an alert. These researchers, using AI, could identify it 30 seconds earlier. That’s not a lot of time, but it’s significant.

[00:16:32] Halil:
Right, but even a short amount of time can matter.

[00:16:35] Lamborghini:
Enough time to sit down or call for help. This is important context for people unfamiliar with shockable rhythms and cardiac arrest.

[00:16:50] Halil:
Every minute counts, every second counts. It’s really important.

[00:16:54] Lamborghini:
I believe survival chances drop by about 10 percent per minute without treatment, so after roughly seven minutes without help, the outcome is usually fatal.

[00:17:03] Halil:
And if someone is in VF or VT, every second in that moment counts. So the detection algorithm has to be extremely fast. There’s no time to analyze for a full minute, it has to happen within seconds.

[00:17:25] Lamborghini:
You need a result quickly, and I assume AI is better at that too. Does it need less time to determine whether something is shockable compared to traditional methods?

[00:17:38] Halil:
Not necessarily, that’s not a general rule. But you can design AI systems that are very fast. You train them on short segments of ECG data, a few seconds’ worth, and the model learns to analyze the characteristics within that short window and still produce a highly accurate result. I think it’s often easier to build a fast system with AI than with a rule-based approach.

[00:18:16] Lamborghini:
That does sound promising. That brings me to another challenge: regulation. AI is a huge buzzword right now, everyone wants to use it, there’s even AI toothbrushes. But from a regulatory perspective, especially in the EU, we have the MDR and the AI Act, which is already in force, plus other complementary regulations still being drafted. From what I understand, AI is currently treated like a fixed software package that can’t be changed once released. But one of the main advantages of AI is that it can be continuously trained and improved. How do you see this tension?

[00:19:15] Halil:
I’m cautious about a system that learns on its own directly on the device while it’s in use, because you can’t fully predict how it will learn or behave. The current standard approach is to train the model offline, arrive at a fixed model, and then implement that fixed model in the device.

[00:19:44] Lamborghini:
And that’s what you register with your notified body: “This is my model, and it’s fixed, no further changes.” Is that right?

[00:19:54] Halil:
Yes, no further changes. As you said, both the MDR and the AI Act specify that models in medical devices should not be trained on the device itself. It has to have fixed, well-understood behavior. You can retrain it offline to build an improved version, and then release that as an update, but not train it directly on the device.

[00:20:20] Lamborghini:
So no continuous training. I think this is handled a bit differently in the US market. The FDA has a tool called PCCP, which allows companies like Corscience to say, “We developed this AI model that determines whether a shock is needed,” and still update it over time. Essentially, the FDA approves a process rather than a fixed software package. Within that process, you can define, for example, “Every six months we’ll update the algorithm, and here’s how it affects the device.” It’s more flexible, partly because AI regulation in the US is generally more flexible. They treat AI more like an evolving system, while the EU takes a stricter, staged approach. I think that’s going to create real differences in how companies handle this between the US and EU markets, and I believe the UK is aligned more closely with EU regulation.

We’ve talked a lot about the regulatory side, how the FDA and the EU MDR handle this differently. What I wanted to ask now is more about the practical side. With traditional software, you document clearly defined rules, for example, “This threshold is used because of this specific reason.” With AI, it’s often more of a black box. How do you handle that from a development perspective? How do you document it, and why is this a challenge from a regulatory standpoint?

[00:22:32] Halil:
For the regulatory side, there are really two questions: is the system safe, and how does it actually work? Safety can be tested with data, but explaining how the system works is harder with AI than with rule-based systems. There are approaches to explainability. This is an active research area that looks inside the AI model to understand which parts of the signal it focuses on and how much weight each part contributes to the final decision. So you can analyze the internal workings of the system step by step. This matters for regulatory purposes because you need to show it’s not just a black box where you feed in data and get an answer without understanding why. It’s also important when a result doesn’t match your expectations. You need the ability to analyze why that happened and what improvements could prevent it in the future.

[00:23:58] Lamborghini:
It’s great that you brought up AI transparency. It’s a major research area right now, and accountability is very important in the EU under the MDR. In general, if your device fails, you want to know why, so that’s definitely a topic to watch. Another question: we’ve talked about how hard it is to get data and use it for training, and touched on the regulatory side. What about hardware limitations? How does the hardware need to adapt to support AI models in an AED, especially since Corscience is currently working on the smallest AED in the world? Can it even fit an AI model?

[00:24:56] Halil:
That’s an important question. The AI model you use has to fit the hardware you have. You can use hardware that’s optimized for AI, but you can’t make the device as large or as powerful as would be ideal for the most capable AI model. So the model has to be adapted to the hardware’s constraints, both in terms of computational power and model size. You need to determine in advance which AI model can realistically run on your target hardware. There are approaches to simplify a model to fit the hardware, but that has limits too, you can’t simplify it so much that it loses the detection accuracy you had before. So the priority is to choose hardware that can actually support the AI model you want to use. There are more powerful controllers available now, including some with dedicated AI cores, but the AI model still has to be matched to the hardware you choose.

[00:26:29] Lamborghini:
Getting a bit more technical here, but if I want to add an AI model to my device, would I need a different processor? Does power consumption roughly double? Does it need a bigger footprint on the PCB? There’s clearly a lot of hardware impact, and I’d guess higher cost too. What typically changes when you’re designing a device meant to support AI models?

[00:27:06] Halil:
It really depends on which AI model you want to use and what detection results you expect. For example, in our case, the controller needs to be fast enough and have enough memory. AI models tend to grow over time and need more memory and processing power, so the controller generally needs to be more capable than a standard one. Some controllers also come with a dedicated AI core, which can help address this. But overall, you need hardware that’s specifically capable of supporting AI systems.

[00:28:07] Lamborghini:
So there you have it. We’ve covered several important topics on AI in AEDs: how data is used for training and why it’s so hard to obtain, some of the regulatory aspects of bringing AI models to market in medical devices, and the hardware limitations involved in adapting device design. Halil, thank you for coming and for joining us.

[00:28:45] Lamborghini (Outro):
Subscribe to MedTech Unfiltered so you never miss an episode, and find more resources in the Corscience Academy.