Introduction

Today I'm going to talk a bit about healthcare in AI and how it can transform AI as we know it today.

A little bit of a short bio about me. I'm João. I'm 36 years old. I'm a father of a boy, three-year-old boy.

I have a master's degree in biomedical engineer from Nova. I did my PhD in electrical computer engineering already in AI applied to medical imaging at Tecnico.

I spent two years in Brussels doing research at the University VUB and also at the university hospital there. In 2017, I moved back to Portugal to work at the Champalimo Foundation. It was the picture that you saw on the first slide. Hopefully, you know it visiting from the outside, not from the inside.

In 2021, in 2023, I was a teaching assistant also at IST, assisting on a deep learning course. And more recently, I became a co-founder. We are trying in the process of spinning off this company from the Champalimo Foundation, such that we can bring some products to the patients and spread it beyond the Champalimo Foundation walls.

Current State of Healthcare

So this is the outline of my presentation and I'm going to start with the current state of healthcare. Currently, we are more than 8 billion people around the world and global life expectancy has been increasing, as you can see on this chart over here. So this is a good thing, I would say,

But with this increase, we also have been having a higher incidence of chronic diseases that put additional pressure on these health care systems. So a growing population and with an increased burden on disease, it's not good if you cannot sustain it.

And at least just in Europe, 36% of adults have a chronic disease, at least one chronic disease. Together with a shortage of healthcare workers that globally is estimated about 15 million healthcare workers that are missing on the systems and just close to 2 million in Europe, this further aggravates the issues.

This picture over here, it's the registration forms for national health system in the UK, the paperwork that the patient has to fill currently. And this further aggravates the burden. We have lots of paperwork that have to be duplicated, replicated, whatever.

Disconnected systems still within the same hospital. Departments that cannot communicate with other departments within the same hospital. And if we go beyond the hospital walls, data doesn't flow as well most of the times.

We have European efforts, like the European health data space, to break these barriers within Europe, but we often travel abroad and we cannot easily bring our health data with us if something happens in another country beyond Europe. So, current practice has all these problems and another one that it's more on disease care than actually taking care of our health.

1And it's very computer centric. When you go to a doctor, normally you are facing the doctor, but the doctor is actually not facing you, it's facing the computer, it's putting the data into the computer and not being attentive to you. So that is also a problem.

And I believe that AI can also be a solution, even though it's a technology, it can be a technology to free up the doctor to actually be taking care of you towards a more human and patient-centered healthcare, which currently we might argue that we don't have.

Transformation of AI in Healthcare

So, the transformation of AI in healthcare, we can say that it started way before 2012, but in 2012 there was something that happened in deep learning, which was the CNN, so Convolutional Neural Networks, that came to revolutionize the image analysis. Traditionally, it was for photographs of the world. But radiology, as it works with medical images, it was a natural candidate for the use of AI, and it was the first domain of healthcare to adopt AI.

And some years later, there was a prediction from one of the godfathers of deep learning Geoffrey Hinton, that said in 2016, four years after this revolution, that people should just stop training radiologists now. It's completely obvious that in five years, deep learning is going to do better than radiologists. Guess what that didn't happen.

There were also other predictions and now we have this meme circulating About radiologists driving regular cars to go to work So

But I'm presenting this but I will also present good things that are happening and showing that we can indeed Have AI being transformative in healthcare So we have this company this startup a bridge in the US that is raising a lot of capital and what is doing is using generative AI and speech-to-text and to actually take notes of the clinical conversations between patients and physicians, such that the physician doesn't have to be typing every information that is listening to the patient, and can save up to 70 hours per month on these tasks from the physician, bringing up not just physicians, but also nurses and other assistants.

And this was, the previous example was on clinical notes, but there are like medication forms, insurance forms that the physician has to duplicate as well. So it has to write once, twice, and as many times as it is needed. And all of this can be assisted through AI, even though if it has to accept and put a final signature, there are a lot of processes within the patient journey and within healthcare that can be sort of automated, even if at the end we have to have a human really accepting it and making sure that it is correct.

Preventive Care and AI

One area that I also like and that I think AI will also have a huge impact is on preventive care. So just in 2019, we had 41 million deaths of chronic diseases that I mentioned earlier. And with these chronic diseases, we have cardiovascular disease, cancer, diabetes, osteoporosis, and many others.

And we have been developing tools to overcome this and to work more towards being preventive with screening programs that will allow the detection of earlier forms of disease that can lead to chances of faster recovery, lowering also the health care costs and lowering the burden on patients and families.

And we already know some of these preventive care and screening programs such as lung cancer, breast cancer, prostate cancer and colon cancer, where we have deployed screening programs across many countries. But we can do more than cancer.

Patients often, when they go to the doctor, they will get an x-ray, a CT scan, and just in the US, each year, more than 70 million chest x-rays are performed and more than 80 million CT scans are performed. And what can we do to extract more information from these scans? So we can have opportunistic screening programs running on these images. So these images might be used to look at the lungs of the patients, but there are other anatomies, other organs present on these images.

the heart, we have the vertebral bodies, the spine of the patient. And for instance, we can run algorithms to predict the risk of this patient developing osteoporosis or whether the patient already has osteoporosis. The same for cardiovascular diseases. And these come at almost zero cost.

So the patient has this information acquired for another specific reason. But there is information that we can extract and be proactive on trying to solve these issues.

The same for CT scans. The same anatomy is being seen. We can see the vertebral bodies. We can see the aorta and measure calcifications that we have on the aorta.

We can see the liver and measure some properties of the liver and understand if this patient has fatty liver diseases or not. and be proactive. So AI can certainly have a role in preventive care and being able to do more.

AI in Practice: Evidence and Impact

These were future prospects of AI, but we can also, and we have already evidence that AI is already transforming our care. So in the context of breast cancer, we have the screening program for breast cancer. And there we already have evidence of the benefits of AI through randomized clinical trials, where it was shown that AI is able to work as a second reader. So in Europe, whenever you are doing a screening program, you have to have two radiologists review the images.

And AI can act as a second reader, reducing the workload by 44% of the radiologists, while increasing the detection rate of cancer. You have other points that it can bring as well. So in terms of the gains of detection rate, you can have more 20% detection rate than just having two human radiologists and having one human and AI on the side. So it's really amazing what AI can already do in the context of breast cancer.

Beyond this, there is also the social impact that AI can have. So AI can bring care to rural areas where we don't easily get access to the best doctors. So we can have AI-driven robots, automation solutions, sophisticated patient remote monitoring, to ensure that we break this division that exists between urban and rural areas, such that we can have people in the rural areas also having the ability to have the best care possible.

This is a paper that I really like. I didn't put the reference, but if you want, I can give you the reference afterwards.

Addressing Biases with AI

So we also have biases in the healthcare knowledge, in the medical knowledge. So this is a new osteoarthritis x-ray. And what we developed in the early 20th century, a score, the Calgary-Lawrence score, to determine the severity of this osteoarthritis and it was developed in the UK in the early 20th century on a Caucasian population. And what we understand nowadays is that it doesn't serve well to predict the pain severity score on underserved and underrepresented populations. And what these researchers were able to do was to reduce this bias that this score had by using AI. And this is already very transformative, such that we can be more fair and equitable towards everyone.

But we also have to have, on the other side, understand what AI is actually looking for. Because there was another study that showed that AI could read from a chest x-ray the race of the patient, which is something that the human cannot do. So what is the AI looking at? And the researchers tried, they destroyed the images with filters and still went at some level of destruction of the images. AI was still understanding what was the risk of the patient. And this can also lead to biases. So we have to look at both sides. It can reduce biases. but it can also propagate biases. And most of the times it propagates biases because data also embeds biases. 1Data is not often representative, or the data that we currently have is not representative. So here you have the example of the distribution of cohorts used to train deep learning in the US, and you can see that there is a huge disparity. So, data is often biased and if we don't put the necessary safeguards in AI, it will likely also become biased.

Integrating AI into Healthcare Systems

But in order to bring all of this to the healthcare, there are some barriers that need to be surpassed, and we need to find ways to bridge the gap. And I would say that the biggest barrier is how to integrate these solutions of AI in healthcare. Most of the times we think that it's like an easy app, just another software that we can provide to the healthcare workers. But if this is not seamlessly integrated, they will never use it.

I've seen a lot of people just having the data, but because this requires another app that you need to go, they will not use it and they will do the regular standard of care. And just to provide you an idea based on images and what radiologists do on a daily basis, they might have to review between 20 to 100 CT scans a day, and each CT scan might have hundreds to thousands of images. So you can just imagine scrolling through this, doing the report,

And this is the pipeline when you come to the hospital, you go to the CT scanner and then the images are archived in the imaging archive and within a couple of days the physician, the radiologist will have a look at your images and create the report. If we add AI, and most of the ways that we have been adding AI is that we create a button in the user interface of the doctor, and if he wants to use AI, he will have to press the button, wait for the results, and then see. But we believe at the Champalimo Foundation that we can do way better because we can create rules and we can understand once an image arrives to the image archive, that this image we have an AI algorithm to run on it, we immediately execute it, we save the results back, and then when the radiologist opens the images, the images are there and the results are already there, making it much simpler and integrated.

And here you have an example of what we have done at the Sean Polimo Foundation, where we open a CT scan and the AI result for osteoporosis opportunistic screening is already there.

We have advanced visualization tools that allow the visualization of volume rendering. So normally you have to scroll through the images to visualize the entire body, but you can have these beautiful reconstructions done. You have multi-planar reconstructions with the three anatomical planes where you can scroll on each of these.

And going back to the other visualization, here you have the results of the opportunistic screening for osteoporosis that we are currently also developing.

Understanding Doctor's Needs and Bridging the Gap

So another thing that it is very important for us when we are developing these technologies and in order to bridge this gap is understand the doctors and understand their needs and try to solve it while reducing their workload.

And here, it is a project that we are running on the breast unit of the Champalimau Foundation, where we want to provide surgical planning to physicians that they don't need to scroll through those black-gray images, but they can have like a segmentation AI algorithm that can annotate the different tissues and can create these 3D models, digital twins of the patient that can allow us to visualize some vessels that we use for reconstruction of breasts on patients that underwent a mastectomy, and that allowed the physician to more easily visualize this, and this can also be superimposed to the patient in the OR, and then we provide an ad set of augmented reality to the surgeon, and he can immediately understand where these vessels are, because these vessels need to be preserved in order to not have skin necrosis.

But beyond this, there are many other needs and opportunities that need to be tackled. And you just have to keep in mind that all of these should be done without further increasing the burden of these patients.

Conclusion

So, going to the end of my talk, the key takeaways that I would like to emphasize is that AI has the potential to democratize access to the best doctors for everyone and do way more and do more preventive care. We need to understand the current needs and processes related to those physicians' needs and healthcare workers. And we cannot create just another more that doctors need to worry about if they want to use it. So we should try for careful and thoughtful implementation of these AI solutions.

And I believe that we can shift from disease care to real health care such that we can have an AI augmented human and patient centered care.

Thank you.