Big Data, AI and the future of hospitals

Peter Seitz, Executive Vice President Surgery, Siemens Healthineers, talks about the latest trends in medical technology and the changes taking place in the era of algorithms and robotic surgery.

2D versus 3D imaging in surgery

In the past, surgery typically has been open surgery without much technology. This has existed for hundreds of years. It's more a craft than a technology driven procedure. Imaging changes this significantly. Imaging is mandatorily necessary to move to minimally invasive surgery. In open surgery, you have the visual information and you have the haptic information. You can feel if in the liver tissue, there's a little of a higher density spot, which could be the tumour. If you lose this information by doing minimally invasive procedures, you need to compensate for this loss of information with imaging.

For very simple fractures, we probably wouldn't use any imaging. But in some, there are studies that show that the application of 3D imaging gives you a better assessment. Entry level imaging is 2D, meaning you get one plane overlaid and the entire information is visible. In that I would not be accurately be able to check where say a screw is. But I would get more information from a 3D point of view. When you apply only 2D imaging, there will be certain procedures you will not be able to reach the best possible outcome, like spinal surgery. In ankle fractures if you don't use 3D there’s a chance you may miss out in the perfect positioning of your screws.

The importance of simulation in training

I think the training of physicians will include technology at a much earlier stage. It will be a very practical aspect of simulation. At the moment, when you want to simulate a procedure, a typical medical student does this in the anatomy department. You cannot train everything in the anatomy department. You especially cannot train on a high number of specimens. Orthopaedic surgery is one area where we have a lot of collaborations. The typical orthopaedic surgeon trains in human specimens.

Once you have a simulator, once the procedure gets less manual, but more guided, maybe more robotic, more endovascular procedure, then I think we will have dedicated simulators where you don't practice on humans on human specimen, but practice on very realistic devices. Maybe you can even practice it in a pure digital way, where you just have the visualisation of the avatar of a patient. I think that is already happening today to some extent. If you or I would require a cardiac procedure, I think both of us would not opt for the student who does it for the first time. Now, how do you get to the stage that you are proficient enough to do a procedure like this? Let’s say someone has done this in a simulator a 1000 times, I think you could discuss that with me. But if someone has done it on just five specimens, I might be more hesitant.

A very practical example of complex spine surgery for example is scoliosis (a sideways curvature of the spine). If it's a simple spinal procedure, a surgeon would do it as it is today. If it’s a complex spinal procedure, you take the high resolution imaging of the spine, you 3D print it and you do a dry run in the real 3D printed anatomy. You double check if your assumed planning, trajectories and pathways work out or not. It's great if you make the first mistake in a 3D printed sample, and not in a real spine. This is not done on a large scale yet, but I think for very complex cases, especially paediatric surgery, I already encounter centres once in a while, which already do this. It's a great example, especially for orthopaedic surgery, because the bony structures you can 3D print and simulate in a really realistic fashion of how a procedure is performed.

Algorithms in the medical industry

The difference of a very junior radiologist of reading a CT (Computed Tomography) scan chest data set or a very senior radiologist reading is in the quality in terms of finding everything and diagnosing everything correctly. But the actual reading component is quite repetitive. In all those transactions where you can do it a million times, I think in the end, the human will not win against an AI algorithm.

Last year we introduced an AI pathway companion that looks at individual procedures. At the hospital, it takes all the data that is available on an individual patient of the Electronic Medical Record (EMR) and it matches it with procedure guidelines. Then for example, if treating breast cancer, the algorithm can tell you with the lab tests and anatomical information you currently have, these could be your three treatment options. And by the way, if you would add this genetic test, a fourth treatment option could become available. So the broader thinking of not missing a gamut of options, I think that is a very appealing use of AI in the diagnostic and the treatment space.

Data and predictive analysis

We have an installed base of 600,000 pieces of equipment worldwide. If we theoretically own all the data, we would drown in it. But data policy in most countries makes not us the owner of the data, but the patient and maybe the hospital. If we want to feed our algorithms for Machine Learning, we typically do research collaborations with institutions, where we have a legal framework on how we can use that data.

But it’s still a sufficient amount of data to be analysed. We have one particular algorithm, for example, when you find a lesion in let's say the liver, then there's a large database that's backing this up and it's comparing the lesion with many other lesions. It gives you recommendations saying this lesion has the highest similarity to the following lesions, in that case, the treatment and the outcome was the following. So you get a second opinion. That I think is a very realistic application. The comparison of anatomy with databases is important. It is of course necessary that your database doesn't only know yes, we have seen this before. It has to know what the treatment was and what the outcome was. Now if you have a large database for millions of lesions, then you can start predicting.

On AI and robots replacing the surgeon

Personally, I don't think it’s on the horizon. That’s not because it's technically not possible. It is. This always brings me to the analogy of autonomous driving. Is it technically possible today that our car drives autonomously? You could say yes, because there are already prototypes. Then the question is, it obviously doesn't work in 100% of the cases. 100% is the theoretical goal, but it probably works in 99% of the cases. That means in 1% of the cases something goes wrong. Who's reliable? The maker of the autonomous driving car or the driver?

This is not even the medical field where regulations are much, much, much stricter. So that's why I think this is not a topic for the near future, because as long as you are assisting the physician, great. As soon as you are flipping towards taking the liability, that's a completely different ballgame. To have an algorithm approved, that can be a second opinion, is a completely different story than to have an algorithm approved, that will be the operator. I would doubt that the robot and the algorithm from a liability point of view could take over. At the moment, I think it's a bit of a philosophical discussion.

But we do take a close look at robotic surgery. Let’s take spine surgery. Here you control the robot, because the robot maybe more precise in drilling than the human hand can be. So you plan based on prior images, what your trajectory will be. A robot with absolute perfection can follow that path. But the robot will not be able to decide the path in the near future. It may give you recommendations where you drill. But the responsibility if you drill two millimetres too far and you hit the spinal cord and the patient is permanently disabled. I think the likelihood that a robotic company will take this liability is pretty small.