Correctly interpreting large amounts of data is becoming increasingly important in the healthcare sector. Deep tech start-ups want to support physicians in doing this.
Signs of the drama that is about to unfold begin to appear almost 30 minutes beforehand. But no one takes any notice, because the monitoring devices don’t sound the alarm until it’s too late. Suddenly, seemingly out of nowhere, the patient experiences a circulatory collapse, and the doctors rush in to resuscitate her.
“When we had stabilized the patient, everyone said: it came as a complete surprise,” recalls Alexander Meyer, a professor at the German Heart Center Berlin who was on duty that evening in 2017 when the incident happened. But a look at the data from that night shows that there were actually enough signs to predict the collapse – they just weren’t recognized.
“In the intensive care unit, we currently work reactively,” says Meyer. Only when the body is in dire need “do the traditional alarms go off. But by then it’s already too late. By then, the consequences for the patient are massive.”
For the 37-year-old physician, who also has a degree in computer science, the solution to this problem lies in artificial intelligence (AI). These algorithms that learn as they are applied can be trained to recognize patterns in values that are measured in order to react to even the smallest deviations; deviations that signal early on that important bodily functions are impaired, that a collapse is possibly imminent.
As a physician who had to care for 20 intensive care patients at a time while on duty, such an early warning system is something Meyer had always wished for. A system that “scans all the data and prioritizes situations for me in real-time, for example, check on this patient now, complications are developing.”
Until a few years ago, this system was still just a dream. But he has since developed it himself. x-cardiac is the name of the start-up through which Meyer wants to save lives while at the same time reducing pressure on medical personnel. His AI has already been approved for predicting post-operative bleeding after heart surgery. “This is a problem we face daily, and everyone longs for a solution,” Meyer explains. In a first step, the system will be tested in various hospitals in Germany. If everything goes according to plan, it could soon be available throughout Europe. “We are also already looking at getting approval in the US,” says Meyer.
The Princely House of Liechtenstein, the owner of LGT, has been successfully pursuing entrepreneurial activities for centuries. Entrepreneurial thinking and actions are deeply rooted in LGT’s DNA.
Around the world, start-ups like x-cardiac are fueling the hope that adaptive computer systems can significantly improve medical care for billions of people. The spectrum of possible applications ranges from automated image analysis for tumor detection, drug development and AI-supported telemedicine, to optimizing processes in hospitals.
“Artificial intelligence offers an opportunity to dramatically improve patient care, early detection of disease and efficiency at hospitals,” says Mariam Kremer, a partner at Global Founders Capital with a focus on artificial intelligence.
Investors are showing particular interest in the potential of AI in medical imaging and drug development. According to market researcher Signify Research, more than 13 billion dollars in start-up capital has flowed into these two areas to date. The lion’s share (10.7 billion dollars) went to start-ups specializing in solutions for the pharmaceutical industry – companies like Relay Therapeutics, BenevolentAI and Berkeley Lights.
Even though each of these companies is pursuing its own approach, the primary goal is usually to identify new active ingredients and develop more effective drugs. And to do so far more quickly than has been possible to date. “The ever-increasing capabilities of artificial intelligence are helping us find molecular compounds that are essential for drugs,” explains Mariam Kremer.
The key here lies in what is the greatest strength of the current systems: pattern recognition. The more information is available in a given field, the more precisely the algorithms can be trained to discover correlations that humans easily overlook – especially when millions of medical studies and terabytes of data are involved. Added to this is the possibility of simulating chemical reactions on a computer to facilitate the subsequent production of drugs.
The power of this approach is exemplified by the Covid Moonshot project, which aims to swiftly develop a “pill that cures Covid”. In March 2020, the start-up PostEra called on pharmacologists around the world to submit suggestions for potential drug combinations that they would then assess using AI to identify promising candidates. Hundreds of scientists, as well as universities and pharmaceutical companies, are now participating in the open-source project: all of the results will be made available to the general public.
“We started this initiative to develop an antiviral drug – without patents, without intellectual property, without profit,” says the CEO of PostEra, Aaron Morris. The Covid Moonshot project recently received nearly ten million euro from the Wellcome Trust to begin clinical trials in 2022. “This would be the first time in history that a crowdsourced drug goes into human trials,” says Morris.
The AI helped in two ways: it sped up the process of identifying promising chemical compounds, reducing it from weeks to just a few hours, and it subsequently helped the researchers to construct the recipes to make the desired molecules. The second step in particular is essential, Morris explains: “Of course, the design is very important. But you then have to produce the molecules by synthesis and then test them – that’s the real challenge.”
With its AI system, PostEra aims to automate this difficult step in order to make drug development more predictable, literally. Because until now, the risk for pharmaceutical companies has been enormous: around 90 percent of all projects fail, often after years of research and huge investments. On average, it costs 1.3 billion dollars to develop a new drug from the laboratory to its market launch.
The interest of established companies in collaborating with AI start-ups is therefore high. New partnerships are being announced continuously, and the pandemic has accelerated this trend. “Drug development was at the forefront of most professionals’ minds as a way to tackle the pandemic directly,” explains Signify Research Analyst Imogen Fitt. But she cautions against people getting their hopes too high: there are no FDA-approved drugs available that have been structurally designed by AI.” And while there are “several ongoing projects which have potential, we’ve got to be careful not to expect too much too soon,” Fitt says.
In contrast, AI systems that are already being used in the real world include those that support radiologists in assessing MRI scans or X-ray images. Here, too, the strength of algorithms in pattern recognition is put to good use. How? Using millions of sample images as a basis, so-called neural networks (Link in German only) are able to detect lung cancer at an early stage – sometimes more reliably than human experts, because computers never tire and examine the thousandth scan just as thoroughly as they do the first.
“Radiology is an area of medicine where AI has started to have a positive impact,” says Sanjay Parekh, an expert on AI in imaging at Signify Research. According to Parekh, the systems have proven particularly useful as digital assistants that support people and make them more productive. “The benefit of AI is to help doctors get to a diagnosis quicker and more accurately by automating time-consuming tasks.”
Parekh doubts that AI alone would be able to take over the role of a doctor and replace humans any time soon. “Medicine is too complex of a field to have an AI algorithm come up with a final diagnosis,” he argues. Especially due to the fact that the susceptibility of today’s systems to irregularities in their training data has been demonstrated time and time again. Even small changes are enough to throw the algorithms off track and produce misdiagnoses.
Making the data easily readable by an algorithm takes a lot of time and effort.
Imogen Fitt, research analyst (Signify)
In addition, there are still many areas where there is no reliable data available to train the digital assistants. With the exception of radiology, which produces plenty of image material, developers have a hard time compiling information and preparing it in a way that it can be used by a computer. “Health data today is still very untidy. It’s hard to access and unstructured. Making the data easily readable by an algorithm takes a lot of time and effort,” he says.
Details of this kind create obstacles for researchers trying to translate solutions from the lab to the real world. Success requires not only capital and expertise, but also a lot of patience. Because unlike e-commerce, delivery services or social media, every product must be certified by the authorities. After all, the technology is supposed to save lives – not endanger them.
Kilian Koepsell has no problem approaching the market gradually, study by study. Koepsell, who was born in Germany, spent years researching neural networks at the University of California, Berkeley, before founding his current company, Caption Health, in 2014. His goal? To detect a wide range of diseases early on by making ultrasound exams supported by artificial intelligence commonplace.
“Ultrasound is a great technology,” Koepsell says. “You can look inside the body, examine virtually every organ, and see if something is going wrong long before symptoms appear.”
However, this requires a lot of expert knowledge. The ultrasound machine must be guided precisely while being held at the right angle to generate images that provide meaningful information – which must then be interpreted.
Caption Health’s AI comes into play during each of these steps. The software guides its users interactively through the examination; shows them where to place the ultrasound transducer; and compares image quality with comparative data to provide feedback to the user. If everything is as it should be, the recording starts automatically and is evaluated at the end of the examination.
“Our goal is for everyone who has medical training to be able to use the device,” Koepsell says. Caption Health has already received approval in the US for a specialized heart exam, and in one study, the AI showed it could also help assistants who had no experience with ultrasound to successfully conduct such examinations. Next year, a study on rheumatic heart disease, which hits children in developing countries particularly hard, is scheduled to start in Uganda.
We want to demonstrate that it’s possible to save millions of lives.
Kilian Koepsell (Caption Health)
“We want to demonstrate that it’s possible to save millions of lives if we can detect the first signs of this disease,” Koepsell explains. For the 49-year-old family man, the project is the perfect example of how medicine should evolve: instead of going to a doctor when something hurts, people should be continuously and predictively monitored.
“Our current healthcare system is set up to be far too reactive,” he says. “There are a lot of serious illnesses that could actually be prevented and that incur unnecessary costs.” Why, Koepsell asks, shouldn’t it be possible to protect the body better – like in the case of cars? “When it comes to our vehicles, we don’t wait for something to start rattling; we have them checked regularly and comprehensively.”
This logic is similar to that of the AI developed by Berlin-based heart specialist Alexander Meyer: check earlier, measure again and again, and automatically detect any deviation from the norm in order to prevent something worse from developing later on.
Fitness trackers also give many people hope that collecting data and digitally monitoring their own bodies will help them lead healthier lives. Unfortunately, it’s not that simple: with much of the data collected in everyday life, “we simply don't know: is this pathological or not?” explains Meyer. “We need to amass more knowledge in this area. It’s a field that’s still in its infancy, and we don’t want to do any harm, we want to heal and help."
In the meantime, it’s reassuring to know that we humans no longer have to deal with the age-old task of keeping our bodies healthy on our own.
Executive Briefing of McKinsey & Company
Transforming healthcare with AI: The impact on the workforce and organizations