Are computers wiser than a team of medical specialists for cancer diagnoses? Doctors doubt this and use computers so far mostly only as supporting assistants. But will it stay that way?
For Dietmar Frey this is not science fiction. The neurosurgeon at Berlin's Charité Hospital and his team of doctors and IT experts are working on computers with artificial intelligence that will be used to treat patients suffering from acute strokes. "This is more than an idea. We have the technology, a prototype and the first machine learning models," says Frey.
Around 270,000 German citizens are hit every year by "the blow". Then time runs out. If the brain is not sufficiently supplied with blood and oxygen, the tissue dies. The consequences can be dramatic, speech loss and paralysis threaten.
Medical guidelines in Germany state that the affected brain tissue is dead after four and a half hours and that side effects of therapy can cause even more damage - bleeding in the head, for example. Therefore, after four and a half hours, the patient is no longer routinely treated, says Frey. "This may be statistically correct, but it's not always the right therapy for the individual patient."
For Frey, a patient is more than a statistical quantity. He suspects that in certain cases therapies still make sense after four and a half hours - and in others, they may no longer make sense after two hours. But how do you know who gets what?
AIs are faster
For Frey, the answer is clear: "A machine could search through thousands of comparative data records on strokes in a matter of minutes. Match them and show samples that could help a doctor at the rescue station make a decision." A job for Dr. Data.
This is a nickname for computers that work with artificial intelligence (AI) in medicine. The doctorate is probably too high for that: Computers could and should not replace doctors at the moment, but support them as assistants in very different areas, says Felix Nensa, radiologist and computer scientist at Essen University Hospital. Especially where the human being is a misplaced person: in boring and tedious activities such as tumor measurement or when saving and scanning thousands of image and text files.
AI does not - yet - mean that computers like the legendary "Hal" in Stanley Kubrick's "Odyssey in Space" strive for domination and make people superfluous. KI stands for databases and computers that are trained to search for programmed patterns.
No good results without sufficient data
Meaningful programming is an art, also in medicine. "We need sufficiently large training data sets and we need to be able to map variability," said bioinformatician Benedikt Brors at the German Cancer Research Centre (DKFZ) in Heidelberg. For example, if a data set is too small, the computer cannot recognize patterns in very rare tumor diseases - too little input.
"In a nutshell, a computer recognizes that there's a cat in this picture. And there's a dog on it." If you present him with a new picture, he would assign a cat and dog correctly. In principle, however, the system can only learn what it is supposed to learn. "It can't do anything with a picture of a horse." In the best case, however, the programs learned with each new input and thus increased their efficiency.
For ten years now, the DKFZ has been developing an AI system that will indicate whether neuroblastomas - very rare tumors in children - are aggressive or rather harmless. Therapy depends on it. But it will still take time with the clinical trials. The requirements for certification are extremely high, says Brors. "Normally, you'd find a company to do it for. But with around 120 new cases in Germany every year, the market is too small." That would not be worthwhile for a company. "So we're gonna have to do this ourselves."
Dietmar Frey at the Charité also has this do-it-yourself tactic in mind. His computers have already loaded over 1400 strokes. They are real patient fates, anonymized and broken down into formulas. This includes information on age, sex, weight, smoking habits, previous illnesses, medication taken and laboratory tests. Frey plans to have this treasure trove of data ready for retrieval in the emergency room when a new patient with an acute stroke is admitted.
The optimal therapy in Minute Fast
Today an average of 10 to 15 minutes remained in the rescue center for diagnostics, another 10 minutes for a decision on a therapy, reports Frey. He would like the computer to analyze in the background immediately after the admission of a patient whether there were comparable symptoms of a stroke, how they were treated and whether this helped. The computer is supposed to spit out the results within minutes - so that the treating physicians can find the best therapy for the patient in front of them.
"The added value would be the weighing of individual benefits and risks," explains the neurosurgeon. "A doctor today has neither the time nor the capacity to roll mountains of files in the emergency room for comparisons. And he can't even calculate it in his head."
If he planned to go into clinical trials, the program could also benefit other clinics. It is docked with EU research and could radiate beyond Germany. Frey thinks about a spin-off, deals with data protection. It could be a big number - but who knows? There are hardly any votes worldwide on who researches what. There's competition.
In Essen, physician Nensa has been building image databases with a technology group for several years: CT and X-ray images of lung disease with clinical data. "We keep getting new cases," he says. "Then we look at the new pictures and sometimes we think, "Whoa, what's this now?" Doctors used to tap into electronic textbooks. "For this, we had to put this picture into words. That's difficult with rare lung conditions."
The AI database is different. "We can now mark areas in a new recording and let the computer search." Has there been a similar case before? "And if so, show me this comparative picture." And because a diagnosis is always associated with the stored case, there is a direct link to the specialist literature. If it says, "If you only see women, but my patient is a man, then I'm on the wrong track," Nensa says. But that, too, is information.
"For me, such AI is the first wave of tools that relieves the doctor's workload," he sums up. "In five to ten years, there will be more of these tools. The profession of the physician will change - more in the direction of merging and interpreting data," Nensa speculates. Maybe radiologists and pathologists would become a discipline. "Or there's Mr. and Mrs. Superdiagnostics."