A research team from China has developed a computer program that learns patient data
to evaluate childhood illnesses ranging from a harmless cold to asthma
to detect life-threatening meningitis. The program could be used to quickly urgent emergency departments
to differentiate from less urgent cases, doctors in training
to assist or diagnose complex diseases
To help symptoms, the researchers around Kang Zhang write from the center
for Pediatric and Women's Medicine at the University Hospital in Guangzhou. Indeed
experts criticize the study because of various shortcomings.
For their study, in the current issue of the science magazine Nature Medicine (Liang et al., 2019), the researchers first fed a computer program with a data set
around 800,000 electronically available patient records of children. This included
the medical history, laboratory results and abnormalities that had been shown in the examination by a doctor – but not the diagnosis. Next, the information was off
translated to medical records using a method called Natural Language Processing. This technique processes natural
machine written or spoken language and bring it into one
structured, processable for other algorithms form. In a further step, an algorithm then always ordered the data
As soon as the classes become finer: first the affected ones
Organ systems, then groups of related diseases to one
How the diagnosis was calculated remains unclear
The researchers then compared the accuracy of the computer-aided diagnosis with that of pediatricians – using a further set of approximately 400,000 data. The result sounds impressive at first: The computer-calculated diagnosis was always correct
Disease in 79 to 98 percent of cases with that of the attending physician
match. In competition with the pediatricians of flesh and blood was the artificial intelligence Superior beginners in making the correct diagnosis, while experienced physicians performed better than the computer program.
However, according to specialist opinion, the program has too many flaws to be used already in clinics or medical practices. According to medical informatics expert Thomas Neumuth from the University of Leipzig, the applied algorithms are not traceable. "This makes it impossible to find out how the decision is made," said the specialist for computer-based medical technology and surgery journalists of the German Science Media Center (SMC).
The biostatist Frank Klawonn from the Helmholtz Center for Infection Research in Braunschweig also criticized the fact that the study did not say how certain values for the detection of diseases were calculated. And he points to a clear mistake: In an attached table, other case numbers are mentioned than in the study itself. Apparently here a few columns have been reversed. "The numbers were handled so sluggishly that one can only very little trust in the entire study," said Klawonn the SMC.
Meanwhile, teams around the world are working on similar diagnostic algorithms using algorithms. For example, there are programs that can use computer tomography or X-rays to detect existing diseases, such as skin cancer, and to do so just as precisely as experienced physicians. Some of these programs are already being used as assistance systems. Although similar approaches could be made with the approach from the current study, experts say. But to judge how far the team from China is, their work ultimately contains too many flaws and is too intransparent. Presumably, the system would have to be significantly expanded.