Australian scientists have checked the symptoms of 1170 medical reports on 36 sites available online diagnostic and found that the correct diagnosis they put only 36 percent of cases, and in ten first-listed diagnoses, he gets 58 percent of cases. Most effectively coped with the sites at diagnosis “using artificial intelligence methods” and consider demographic data of the user. More important was inaccurate determination of priority of treatment for the patient, write the scientists in The Medical Journal of Australia.
Often the inability (or unwillingness) to seek medical help when unwell people trying to diagnose themselves using the Internet. Look for symptoms separately and sum them to determine the condition or illness that makes little sense, therefore, most people use diagnostic sites and applications (one of the most famous in the world — the English-language service of WebMD, and somen-language sites promise the accuracy of such diagnosis to 67 percent). Most of these services use quite large data arrays with the doctors and put the statistically probable diagnosis.
Of course, automatic online diagnostics does not replace seeking medical help, but for some can be a reason to go to the doctor, so it is important to evaluate their effectiveness. Decided to do this, researchers from the Edith Cowan University under the leadership of mills Brennen (Brennen Mills). For analysis they used the 36 most popular in Australia automatic online diagnostics — applications and websites are found through search engine queries: 10 of them were given the diagnosis and triage (prioritization of the patient in treatment), 17 were given only the diagnosis, but 9 — only triage.
To check the scientists collected 1170 medical reports and checked the symptoms in each of the selected online diagnostics. The correct diagnosis of the system was determined immediately for the 421 opinions (36 percent) in the first three fell to 606 opinions, and was listed among the first ten to 680 conclusions. System, which according to the creators “using artificial intelligence methods” for diagnosis and also take into account the demographic data was the most truthful and correctly diagnosed immediately for 46 percent of the conclusions.
Next, the researchers tested how effectively the system determine triage (this used 19 system, in which triage was included). The priority of the patient was correctly identified for 338 of 688 conclusions. For those cases where the patient needed urgent medical care, the priority was correctly issued most often (68 percent), and averaged over all conclusions was correct in 49% of the findings.
In addition, for 40 percent of the opinions in which the priority of the patient was low and recommended self-treatment system were given a triage emergency room. By contrast, 10 percent of the conclusions, which in fact required an emergency system recommended self-treatment and non-urgent medical care: in the list of such conclusions, for example, got severe liver damage and stroke.
Obtained by the authors results should be compared with those received in a similar the work of their American colleagues: in 2015, they found outthat the Americans used the system to give a correct diagnosis in 34 percent of cases and put a proper triage in 57 percent. Triage, according to scientists, is much more important: as all systems specify that online-diagnosis does not replace a trip to the doctor exhibited triage may be the reason to go to the doctor as soon as possible and get medical help. Therefore, it is very important to present correct: does not lead to unnecessary workload of doctors in cases when you can do self-treatment, and did not put human life in danger in cases where no assistance was required.
It should be clarified that the modern methods of machine learning can actually improve self-diagnosis in some cases. For example, a year ago, scientists demonstratedhow to use smartphone application and a paper cone can be used to diagnose otitis media in children.