Study: AI can significantly reduce misprescriptions for antibiotics


Researchers at the German Institute for Economic Research (DIW) show in a recently published study that the prescription of antibiotics can be reduced by up to 10.2 percent using machine learning methods without negative effects for patients. Wrong prescriptions of such drugs against bacteria and the inflammation caused by them are common. Due to increasing resistance, they cause considerable social costs, but according to the analysis they can be reduced by data-based decision-making aids.

The investigation, which is funded by the European Research Council, is based on extensive personal data from the administrative area, which has been linked for research purposes with laboratory data for diagnosing urinary tract infections. For this purpose, the DIW researchers used mostly personal information from Denmark, as it is comparatively easily available there.

Bacterial urinary tract infections are usually treated effectively with antibiotics. However, patients can show symptoms of such inflammation unless bacteria are the cause. In this case, treatment with the medicine will not be effective. At the same time, urinary tract infections are one of the main reasons for prescribing antibiotics in the population. Respiratory infections, which, due to the corona pandemic, are currently the focus of public attention due to the virus, are also well represented.

The temporarily less noticed trend of increasing antibiotic-resistant bacteria is continuing, warn the authors. Mainly due to a lack of financial incentives, however, hardly any new active ingredients have been developed for some time, so that treatment options have become fewer over time due to the resistance. British health experts estimated the cost of resilient bacteria to be US $ 100 trillion in 2050.

Excessive antibiotic treatments are seen as the main reason for the increase in resistant pathogens. A big problem, however, is that doctors usually have to determine the medication by visual inspection. This is because the laboratory analyzes required for an accurate diagnosis are only available after several days. Data-based predictions should therefore help to reduce temporary uncertainty and improve quick decisions.

With so-called ensemble methods it is the Scientists in the present case managed to predictthe probability of a patient’s laboratory findings containing bacterial pathogens at the time the sample was taken. To do this, they initially used large amounts of data from past patients for which reliable individual test results were available. They linked these with individual demographic data from the Danish administration. In doing so, they uncovered statistical relationships between laboratory results and other available personal information such as medical treatment histories.

In total, the team considered 95,594 acute treatment situations that were not immediately preceded by antibiotic treatment or diagnostics. From these, they predicted laboratory results in 42,480 cases on the basis of existing data and then compared them with their actual test results.

In order to measure the gain in forecast quality by successively linking additional personal data, the scientists divided the available data into the five sub-segments of time and region, age and gender, detailed personal characteristics, health data and the medical decision. They made this available to the algorithm one after the other. Adding the first to the last segment reduced antibiotic prescriptions by 1.18 percent in the experiment, and 7.42 percent when incorporating health data.

Using all available information, the experts achieved the maximum rate of 10.22 percent. “If all the predictions were perfect, 39 percent fewer antibiotics could be prescribed,” they write. “Using all the data, a good quarter of this maximum possible reduction could be achieved.” The reduction in antibiotics takes place only in the case of incorrect prescriptions.

The approach “stands in contrast to the negative examples of large, sometimes failed digital projects from the private sector,” the authors refer to, for example, to experiments with Watson Health from IBM, which would have provided little added value. The SARS-CoV-2 pandemic has now made the general public aware of the relevance of high-quality data. This dynamic could “spur applications in the healthcare sector and beyond, in order to master societal challenges in a more informed manner”.

What is necessary for this is an “infrastructure for data linking and provision that complies with socially accepted data protection and ethical standards,” emphasize the researchers. If such services were available to science and the health sector, “numerous opportunities could open up to significantly improve health care”. However, various ethical questions would also have to be discussed and weighed beforehand.


To home page