Battling antibiotic resistance: can machine learning improve prescribing?

By Michael A. Ribers and Hannes Ullrich - Version in Deutsch

Antibiotic resistance is a major health policy challenge. As resistance has developed to all newly developed antibiotics, their therapeutic value has diminished over time. Although a lack of financial incentives for developing new antibiotics is a large problem for the health care sector, it can be addressed by implementing supply-side policy measures, such as subsidies or exceptions within the patent system. Another way to address this problem is to focus on physicians’ prescription practices, which can play a leading role in preserving the therapeutic benefit of existing medications. Both unnecessary over-prescribing that exposes bacteria to active ingredients and under-prescribing that enables resistant bacteria to survive and evolve support resistance.

BCCP Fellows Michael A. Ribers and Hannes Ullrich consider the case of suspected urinary tract infections in Denmark to show how policy measures based on comprehensive individual data and machine learning can reduce the number of unnecessary prescriptions and increase the number of expedient prescriptions. They show that data-based predictions can differentiate among bacterial and non-bacterial infections so well that a reduction in written prescriptions of around 7.42 percent is possible without reducing the number of treated bacterial infections.

Currently available diagnostic methods in general practitioners’ practices result in uncertain diagnoses, particularly for first examinations, leading to over- and under-prescribing. Detailed microbiological laboratory tests can identify possible pathogens and provide information on their resistance profiles. However, this information is usually available after a several-day delay, often equal to the duration of a treatment unit with an antibiotic. For many prescription decisions, information on the probability of bacterial infection can be helpful in allowing informed decisions on whether or not treatment with antibiotics could be delayed until detailed test results are available or should be initiated immediately.

To predict bacterial causes using a random forest, the authors link the anonymized laboratory test results of patients with presumed urinary tract infections from the largest medical laboratory in Denmark to rich administrative data. The data include complete prescription and treatment histories, past hospital stays, laboratory results, and demographics such as age, gender, profession, municipality of residential address, household size and type, and marital status, among others. Comparing data-based predictions with physicians’ prescription decisions, they propose policy rules that lead an improvement in antibiotic prescriptions based on a social objective function trading off the benefit of antibiotic treatment and the negative externality of an expected increase in antibiotic resistance.

The results show the potential for data-based predictions using machine learning methods. The available mass and complexity of data can be reduced to simple, useful information, thus contributing to resolving one of the major challenges of today’s health care system: the efficient use of antibiotics given the trend of increasing resistance. To assess its full potential and assess potential risks, further testing of the proposed method in general practice in collaboration with domain experts is necessary.

A longer version of this report is available for download as DIW Weekly Report 19/2019. The full paper is available as DIW Discussion Paper 1803 (open access pdf download).