4.1 Estimating the proportion and frequency of AMR

The following list reviews the data sources available for estimating the proportion and frequency of AMR, including information on the analytical approaches used for each source, as well as their strengths and limitations:

• Isolate-based surveillance data, such as routinely collected microbiology results, is often used to calculate the proportion of AMR infections. The data is often available in electronic format, making it easy to access and analyse with minimal effort. However, it usually relies on blood samples, which may not capture the full range of infections like respiratory or urinary tract infections – especially in settings where clinical data is limited or not well structured. Additionally, low use of microbiology testing services can introduce bias, and hospital data is often needed to estimate infection frequency and to break down results by infection origin, age or sex.
• Prospective case-based surveillance data involves actively collecting data on patients who meet specific infection definitions. It allows for direct calculation of both the proportion and frequency of AMR infections and provides a detailed understanding of different infection types. However, it is resource-intensive, requiring significant investment in staff and infrastructure to identify and enrol patients and collect data.
• Point-prevalence surveillance survey data is collected at a specific point in time and offers a snapshot of AMR infection rates at a specific time, using moderate resources. It can complement routine data and help estimate both proportion and frequency of infections. However, results may not represent the full year, and sample sizes can be small. These surveys often underrepresent community-origin infections and short-stay hospital patients, leading to potential sampling bias.
• Systematic reviews use meta-analysis, a statistical method that combines and analyses the results of multiple studies to summarise AMR infection data from multiple studies. They are cost-effective, since they rely on existing data. However, differences in study quality, settings and methods can introduce bias. Additionally, most studies come from tertiary hospitals, which may not reflect national or regional patterns.
• Combining multiple data sources using advanced statistical models (e.g. machine learning or decision trees) can help estimate AMR burden in settings with limited data. This approach benefits from large data volumes, but faces challenges like data inconsistency, varying quality and assumptions about similarities between countries that may not be accurate. Community-origin infections and non-tertiary hospitals are often underrepresented.