Benchmarking of healthcare providers based on claims data

Dutch Healthcare Authority (NZa)

Benchmarking healthcare providers based on claims data

Last change on 10th of December 2025, at 12:41 (CET) | Publication Standard 1.0

• Publication category: Other algorithms
• Impact assessment: Privacy Quickscan
• Status: In use

General information

Theme

Health and Healthcare

Begin date

2023-01

Contact information

https://www.nza.nl/contact

Responsible use

Goal and impact

Purpose: To compare healthcare providers' claiming behaviour to support surveillance investigations.

Impact: In surveillance investigations into signals and findings about healthcare providers, we use analyses with claims data to enrich the investigation. One way is to compare a healthcare provider with other (comparable) healthcare providers in the healthcare market through benchmarking. The degree of deviation from other healthcare providers from such a benchmark can be a reason to start more extensive research, but is not the only decision factor to do so

Considerations

"It is not possible on the basis of the claim data in most cases to give a definite indication of a violation by a healthcare provider. However, we can give an indication of whether claiming behaviour is very different from that of other healthcare providers. So we only use benchmarking to confirm signals (internal or external) and not as a source to generate signals.

Using personal data (age, gender) in correcting benchmarks (see Technical operation) makes this comparison of healthcare providers more accurate, as long as careful thought is given to which variables are corrected for. This is up to the analyst and is not determined by the algorithm."

Human intervention

What the benchmarking is applied to (which healthcare providers, indicator, adjustment variables) is always determined by the analyst and supervisor. The interpretation of the benchmark is then also always done by the relevant analyst and supervisor.

Risk management

The risks in terms of technology, legislation, costs, ethics, explainability and other issues are very limited. The algorithm is not critical, no direct decisions follow from it and the technique is understandable and explainable.

Elaboration on impact assessments

A legitimacy assessment has been done on the processes in which this algorithm plays a role according to our standard procedure. The impact on privacy of this algorithm is negligible.

Impact assessment

Privacy Quickscan

Operations

Data

The method is not tied to specific data, but is deployed at the NZa on claims data received from Vektis. Commonly, but not exclusively, used are: AGB code healthcare provider/carer, UZOVI, performance code, number of services performed, amount declared/allocated and patient data (pseudo BSN for indicator calculations that must first be determined per patient and possibly age/sex for corrected benchmarks).

Technical design

Benchmarking is applied given a predefined indicator and healthcare provider and group of healthcare providers to be investigated with which it is compared. For example: the average turnover per patient per year (indicator) calculated per healthcare provider, where the question is how a specific healthcare provider (the healthcare provider under study) compares with the rest. The healthcare providers are compared in a density plot (a representation showing the distribution of the calculated values) with a percentile score for the given entity, which are interpreted by the analyst & supervisor.

Within this method, it is also possible to correct indicator values for patient characteristics, such as age and gender. This can help correct for differences in patient groups within healthcare providers. Indeed, a healthcare practice with mostly older patients will (depending on the indicator) not be fully comparable to a healthcare practice with mostly young people. This correction is applied using linear regression. By predicting the indicator value based on categorical variables (categorical to avoid making assumptions about linearity), it is possible to distinguish the effect not explained by patient characteristics. This new value can then be used in the density plot as described above. This is done with the following steps:

1. Calculate indicator values per patient.

2. Fit a linear model to the indicator values, using the categorical adjustment variables as predictors.

3. Take the values of the residuals as new values.

4. Create the density plot using these new values.