Prototype Prescribing Outlier Dashboard for Gp Connect PCN
At OpenPrescribing we are piloting a number of data-driven approaches to identify unusual prescribing and collect feedback on this prescribing to inform development of new tools to support prescribers and organisations to audit and review prescribing. These pilot results are provided for the interest of advanced users, although we don't know how relevant they are in practice. There is substantial variation in prescribing behaviours, across various different areas of medicine. Some variation can be explained by demographic changes, or local policies or guidelines, but much of the remaining variation is less easy to explain.
The DataLab is keen to hear your feedback on the results. You can do this by completing the following survey or emailing us at [email protected]. Please DO NOT INCLUDE IDENTIFIABLE PATIENT information in your feedback. All feedback is helpful, you can send short or detailed feedback.
This report has been developed to automatically identify prescribing patterns at a chemical level which are furthest away from “typical prescribing” and can be classified as an “outlier”. We calculate the number of prescriptions for each chemical in the BNF coding system, the count of all prescriptions within that chemical's BNF subparagraph, for prescriptions dispensed between June 2021 and December 2021. We then calculate the ratio of these counts along with the mean and standard deviation of those ratios across all PCNs. From this we can calculate the “z-score”, which is a measure of how many standard deviations a given PCN is from the population mean. We then rank your “z-scores” to find the top 10 results where prescribing is an outlier for prescribing higher than its peers and those where it is an outlier for prescribing lower than its peers.
For each outlier chemical, a kernel density estimation plot of all PCN's chemical:subparagraph ratios is provided, with this PCN's ratio overlaid in red.
It is important to remember that this information was generated automatically and it is therefore likely that some of the behaviour is warranted. This report seeks only to collect information about where this variation may be warranted and where it might not, to inform research on this topic. Our full analytical method code is openly available on GitHub here.
This is a new, experimental feature. We'd love to hear your feedback .
