Automated quality checks on repeat prescribing.

Background: Good clinical practice in primary care includes periodic review of repeat prescriptions. Markers of prescriptions that may need review have been described, but manually checking all repeat prescriptions against the markers would be impractical.

Aim: To investigate the feasibility of computerising the application of repeat prescribing quality checks to electronic patient records in United Kingdom (UK) primary care.

Design Of Study: Software performance test against benchmark manual analysis of cross-sectional convenience sample of prescribing documentation.

Setting: Three general practices in Greater Manchester, in the north west of England, during a 4-month period in 2001.

Method: A machine-readable drug information resource, based on the British National Formulary (BNF) as the 'gold standard' for valid drug indications, was installed in three practices. Software raised alerts for each repeat prescribed item where the electronic patient record contained no valid indication for the medication. Alerts raised by the software in two practices were analysed manually. Clinical reaction to the software was assessed by semi-structured interviews in three practices.

Results: There was no valid indication in the electronic medical records for 14.8% of repeat prescribed items. Sixty-two per cent of all alerts generated were incorrect. Forty-three per cent of all incorrect alerts were as a result of errors in the drug information resource, 44% to locally idiosyncratic clinical coding, 8% to the use of the BNF without adaptation as a gold standard, and 5% to the inability of the system to infer diagnoses that, although unrecorded, would be 'obvious' to a clinical reading the record. The interviewed clinicians supported the goals of the software.

Conclusion: Using electronic records for secondary decision support purposes will benefit from (and may require) both more consistent electronic clinical data collection across multiple sites, and reconciling clinicians' willingness to infer unstated but 'obvious' diagnoses with the machine's inability to do the same.