The safety and quality of decision support systems.

OBJECTIVES: The use of clinical decision support systems (CDSS) can improve the overall safety and quality of health care delivery, but may also introduce machine-related errors. Recent concerns about the potential for CDSS to harm patients have generated much debate, but there is little research available to identify the nature of such errors, or quantify their frequency or clinical impact.
METHODS: A review of recent literature into electronic prescribing systems, as well as related literature in decision support.
RESULTS: There seems to be some evidence for variation in the outcomes of using CDSS, most likely reflecting variations in clinical setting, culture, training and organizational process, independent of technical variables. There is also preliminary evidence that poorly implemented CDSS can lead to increased mortality in some settings. Studies in the US, UK and Australia have found commercial prescribing systems often fail to uniformly detect significant drug interactions, probably because of errors in their knowledge base. Electronic medication management systems may generate new types of error because of user-interface design, but also because of events in the workplace such as distraction affecting the actions of system users. Another potential source of CDSS influenced errors are automation biases, including errors of omission where individuals miss important data because the system does not prompt them to notice them, and errors of commission where individuals do what the decision aid tells to do, even when this contradicts their training and other available data. Errors of dismissal occur when relevant alerts are ignored. On-line decision support systems may also result in errors where clinicians come to an incorrect assessment of the evidence, possibly shaped in part by cognitive decision biases.
CONCLUSIONS: The effectiveness of decision support systems, like all other health IT, cannot be assessed purely by evaluating the usability and performance of the software, but is the outcome of a complex set of cognitive and socio-technical interactions. A deeper understanding of these issues can result in the design of systems which are not just intrinsically 'safe' but which also result in safe outcomes in the hands of busy or poorly resourced clinicians.