Performance, Human Factors and EMR

This weekend my classmates and I are working on various projects and studying for pending exams. One project coming due is development of a Request for Proposal for acquisition, installation and support of an electronic medical records product. Some of the communications with my project team-mates concerned issues of hardware configurations and what we should assume. This left me thinking about frameworks for performance benchmarking for EMR systems and metrics for usability.

Much work has been done developing performance benchmarks for high end computing systems. The most widely used framework is Linpack. The defacto standard has several flaws, but serves as the one widely understood performance benchmark. Vendors tend to both tout their performance as part of marketing information, and optimize their codes to maximize scores.

I went to the NIST and CMS web sites in search of information on their EMR certification programs and scanned the test outlines descriptions. All were focused on basic functionality, and none that I found spoke to performance or usability issues. At this stage of the game, focus seems to be all on doing transactions at all. Efficiency will follow later.

A few blog postings popped up when searching the web. Generally they involved discussions of contributors to poor performance. Basic issues tended to center around inadequate infrastructure such as slow or undersized disk, network choke points, under provisioned system DRAM and undersized CPUs. No discussion of benchmarking.

This is not an exhaustive search, but it suggests a void that needs to be filled. I can imagine a framework of transactions segmented by clinic types. A survey process could identify the basic statistical profile of transaction types. Various clinicians (MD, NP, RN, etc.) could be interviewed and and perhaps shadowed to find the types and frequencies of transactions. Time required to complete the transactions and keystroke and mouse clicks could be recorded. With a sufficiently large library of user profiles, model practices could be defined, say ranging from small single physician clinics to large academic medical centers. In principle it would be possible to use this data to create stochastic models of traffic flows between various nodes, such as data entry stations, servers, backup storage, remote sites, etc. Similar techniques are have been used for decades to model network systems and to predict performance in normal operations and in the face of traffic surges and outages.

System vendors would benefit from such systems. It would allow them to evaluate hardware configuration options for customers and predict performance prior to first installations and planned upgrades. This should enabling them to right size configurations and avoid frustrations from under provisioned systems. It would also allow them to optimize their software implementations.