Mapping point-of-care performance using locally-smoothed median and maximum absolute difference curves

Abstract

Background: The goal is to introduce visual performance mapping efficient for establishing acceptance criteria and facilitating decisions regarding the utility of hospital point-of-care devices. This approach uniquely reveals the quality of performance locally, as opposed to globally.

Methods: After presenting theoretical foundations, this study illustrates the approach by applying it to six hospital glucose meter systems (GMSs) using clinical multi-center (n=2767) and multi-system (n=613, n=100) observations.

Results: LS MAD curves identified breakouts, that is, points where the locally-smoothed median absolute difference (LS MAD) curve exceeds the recommended error tolerance limit of 5 mg/dL (0.28 mmol/L). LS maximum absolute difference (MaxAD) breakthroughs, which occur where the LS MaxAD curve exceeds the 99th percentile of MaxADs from x=30-200 mg/dL (1.67-11.10 mmol/L), showed extreme error locations. A multi-sensor interference- and hematocrirt-correcting GMS displayed a flat LS MAD curve until it reached a breakout of 179 mg/dL (9.94 mmol/L) and generated breakthroughs that could affect bedside decision-making, but less erratically than other systems with inadequate performance for hospital critical care. We discovered Class I (meter high, reference low) and Class II (converse) discrepant values in some systems. Class I errors could lead to inappropriate insulin dosing and hypoglycemic episodes in tight glucose control.

Conclusions: LS MAD-MaxAD curves help assess the performance of point-of-care testing. Visual mapping of systematic and random errors locally over the entire analyte measurement range in a single integrated display is an advantage when considering the adverse impact of zones of poor quantitative performance on specific clinical applications, threshold-driven bedside decisions and the care of critically ill patients.