Evaluation of point-of-care glucose testing accuracy using locally-smoothed median absolute difference curves.

BACKGROUND: We introduce locally-smoothed (LS) median absolute difference (MAD) curves for the evaluation of hospital point-of-care (POC) glucose testing accuracy.
METHODS: Arterial blood samples (613) were obtained from a university hospital blood gas laboratory. Four hospital glucose meter systems (GMS) were tested against the YSI 2300 glucose analyzer for paired reference observations. We made statistical comparisons using conventional methods (e.g., linear regression, mean absolute differences).
RESULTS: Difference plots with superimposed ISO 15197 tolerance bands showed bias, scatter, heteroscedasticity, and erroneous results well. LS MAD curves readily revealed GMS accuracy patterns. Performance in hypoglycemic and hyperglycemic ranges erratically exceeded the recommended LS MAD error tolerance limit (5 mg/dl). Some systems showed acceptable (within LS MAD tolerance) or nearly acceptable performance in and around a tight glycemic control (TGC) interval of 80-110 mg/dl. Performance patterns varied in this interval, creating potential for discrepant therapeutic decisions.
CONCLUSIONS: Erroneous results demonstrated by ISO 15197-difference plots must be carefully considered. LS MAD curves draw on the unique human ability to recognize patterns quickly and discriminate accuracy visually. Performance standards should incorporate LS MAD curves and the recommended error tolerance limit of 5 mg/dl for hospital bedside glucose testing. Each GMS must be considered individually when assessing overall performance for therapeutic decision making in TGC.