Abstract P402: A Rapid Automated Method Detects Probable Errors Among Sequential Clinical Measures Recorded During Clinical Care

Abstract

Information collected during clinical care is increasingly used for research. Such data include errors and a rapid automated method for identifying likely errors is needed. This study aimed to develop a rapid automated method for identifying extreme values among sequential measures recorded during clinical care and to compare this method with the extreme studentized deviate many outlier procedure (ESD). Data were weights (W, n=1,151,434), heights (H, n=760,924), pulses (P, n=1,369,484), and systolic (S, n=1,680,583) and diastolic (D, n=1,680,503) blood pressures for 88,068 Veterans (82,874 men and 5,194 women) receiving care from the Phoenix VA Health Care System who had ≥5 five sequential values for these measures recorded in the electronic medical record during primary care visits from October 1999- September 2013. The new method identified extreme values by comparing individual values with three within-person metrics: 1. median (PMed), 2. interquartile range (PIQR), and 3. modified IQR [ModPIQR: PMed + / - smaller of (PMed - 25th percentile) and (75th percentile - PMed)]. These measures were selected because they were expected to be less perturbed by extreme values than the mean and standard deviation. Values exceeding cut-points (> 99th or < 1st population percentile) for ≥ 2 metrics were considered errors. High errors (0.46% W, 0.40% H, 0.55% P, 0.48% S, 0.50% D) were median 125%, 107%, 156%, 135% and 140% of PMed; median 3.3, 2.3, 2.8, 2.0 and 2.1 fold above PIQR; and median 5.0, 20, 3.4, 2.6, and 2.6 fold above ModPIQR for W, H, P, S, and D, respectively. Low errors (0.48% W, 0.43% H, 0.48% P, 0.46% S, 0.47% D) were 73%, 86%, 62%, 70%, and 65% of PMed; median 4.3, 5.0, 2.0, 1.8, and 2.1 fold below PIQR; and median 6.4, 50, 2.6, 2.3, and 2.4 fold below ModPIQR for W, H, P, S, and D. Compared with the new method, ESD (alpha = 0.01) identified fewer total (high+low) outliers for D (0.21% vs. 0.97%), P (0.58% vs. 1.03%), and S (0.14% vs. 0.94%), more outliers for H (2.40% vs. 0.83%), and similar numbers for W (0.70% vs. 0.94%). Both high (> 99th population percentile) and low (< 1st population percentile) extreme values for percent of PMed were detected with greater sensitivity by the new method (39%, 39%, 50%, 45%, and 46%; 41%, 40%, 43%, 41%, and 42%, respectively for high and low extremes for W, H, P, S, D) than by ESD (19%, 20%, 22%, 6%, 9% and 27%, 31%, 11%, 3%, 7%, respectively) while specificity was equally high (>98%) for both methods. The new method can be easily implemented and effectively identifies extreme values likely to be errors from among sequential clinical measures and could help reveal underlying longitudinal trends in weight/BMI, blood pressure or other clinical measures.