Developing an advisor predicting inpatient hypokalemia: a negative study

Developing An Advisor Predicting Inpatient Hypokalemia: A Negative Study
Kou-Wei Chiu, MD and Randolph A Miller, MD Department of Biomedical Informatics, Vanderbilt University School of Medicine Introduction
Hypokalemia, defined as serum potassium (“K”) Data from the EHR system indicated 57,839 inpatients levels below 3.5mEq/L, is important clinically. It can from 8/1999–7/2003 stayed at least 48 hours. Of those, cause muscle weakness, renal tubular defects, 53,643 patients had at least one serum K measurement. A arrhythmias, and even death. A complex buffering total of 18828 patients received furosemide (34%), and system regulates serum K levels. Diuretics, like their relative risk of lowK (at any time, including furosemide, cause renally mediated K depletion [1]. admission) was 1.53. Approximately 10% of the 57,839 A previous study detected hypokalemia (lowK) in patients had initial (admission) K < 3.5mEq/L. Of patients 21% of inpatients taking furosemide [2]. Despite not receiving furosemide, 17.5% had simple lowK hypokalemia's significance, a search of PubMed crossings a median of 76 hours into the hospital stay; for reveals few articles that provide algorithms for patients receiving furosemide, 38.9% had such crossings at forecasting lowK. With the ultimate goal of building median 105 hours. On furosemide, the relative risk of an informatics advisor to help clinicians avoid lowK simple lowK crossing after admission was 2.2. in their patients, authors reasoned that linear For the patients receiving furosemide, using "simple regression using recent K measurements might lowK crossing" as the gold standard, the regression predict incipient hypokalemia, since serum K levels algorithm yielded sensitivities ranging from 52% to 55% might fall at a different rate after K store depletion.
(for 24, 48, and 72 hours into the future) and positive predictive value(PPV) ranging from 18% to 34%. Using Vanderbilt University Hospital (VUH) is an "significant lowK crossings", the algorithm yielded academic, tertiary care hospital in Nashville, TN. sensitivities ranging from 54-57% and PPV ranging from Since 1995, VUH has maintained a locally developed electronic health record (EHR) that includes care For patients who never received furosemide, using provider order entry (CPOE) and patient data "simple lowK crossing" as gold standard, the algorithm repository components. Authors completed a simple yielded sensitivities ranging from 46% to 47%, with PPV epidemiological analysis of the frequency of ranging from 15% to 30%. Using "significant lowK hypokalemia in patients receiving and not receiving crossing", the algorithm yielded sensitivities ranging from furosemide. With the goal of correlating serum K 46% to 48% and PPV ranging from 8% to 16%. levels and furosemide ordering, and with IRB Analysis and Discussion
approval, authors extracted anonymized data from the The algorithm's sensitivity for predicting lowK data repository systems spanning 8/1999-7/2003 crossings improved 7-11% when applied to patients inpatients whose admissions lasted at least 48 hours. given furosemide versus not. Sensitivities further Authors defined "simple lowK crossing" as improve 2-4% using the gold standard of "significant sequential serum K measurements, first above, then lowK crossing." Thus, simple linear regression upon below 3.5 mEq/L within a given time period, and recent 48 hours of historical serum K measurements “clinically significant lowK crossing" as a serum K remains a poor predictor (not much better than measurement above 3.5mEq/L followed immediately measured prevalence) for lowK crossings at our by either one measurement below 3.2 mEq/L or followed by 2 or more measurements below 3.5 sensitivities of algorithm may include incorporating other patient variables such as other labs and Authors wrote a PERL script attempting to medications. The same approach at different predict, using linear regression, lowK crossings at institutions (using local data) can produce different, “starting points” of 48, 72, 96, 120, and 144 hours locally appropriate monitoring recommendations. At into a hospital admission. The linear regression used present, our informatics advisor for hypokalemia all measured K levels during the 48 hours prior to prevention would recommend: measure serum K each starting point, and projected 24, 48, and 72 every 48 hours if initial value >= 4.0 mEq/L, and hours into the future (after the staring point). If the change monitoring to once every 24 hours if K < 4.0 regression line crossed 3.5mEq/L from above to below, a forecast of “impending lowK” during that References:
future 24-hour time interval was recorded. The [1] Brenner, B. "Brenner and Rector's The Kidney", regression predictions were compared to the “gold standard” of actual recorded lowK crossings of both [2] Lowe J, Gray J, Henry DA, Lawson DH. Adverse types. Algorithm performance was measured through reactions to furosemide in hospital inpatients., Br sensitivity, specificity, positive predictive value, and Supported by NLM Grant R01-LM-007995.
AMIA 2007 Symposium Proceedings Page - 910

Source: http://telemedicina.unifesp.br/pub/AMIA/2007%20AMIA%20Proceedings/data/papers/posters/AMIA-0473-S2007.pdf