Analysis of electronic medication orders with large overdoses: opportunities for mitigating dosing errors

Abstract

Background: Users of electronic health record (EHR) systems frequently prescribe doses outside recommended dose ranges, and tend to ignore the alerts that result. Since some of these dosing errors are the result of system design flaws, analysis of large overdoses can lead to the discovery of needed system changes.

Objectives: To develop database techniques for detecting and extracting large overdose orders from our EHR. To identify and characterize users' responses to these large overdoses. To identify possible causes of large-overdose errors and to mitigate them.

Methods: We constructed a data mart of medication-order and dosing-alert data from a quaternary pediatric hospital from June 2011 to May 2013. The data mart was used along with a test version of the EHR to explain how orders were processed and alerts were generated for large (>500%) and extreme (>10,000%) overdoses. User response was characterized by the dosing alert salience rate, which expresses the proportion of time users take corrective action.

Results: We constructed an advanced analytic framework based on workflow analysis and order simulation, and evaluated all 5,402,504 medication orders placed within the 2 year timeframe as well as 2,232,492 dose alerts associated with some of the orders. 8% of orders generated a visible alert, with ¼ of these related to overdosing. Alerts presented to trainees had higher salience rates than those presented to senior colleagues. Salience rates were low, varying between 4-10%, and were lower with larger overdoses. Extreme overdoses fell into eight causal categories, each with a system design mitigation.

Conclusions: Novel analytic systems are required to accurately understand prescriber behavior and interactions with medication-dosing CDS. We described a novel analytic system that can detect apparent large overdoses (≥500%) and explain the sociotechnical factors that drove the error. Some of these large overdoses can be mitigated by system changes. EHR design should prospectively mitigate these errors.

Keywords: CPOE; Electronic health record; clinical decision support systems; electronic medical record; medical order entry system.

Fig. 1

Clinical workflow and its effect on data recorded for each alert. Orders are typically placed in batches and signed in one operation. Alerts fire at the time of signing. This diagram indicates how user actions are reflected in the data about the alert. Alert status is recorded as one of five categories in the system; (1) filtered (user is not automatically shown the alert), (2) removed, (3) cancelled, (4) overridden, or (5) viewed (typically a nursing action and not part of the prescribing process).

Fig. 2

Datamart schema. Multiple tables are joined, linking retrospective order and alert data to create a medication dosing alert analytic framework.

Fig. 3

Swim lane diagram of the CPOE dosing rules evaluation system. The top lane represents the production system and its data warehouse, which provides data to the study data mart. The lower lane represents the test system. The data analyst interprets output from both the data mart and the test system after running medication-dosing scenarios.

Fig. 4

Order-Alert rates. Rate chart of medication order alerts over the study period, by % visible alerts, % visible dosing alerts, and % visible overdose alerts in the Enter Orders phase of CPOE prescribing. Visible alerts are those alerts that are not filtered from view of the prescriber. The erratic behavior of this metric in the first 6 months can be attributed to discrepancies in order counts between datamart extracted data and the EHR utility-generated order counts. After some investigation, it was concluded that missing orders from the datamart are related to internal changes in how the EHR stores one category of the inpatient orders.

Fig. 5

Total medication dose alerts and large overdose alerts (≥500% overdoses) over the study period. The total alert rate remained relatively stable over the study perior, while the large overdose alerts decreased after October 2011, then stabilized.

Fig. 6

Salience rate of overdose alerts in the “enter orders” phase by user action. Salience rates are shown for all single-dose overdose alerts and for ≥500% overdoses. A single spike in salience rate was attributed to one user.

Fig. 7

Salience metric decreases with magnitude of overdose. As overdoses for a higher percentage over the upper dose range limit are ordered, users tend to override the alerts more often. This trend varies by therapeutic class, as shown by the contrasting patterns of salience for analgesics/anesthetics and anti-neoplastics (no variation in salience across the range of alerted high doses). Error bars were computed using the Clopper-Pearson method based on the binomial distribution.