Drug addiction in health care workers is a significant public health concern for anesthesiologists (1) and anesthesiology residents (2), possibly even more so than for their peers in other specialties. Documented diversion of controlled substances from the operating room (OR) by anesthesia care providers has a reported incidence of 1.0% for faculty and 1.6% for residents (3). The implication of this finding is that in a 3-yr training program with 13 residents per year, one would expect to discover a resident diverting drugs approximately once every 3 yr. In a recent survey, 19% of 111 responding program directors reported at least one fatality (overdose or suicide) in their training program between 1991 and 2001 (4). The initial recognition of the problem in 18% of cases of addiction occurred when there was either a fatal or near-fatal overdose (3), making earlier detection of particular importance.
Prevention of drug addiction in the field of anesthesiology has centered on education, awareness, and tightening of control policies for narcotic availability, but these efforts appear to be ineffective (3). The presence of automated drug dispensing systems and information management systems in the OR offers the promise of earlier and more efficient screening and detection of drug diversion compared with traditional manual methods (e.g., audit of dispensed drugs versus documentation in the anesthesia record, random testing of returned drugs, random urine testing). Because events recorded by the various systems to their databases include information as to the user, location, and a timestamp, an extremely detailed log of all drug- and case-related information is readily available for analysis.
We describe our experience in developing and applying such a system. Our approach should be useful to other departments interested in developing similar methods using their own information systems; sufficient details are provided for a skilled data analyst to be able to replicate our work.
Over the past several years, we identified two nonfaculty providers in our department who were diverting controlled substances in the OR. After the second case was uncovered, we decided to analyze the databases from our anesthesia information management system (AIMS), our pharmacy information management system (PIMS), and our OR information management system (ORIMS) to determine if an automated method could be developed to screen for a variety of potential drug-diverting subterfuges. A brief description of these two index cases follows:
The department chairman received a phone call from his counterpart at a nearby hospital reporting that a provider who worked at both institutions had been discovered diverting medications. To determine if similar activity had taken place at our hospital, a report was generated from the ORIMS listing the OR entry and exit times for all cases done by the provider for the previous 2 mo. A comparison to records from the PIMS indicated that the provider, on a daily basis, had been dispensing fentanyl from Pyxis machines at various locations to patients who were undergoing procedures at locations different from the site of these transactions. For example, on one day, drugs were withdrawn from the Pyxis machine in OR 32, where the provider was working, and assigned to patients in OR 8 and 15. The provider had not been involved in the care of the patients at the remote sites. There was no plausible explanation for this behavior other than diversion, which was acknowledged subsequently by the provider.
Changes in affect and reliability, and excessive absences were observed in an anesthesia provider by several members of the department and reported to the chairman, prompting concern of drug abuse. The provider’s anesthesia records for the previous month were reprinted from the anesthesia information system and examined, but no unusual patterns of scheduled drug administration were noted. The drug-dispensing records from the pharmacy system for the provider were also printed and manually compared with the anesthesia records. There were multiple instances in which fentanyl had been dispensed for cases in which the provider had no involvement, and where the quantity of vials removed in a single transaction appeared to be excessive. In several instances, the timestamp on the pharmacy records indicated that the patient had not been on the OR schedule on the date in question but, rather, had undergone surgery during the prior few days. The department’s substance abuse committee met to consider the evidence, and concluded that the provider had been diverting fentanyl. Arrangements were made for admission to a nearby inpatient substance abuse rehabilitation facility, and the provider was then confronted according to the established protocol in our department. The provider acknowledged the drug diversion and agreed to be admitted to the facility.
We used a data mining approach to explore potential relationships among data stored in our ORIMS (ORSOS, Per-Sé Technologies, Alpharetta, GA), AIMS (Innovian, Dräger Medical, Inc., Telford, PA), and PIMS (Pyxis, Cardinal Health, San Diego, CA) that might be useful for identifying scheduled drug diversion in the OR. Access™ 2003 (Microsoft, Redmond, WA) was used as the front-end database to perform the data exploration of the ORIMS Oracle® database (Oracle, Redwood Shores, CA) and the AIMS SQL Server™ database (Microsoft, Redmond, WA). Data were exported from the PIMS SQL Server™ database to a spreadsheet using PyxisConsultant™ (CardinalHealth, San Diego, CA) and then imported into Access. This indirect process was necessary because we were not granted access to the primary Pyxis database. Further data analysis was conducted and graphs were prepared using Excel™ 2003 (Microsoft, Redmond, WA). This study was approved by the IRB of Thomas Jefferson University.
We considered transactions involving withdrawal of opiates1 and midazolam from our Pyxis anesthesia drug dispensing units (CardinalHealth, San Diego, CA), one of which is located at each anesthetizing location. Access to a Pyxis cart requires individual user logon with a biometric fingerprint as authentication, and scheduled drugs are dispensed at the unit level (i.e., there are no open stock bins). As implemented, dispensed drugs can either be assigned directly to a patient, or dispensed to the provider and then subsequently assigned to the patient receiving the medications. Both inpatient and outpatient medical and surgical patients stay in the Pyxis active patient list until they are discharged from the hospital.
PIMS transactions were matched against the ORIMS and/or the AIMS databases. There were 21 mo of data in the first two databases, and 11 mo of data in the AIMS database, which is our most recent system. Details of the query process are provided in the Appendix.
There are a variety of atypical drug transactions that occur occasionally in the course of normal clinical care that might be exploited by a provider seeking to divert drugs for personal use. The occurrence of innocent atypical transactions creates a background level of noise, which is also contributed to by the method of matching records (Appendix). Our hypothesis was that we would be able to distinguish providers who were systematically diverting drugs via atypical transactions on a regular basis.
The goal of the data exploration was to develop a series of database queries, using the entire dataset, which would identify providers who entered into such transactions at an unusually high rate. We considered a query to be potentially useful for screening purposes if it clearly distinguished at least one of our index cases from the noise, and if the query, applied retrospectively on a monthly basis, indicated when the diverting behavior began. The frequency of atypical transactions by provider was normalized by dividing the transaction count by the number of days when the provider entered at least one transaction in the PIMS system. This normalization process permits comparison of full and part-time providers and considers absences due to vacation and meeting time, non-OR assignments, and rotations at other facilities.
Following is a list of potentially suspect transactions we considered for our analysis, with an explanation of why the transaction might indicate diversion as well as how such behavior might occur innocently. Database queries were developed using Structure Query Language (SQL) to extract records meeting the specified criteria, and are described using pseudocode for those interested in developing their own systems. In the SQL descriptions, the phrase “and normalize” means “divide by the number of days during which a provider executed at least one Pyxis transaction.” The queries are executed against a date range of the dataset (e.g., entire dataset, 1 mo, 2 mo).
- High Use of Opiates. Outliers in the use of such drugs might be diverting, but this could also result from being assigned to cases in which narcotic-based techniques are common (e.g., cardiac, neurosurgery). Such behavior might also reflect a practice pattern favoring balanced anesthetic techniques over the primary use of inhaled anesthetics. SQL: Compute average daily total use of fentanyl equivalents by provider.
- High Wastage of Controlled Substances. Diverters might be dispensing extra drug and then pretending to waste it in an appropriately labeled syringe containing a substitute fluid. Since the witness cannot assay the syringe contents, this artifice would evade easy detection. However, wastage is a normal part of clinical practice and providers are expected to have a certain amount of leftover drug that needs to be disposed of. SQL: Compute separately average daily total wastage of fentanyl equivalents and of midazolam by provider.
- Transactions on Cancelled Cases. Diverters might assign drugs to cancelled cases, which would not generate anesthetic records subject to random review by the pharmacy department. However, providers sometime draw drugs up in advance of cases that are subsequently cancelled. SQL: Compute total number of individual PIMS transactions executed on cancelled cases and normalize.
- Late Pyxis Transactions. A diverter might withdraw drugs on cases that had already ended on the day of surgery, or from a prior date, if the patient was still in the Pyxis system. These drugs would not appear on the printed anesthesia record, and thus not arouse suspicion in the postanesthesia care unit that excessive drug had been administered (e.g., 500 μg of fentanyl for a septoplasty). Legitimate transactions of this type might involve providers reconciling their administered and dispensed drug records after the end of the case, for example, if drugs originally assigned to the provider were then correctly allocated to the patient who received them. SQL: Compute total number of individual PIMS transactions executed after the end of the case and normalize.
- Mismatched Location Transactions (MLT), Mismatched Location Quantities (MLQ). A Diverter might be dispensing and assigning drugs from one OR to patients actually undergoing care in another OR and in whose care the provider was not involved. However, sometimes providers draw-up drugs from one workstation, after which the patient is moved to another OR. Also, the provider may be drawing-up his or her drugs in advance from another location as a matter of convenience. Finally, the incorrect room could have been specified in the ORIMS or AIMS system. A diverter might begin withdrawing more than the usual number of vials during mismatched transactions to satisfy an increasing need for more drug as a result of tolerance to its effect. Although our providers do dispense multiple vials at a time (e.g., if a narcotic-based technique is to be used), they more typically dispense a single vial of a controlled substance at a time. SQLs: Compute total number of individual PIMS transactions executed on workstations in an OR location different from where the case was actually performed and normalize. Compute total number of individual vials removed from workstations in an OR location different from where the case was actually performed and normalize.
A nonparametric, graphical approach rather than a formal statistical analysis was performed because of the small number of index cases and the relative infrequent nature of atypical transactions (i.e., we were trying to find outliers among a distribution of outliers). The purpose of the screening process was not to prove statistically that diversion was taking place but, rather, to identify providers who would undergo a more detailed audit of their dispensing and administration behavior to determine if there was reason for concern. We generate a report that lists all cases that the provider was involved in, the procedure performed, the start and end time of the case, the timestamp and amount of each scheduled drug administered during the case, and the Pyxis log of all drug transactions by the provider during the specified data range. We then attempt to reconcile the Pyxis transaction logs with the AIMS records. For patients with scheduled drug transactions executed by the provider in whose care the provider does not appear to have been involved, we reprint those anesthesia records in their entirety to determine if there is any other indication (e.g., through a comment), that the provider was actually present at some time during the case.
Continuous variables are presented as mean ± sd.
High Use of Opiates
We calculated from the Innovian anesthesia records the average daily administration of opiates available in our Pyxis machines, converted to fentanyl equivalents as follows: sufentanil 5 μg: 50 μg, morphine 1 mg: 50 μg, We arbitrarily considered outliers to be providers who used more than the 90th percentile of fentanyl equivalent (1050 μg) on a given day on more than 33.2% of days that they were in the OR (the 95th percentile for the fraction of days among all providers where daily use was above 1050 mg). The distribution of daily use was lognormal, and examination of the anesthesia records of the five outliers identified did not reveal any apparent irregularities. Neither of the index cases was identified by this query.
No outliers were detected looking at the average daily wastage of fentanyl equivalent (101 ± 48 μg) or midazolam (0.5 ± 0.4 mg). All providers wasted, on average, <250 μg fentanyl per day, and <1 mg of midazolam per day.
Transactions on Cancelled Cases
There were no outliers found for Pyxis transactions on cancelled cases using either the AIMS or ORIMS databases.
Late Pyxis Transactions
Both index cases were identified using the ORIMS data to determine the end of case by this query, which looked at Pyxis transactions executed after the case ended (Fig. 1). The more recent index case was also identified using the AIMS data as the end time (not shown). When the query was applied on a monthly basis, the two index cases were also evident using either case end data from the AIMS (Fig. 2) or ORIMS databases (Fig. 3). (Providers working <6 days during the month were excluded, since inspection of such records revealed magnification of a small number of errors by the small denominator.) The anesthesia records and Pyxis logs of the two index cases’ anesthesia records and Pyxis transaction logs were retrospectively audited for the month when the provider was first identified as an outlier. It was clear that each had diverted drugs on multiple occasions.
An example of such an audit follows: A patient underwent a parathyroidectomy from 14:32 until 16:34. Fentanyl, 250 μg, and midazolam, 2 mg, had been drawn-up by the provider just before the case and were charted on the anesthesia record. However, between 21:58 and 21:59, over 5 h after the patient left the OR, the provider returned to the same Pyxis machine and withdrew fentanyl, 250 μg, morphine, 10 mg, and midazolam, 2 mg, assigning the drugs to the patient who had undergone the parathyroidectomy earlier that day. Examination of the call schedule revealed that the provider was on call that night.
When we extended the definition of “after the end of the case” from any time after the case ended to more than 60, 120, or 240 min later using the AIMS system as the data source, there was no apparent increase in sensitivity, as the ratio of the provider with the highest frequency of late transactions (always the index case) to the next highest provider remained approximate the same (mean 4.8 ± 0.6, range 4.2–5.6).
MLT and MLQ
Both queries identified the second index case using the AIMS database as the source of the actual patient OR location (Fig. 4) and both also identified the first index case using the ORIMS database for this location (not shown). The higher value of the MLQ compared with that of the MLT reflects the fact that the provider was withdrawing multiple vials during each transaction (e.g., sometimes three or four vials of fentanyl at a time). When applied monthly using the AIMS database as the data source, the query identified the onset of diversion using both the MLQ (Fig. 5) and the MLT (not shown). (Providers working <6 days during the month were excluded for the analysis for the same reason noted above.)
We successfully formulated several queries using the PIMS database and either the AIMS or ORIMS database that identified our two index cases involving diversion of opiates and midazolam. We did not discover any other providers who appeared to be diverting during the period of the data analysis. Despite the presence of “noise” in the database arising from innocent transactions and errors in matching cases, the outliers were easily identified from the graphs. Had we been prospectively applying these queries, we would have detected the diversion several months earlier than when they were actually discovered. We do not know the optimal interval over which to execute the screening queries, although too short an interval will likely increase the level of noise and obscure the signal. Too long of an interval may decrease the sensitivity of the test by including a period of time when a provider had not yet started diverting drugs. We currently run our queries looking at the previous 1, 2 and 4 mo of data. We hypothesize that the longer intervals may improve the sensitivity for frequencies that occur just above the rest of the group for several months in a row. The rationale is that the other providers’ frequencies of atypical transactions will regress toward the mean over time, whereas the diverter’s frequencies will likely increase.
A limitation of the methodology we developed is that we only had two index cases, and have not detected anyone else diverting drugs in our department over the past 8 mo. We currently are running monthly reports based on the screening queries described in this manuscript, as well as some additional queries whose usefulness we have not yet determined. Applying these queries to data from other hospitals with PIMS and either ORIMS or AIMS systems, and being able to identify their known diverters in a blinded fashion, will be necessary to validate the methodology more fully.
Another weakness of our detection logic is that there are methods of diversion that we have not considered or which may not be easily uncovered based on examination of the databases. We choose not to enumerate these methods so as to avoid providing insight to drug diverters about techniques that might escape detection. No claims are made that our procedures represent a definitive approach to the detection of drug diversion in the OR, and more complicated approaches looking at diversionary behavior will be necessary to uncover more sophisticated diverters (e.g., looking at the distribution of time intervals between successive controlled substance transactions, examining the frequency of transactions from locations at unexpected times of the day). The difficulty in this effort is that no single anesthesia program has had a sufficient number of providers diverting drugs to identify all the different schemes that might be used. We therefore view our work as a first step in what we hope will be a collaborative effort to improve detection of drug diversion in the OR.
For hospitals with PIMS that are not present in each OR, but rather in central locations, the query looking for mismatches between OR dispensing and administration locations will not work. In this circumstance, records could be identified by an SQL procedure finding transactions in which the person dispensing the drug is not involved in the patient’s care. We choose not to try this approach because such data are not recorded in the ORIMS for at least 30% of our cases, and because providers involved inpatient care do not always remember to log their names into the AIMS. This level of noise would have overwhelmed the signal in our dataset.
The applicability of the methods of detection described is limited to facilities having a PIMS and either an AIMS, ORIMS, or some other method of recording case information that can be retrieved electronically (e.g., a manual OR log that is subsequently entered into a spreadsheet). In addition, a skilled analyst who understands how to query databases must be available.
One caveat that users of automated anesthesia drug dispensing systems need to consider is that transactions executed are attributed to the person logged in to the machine. Thus, during breaks, if the primary provider stays logged on and the relief person withdraws drugs, the audit trail will point to the primary provider in the event that diversion takes place. This discrepancy might be resolvable by detecting a pattern of such transactions, but this would be a time-consuming forensic process. We suggest that manufacturers of automated anesthesia drug dispensing systems provide an option to require reauthentication whenever scheduled drugs are dispensed. This would greatly tighten up what is currently a security loophole in the system, as all controlled substances would then be attributable to the actual person withdrawing the drug.
In addition to the screening methods we have described, our pharmacy department runs other controlled substance reports on a regular basis that may detect diversion. For example, they look at providers who have dispensed drugs to themselves but not assigned them to patients and ask them to reconcile their records. We believe this serves as a strong deterrent for this obvious type of diversionary behavior, as these reports are also sent to the department’s executive vice chairman. Anesthesia records are also randomly audited by our pharmacy department, but the effectiveness of this in detecting diversion is unknown.
We stress that the methodology described is meant only to be a screening tool, and should not be used as the sole basis to support a confrontation. A manual inspection of the actual drug transaction logs and the anesthesia records is required to evaluate the likelihood that the behavior represents diversion. We feel it is important that providers who appear to be outliers not be asked casually to explain apparent discrepancies, but rather that the same preparatory steps for confrontation and intervention be put in place as is the case when definitive evidence of diversion is present. If the evidence that diversion is taking place is equivocal, a heightened level of surveillance could be instituted for that provider. Also, the fact that an individual provider is being audited should be kept confidential. It is possible that innocent providers may occasionally be subject to scrutiny, but we feel that sacrificing some degree of specificity for sensitivity is important for the earliest detection of diversion. When an audit indicates that no diversion took place, we destroy the records related to that investigation and delete any related files on the computer. Full records, however, are kept for individuals found to have been diverting.
In summary, using data mining techniques, we developed several queries that, when applied retrospectively against the databases of our anesthesia drug dispensing system and either our AIMS or ORIMS, identified drug diversion in two known cases. We consider this approach to be potentially useful for earlier detection of drug diversion than is possible based on observation of behavioral changes deemed suspicious of abuse. Widespread adoption of this methodology, details of which are provided in the manuscript, requires validation at other facilities and with other logic systems to detect additional methods of diversion. Better and earlier detection of drug diversion may help prevent the major morbidity and mortality associated with addiction to opiates by anesthesia providers, and the regular application of methods screening for diversion may serve as a disincentive to those with addictive personalities from entering anesthesiology.
APPENDIX: TECHNICAL DETAILS OF QUERY PROCESS
Details of the database query process are provided below for the benefit of data analysts who may wish to replicate our methodology.
Because PIMS records did not contain a field that could be directly linked reliably to cases in the AIMS or operating room (ORIMS databases, a hash was constructed using the patient’s name (Last Name & initial two letters of First Name) and the date of transaction (mm/dd/yyyy) so that records could be aligned (e.g., EPSTEINRI01/01/2007). We did not use the medical record number as part of this hash because this was not always present in the PIMS, whereas the patient’s name was always listed. This method creates a small inaccuracy for the small number of our patients who undergo multiple separate operations on the same day (0.3% in the AIMS database), and for patients whose procedure started before midnight and ended the next day (0.7%). Among the over 31,000 records in the database, there were no unique patients with identical hashes, although this can occur as a very rare event.
Our method of creating a hash based on elements of the patient’s name and the date of surgery to link records among the various systems is not the only way that this can be done. Others may find it more convenient to use the patient’s medical record, social security number, or account number in association with the date of surgery.
The following data were extracted from the databases for use in the exploratory data analysis.
- PIMS: User, drug name, amount withdrawn, patient name to whom the drug was assigned, provider name executing the transaction, location of pyxis station, transaction date and time
- AIMS: Date of surgery, patient name, start and end of the intraoperative anesthesia record, OR location
- ORIMS: Date of surgery, patient name, enter procedure room time, leave procedure room time, OR location
We chose to use the start and end of the printed intraoperative record times in our AIMS as surrogates for the case start and end time, respectively, for our analysis. The printed record events are present on every AIMS record, being automatically applied by the system, whereas users sometimes forget to enter other events marking the case duration. Other case start and end times from the AIMS can be used without altering the results, e.g., enter OR and leave OR.
1. Gallegos KV, Browne CH, Veit FW, Talbott GD. Addiction in anesthesiologists: drug access and patterns of substance abuse. QRB Qual Rev Bull 1988;14:116–22
2. Talbott GD, Gallegos KV, Wilson PO, Porter TL. The Medical Association of Georgia’s Impaired Physician Program. Review of the first 1000 physicians: analysis of specialty. JAMA 1987; 257:2297–30
3. Booth JV, Grossman D, Moore J, Lineberger C, Reynolds JD, Reves JG, Sheffield D. Substance abuse among physicians: a survey of academic anesthesiology programs. Anesth Analg 2002;95:1024–30
4. Collins GB, McAllister MS, Jensen M, Gooden T. Chemical dependency treatment outcomes of residents in anesthesiology: results of a survey. Anesth Analg 2005;101:1457–62
1Fentanyl, sufentanil, morphine, and remifentanil are stored in our anesthesia Pyxis carts. Meperidine and hydromorphone are not available in these units.© 2007 International Anesthesia Research Society