Surveillance of poisoning and drug overdose through hospital discharge coding, poison control center reporting, and the Drug Abuse Warning Network

Surveillance of Poisoning and Drug Overdose Through Hospital Discharge Coding, Poison Control Center Reporting, and the Drug Abuse Warning Network PAUL D. BLANC, MD, MSPH,*$§ KENT R. OLSON, MDt$§

MATTHEW R. JONES, BS,+§

There is no gold standard for determining poisoning incidence. We wished to compare four measures of poisoning incidence: International Classification of Diseases 9th Revision (ICD-9) principal (N-code) and supplemental external cause of injury (E-code) designations, poison control center (PCC) reporling, and detection by the Drug Abuse Warning Network (DAWN). We studied a case series at two urban hospitals. We assigned ICD-9 N-code and E-code classifications, determining whether these matched with medical records. We ascertained PCC and DAWN system reporting. A total of 724 subjects met entry criteria; 533 were studied (74%). We matched poisoning N-codes for 278 patients (52%) E-code by cause in 306 patients (57%), and E-code by intent in 171 patients (32%). A total of 383 patients (72%) received any poisoning N-codeor any E-code. We found that PCC and DAWN reporting occurred for 123 of all patients (23%) and 399 of 487 eligible patients (82%), respectively. In multiple logistic regression, factors of age, hospital admission, suicidal intent, principal poisoning or overdose type, and mixed drug overdose were statistically significant predictors of case match or report varying by surveillance measure. Our findings indicate that common surveillance measures of poisoning and drug overdose may systematically undercount morbidity. (Am J Emerg Med 1993;11:14-19. Copyright 0 1993 by W.B. Saunders Company)

Poisoning and drug overdose are major causes of morbidity and mortality in the United States, with temporal trends suggesting that the number of such injuries may be sharply increasing. I-3 Despite the importance of this problem, epidemiologic study of poisoning and drug overdose has been hampered by lack of reliable surveillance measures. Hospital discharge data bases using International Classification of Diseases, 9th Revision (ICD-9) coding have pro-

From *the Division of Occupational and Environmental Medicine, Tthe Schools of Pharmacy and Medicine, and *the Injury Prevention Research Center, University of California San Francisco, San Francisco, CA, and @he San Francisco Bay Area Regional Poison Control Center, San Francisco, CA. Manuscript received April 15, 1992; revision accepted August 21, 1992. This study was supported in part by Grant No. R49CCR903697 from the National Center for Injury Prevention and Research, Centers for Disease Control, Atlanta, GA. Presented at the 1991 meeting of the American Association of Poison Control Centers. Address reprint requests to Dr Blanc, Box 0924, University of California San Francisco, San Francisco, CA 941430924. Key Words: Poisoning, drug overdose, injury surveillance, poison control, International Classification of Diseases 9th Revision, external cause of injury. Copyright 0 1993 by W.B. Saunders Company 07356757/93/l 101-0005$5.00/O 14

vided an important tool for disease surveillance. Discharge coding has been made more attractive as an injury surveillance measure by the increasing use of supplemental External Cause of Injury codes (E-codes).4-6 Nonetheless, potential shortcomings in general injury surveillance relying on such coding are recognized, although studies focusing specifically on the accuracy or effectiveness of poisoning and drug overdose case capture using ICD-9 coding (N-coding) have not been reported to our knowledge. As a more specific measure, poison control center (PCC) case reporting through the American Association of Poison Control Centers has provided one of the most widely used indices of poisoning incidence. Unlike hospital discharge coding, the PCC system includes a much wider population for potential surveillance, including cases treated at home without visiting a health care facility. In 1990, over 1.7 million poisoning cases were reported through this system, with 25% of these of sufficient severity to be treated in a health care facility, although approximately 20% of the US population is not served by a participating PCC.3 Furthermore, little is known about those cases that are not detected through PCC surveillance, and the potential biases that may contribute to underreporting in that system.’ Similarly, the Drug Abuse Warning Network (DAWN) system of the National institute on Drug Abuse has provided another incidence measure, based on an active surveillance program using a sampling scheme involving 770 hospital emergency departments in 21 metropolitan areas.’ Eligibility for reporting through that system is not limited to drugs of abuse, also including emergency treatment of adverse effects attributed to prescription and over-the-counter medications. Once again, its potential reporting biases have not been well delineated. We performed this study to examine the detection of poisoning and drug overdose at the level of hospital-based reporting by each of these four potential surveillance measures: principal N-coding, supplemental E-Coding, PCC reporting, and reporting through the DAWN system. We hypothesized that surveillance dependent on each of these measures might have inherent biases affecting case detection and resulting in a differing incidence data reflecting the approach taken. Because there is no “gold standard” for poisoning and drug overdose incidence, we directly assessed emergency room visits, tracking cases to determine the N-codes and E-codes that they were assigned by routine medical recording procedures and to assess whether the cases were reported either to the PCC or through the DAWN system.

BLANC ET AL n SURVEILLANCE

OF POISONING

15

METHODS

Medical Record Review

Study Sites

We requested complete medical records for review in all subjects at each hospital site meeting entry criteria. We extracted from the medical record face sheet the subject’s age, sex, length of stay (if admitted), and (if present) the ICD-9 codes present and the E-coding assigned by the medical records department to the case. We noted only the first four assigned codes, except for the E-code which we extracted regardless of the number of codes present in the record. We also noted the relative rank of the E-code assigned. We then reviewed the body of the medical record, in particular the medical admission and discharge notes. Based on this information, we independently assigned the case a principal poisoning or drug overdose ICD-9 N-code and an E-code, including a determination of intentionality.’ Following accepted coding guidelines. we classified as “accidental” all unintended overdose cases related to illicit drug use, even though the abuse itself can be considered intentional. We also noted from the medical record whether the subject had been treated in an intensive care unit. We identified all subjects in whom multiple drug ingestion as a part of the reported overdose had occurred. In those cases, each subject’s vital signs and other relevant diagnostic information, including drug levels when available, were presented to one of the authors (P.D.B.) for review, who determined a principal ICD-9 code without knowledge of that assigned on the medical record. A similar approach was undertaken for combination drug products (eg, acetaminophen with opiate) although ingestions of such formulations were not considered multidrug overdoses per se. We also did not consider as multidrug overdose ingestions of ethanol along with a single drug formulation.

We studied poisoning and drug overdose cases from each of two hospitals. Hospital One is a 582-bed, urban public hospital with approximately 73,000 emergency department visits yearly. Hospital Two is a X&bed university teaching hospital with approximately 22.000 emerg,ency department visits each year. Both hospitals are served by the same medical house staff training programs.

Case Selection At Hospital One. we manually reviewed all adult emergency department case record cover sheets, which included patient identifiers and a brief narrative describing the principal diagnosis completed by the treating physician. We considered eligible for study all subjects with a principal emergency department diagnosis on this cover sheet of presumed or suspected drug overdose or poisoning (including symptomatic toxic chemical exposures). We did not study cases where the principal diagnosis was worded as “adverse drug effect or reaction,” “drug abuse,” “drug-seeking behavior,” or cases where complications of drug use were listed (for example. infections secondary to illicit intravenous drug abuse). We also did not study cases with a principal emergency department diagnosis of ethyl alcohol intoxication. We adopted this approach to case selection to limit study to those cases in which toxic or supratherapeutic drug effects were suspected by the treating medical staff, under the assumption that such cases would be less ambiguous for medical records coding purposes. In adopting these restrictive criteria we realized that some of the “drug abuse” and other cases that we excluded may indeed have manifest toxic drug-related syndromes. We supplemented our review of the emergency department case cover sheets with review of the adult emergency department case admission log, which included only those cases that had been admitted as inpatients. We used similar case selection criteria based on the principal diagnosis listed in that log. Our rationale for review of this supplemental source for case finding was the observation that the emergenc:y department case cover sheets tended to inadvertently follow admissions to the inpatient wards and thus be missing from the cover sheet files that we reviewed for subject identification. Because it was standard procedure in the medical record department of Hospital One to apply routine ICD-9 N-coding and E-coding to most, although not all, nonadmitted emergency departmenttreated cases, we did not limit subject eligibility at that site to hospital-admitted cases only. We entered all subjects meeting study criteria who were seen at the adult emergency department at Hospital One between January I and June 30, 1990. At Hospital Two we used a different case identification strategy with similar goals to that used in Hospital One. At Hospital Two we identified all emergency department-treated cases with a principal diagnosis of poisoning or drug overdose first by searching a computerized emergency department data base maintained independently of the hospital medical records department (using the key “poisoning.” “intoxication,” “inhalation,” words “overdose.” “abuse,” and “drug effect” ), and then by applying the selection criteria used at Hospital One to the principal diagnosis as worded in the data base. For example, a diagnosis that appeared in the data base as “drug abuse cocaine injection” would not have been eligible for entry; in contrast, “drug overdose, cocaine injection” would have been. The rationale for this restrictive approach, as with Hospital One, was in an attempt to exclude cases in which coding error would more likely be due to ambiguity between toxic and “therapeutic” drug effects. Because the Hospital Two medical records department carried out ICD-9 N-coding and E-coding only for cases admitted to hospital, we studied only admitted cases at that site. We entered all subjects meeting the criteria admitted over 15 months between October 1. 1989 and December 31, 1990.

Poison Control Center and Drug Abuse Warning Network Reporting After examination of the medical records, we manually reviewed case logs of the San Francisco Bay Area Regional PCC, identifying cases matching study subjects either by name or, if no name was noted in the PCC log, by ingestion, date, and hospital location. The PCC serves a multicounty region including the city in which the study hospitals are Iccated. Moreover. the PCC is physically located within the emergency department of Hospital One. We also reviewed the DAWN case log for Hospital One only, using matching criteria based on age, sex, and drug type. The DAWN record in that hospital is managed by PCC staff, but case identification in that system is unrelated to the routine PCC consultation activities reflected by the PCC case log. Although DAWN reporting is also done by Hospital Two. its record logs were incomplete for the study period and were not used in this study for that reason.

Data Analysis Data were key entered and analyzed using a standard computerized statistical package. We studied predictors of subject identification by each of four coding or reporting methods: ICD-9 N-code or E-code, PCC report, or DAWN report. We defined successful N-coding as a match between the N-code that we assigned upon chart review with that assigned in the medical record either as the primary or one of the first three secondary diagnostic codes given. We also examined a more broadly defined matching criterion of N-code match independent of diagnostic accuracy or specificity. Under those broader criteria we accepted as “matching” any drug or poisoning ICD-9 N-code (codes 960-989.9) appearing as any of the first four codes assigned in the medical record. We defined successful E-coding as a match between the E-code we had assigned and that assigned by medical records. We considered E-codes matched regardless of the rank among secondary diagnoses of the medical record assigned E-code; if more than one E-code had been assigned by medical records we chose the best match. In addition to overall E-code match, we further noted whether matching was present

16

AMERICANJOURNAL OF EMERGENCY MEDICINEn Volume 11, Number 1 n January 1993

based at least on E-code substance or E-code intentionality ( “accidental, ” “suicide or intentional,” “assault by poisoning,” or “intentionality undetermined” ). We defined successful PCC reporting as the presence of a matching case in the PCC log and similarly defined successful DAWN reporting as a matching logged case in that system (Hospital One only). We analyzed predictors of subject identification in each of the four systems using multiple logistic regression. We included a priori in the logistic model the following variables: hospital site studied, age, sex, hospital admission (compared with emergency department treatment without admission), suicide (compared with accidental or assault), multidrug overdose (yes or no), and principal drug or exposure causing the poisoning or overdose (for example, antidepressant). We also repeated the multiple logistic regression limited to hospitalized cases, adding intensive care treatment to the predictive model. For principal drug cause we assigned dummy variables (1.0) for each of the following common substances (defined as 1 5% frequency in our study population): opiates, benzodiazepines, antidepressants, cocaine, phencyclidine (PCP), and acetaminophen. As noted previously, if more than one substance had been ingested, one and only one (mutually exclusive) drug cause received a “I” value as a dummy variable. The comparison (default) category for poisoning and drug overdose cause included all other, less common causes of poisonings and drug overdose among our subject series. Although the principal drug causes that we assigned were mutually exclusive, we were concerned about the potential impact of collinearity with other predictors (eg, antidepressant as the cause of poisoning and suicidal intent). We analyzed both of our predictor models (all subjects, “model one” ; hospitalized subjects only including intensive care treatment in the model, “model two”) according to the criteria for multicollinearity outlined by Sinker and Glantz.” The highest correlations between predictors were 0.54 for hospital site studied and hospital admission (model one, consistent with design-limiting study to admitted subjects at Hospital Two) and - 0.45 for opiate cause and suicidal intent (model two). The lowest eigen value and the ratio of highest to lowest eigen values was 0.25 and 10.1 (model one) and 0.34 and 5.5 (model two). These findings do not suggest that major collinearity is present in these models.

RESULTS We identified 724 subjects eligible for study. We were unable to locate medical records for 46 (6%) of the cases. Upon retrieval of medical records, we found that no ICD-9 coding at all had been assigned in 145 (20%) of the cases. All but one of these were among cases treated and released from the emergency department of Hospital One, consistent with that hospital’s medical record policies regarding uninsured outpatients. We excluded from further analysis these subjects. After lost records and exclusions 533 subjects (74% of those eligible) remained for final study. Demographic and other descriptive data are presented in Table 1. Overall, two thirds of the cases were male, with a median age among the subjects of 32 years (90% of the subjects ranged between 17 and 53 years, inclusively). The extremes of age were uncommon: five (1%) of the subjects were under age 14 years while eight (2%) were over age 65 years. Although Hospital Two contributed only 9% of the subjects studied, because these were limited to hospital admissions those 46 subjects represented 35% of the admitted cases studied. Among admitted subjects intensive care treatment was common, occurring in one out of three cases, while length of hospital stay was, in general, brief. We noted only three (0.6%) deaths among our subjects despite suicidal intent among 48% of the subjects.

TABLE 1.

Poisoning

and Drug Overdose

by Type Among

533 Subjects Case Type

Frequency

Hospital location Hospital One, municipal public hospital, n (%) Hospital Two, university hospital, n (%) Male n (%) Age (y)’ median, (90% range), (n = Admitted Treated n (% Length

519) to hospital, n (%) in intensive care unit, of admits) of stay in days, median

(range) Suicidal intent, n (%) Principal poisoning by drug or exposure cause Opiate n (%) Benzodiazepines Cocaine n (%) Phencyclidine n Antidepressants Acetaminophen All other causes Mixed drug overdose l

Excludes

487 (91) 46 (9) 349 (66) 32 (17 to 53) 131 (25) 44 (34) 2 (1 to 62) 254 (46)

179 (34) n (%) (%) n (%) n (%) n (%) n (%)

14 subjects

for whom

50 39 39 30 26 170 i ia

(9) (7) (7) (6) (5) (32) (22)

age was not available.

As shown in Table 1, the principal specific poisoning or drug overdose cause was opiates, including both illicit drugs (heroin) and controlled prescription medications such as oxycodone. The remaining common causes of poisoning we identified included illicit drugs (cocaine and PCP), other controlled substances (benzodiazepines), prescription medications (antidepressants), and acetaminophen. No other single agent among the remaining 170 (32%) of the subjects accounted for 5% or more of the cases. Although in all subjects we assigned a principal cause of poisoning consistent with the clinical presentation and management, mixed drug overdose was noted among 118 (22%) of the subjects. Table 2 delineates the success of case detection by each of four systems that may serve as potential surveillance measures for poisoning and drug overdose. The systems appeared to have variable success at case identification. We identified a matching poisoning and drug overdose principal N-code in just under half of the subjects; broadening criteria to include secondary diagnoses, we found a match in 278 (52%) of the subjects. Regardless of accuracy, 361 (68%) of subjects had been assigned any N-code consistent with poisoning or overdose (N-codes 960-989.9). When analyzed by intent E-coding appeared to provide a rather poor match, in part because of a tendency in medical records to assign “intent not determined” codings, especially in cases involving drugs of abuse. Medical records E-code assignment by substance cause matched more frequently with the E-code that we assigned. Although in 327 (61%) of all subjects an E-code of some sort had been assigned by medical records, in 15 (5% of those with any E-code) the E-code assigned was for an external cause of injury unrelated to poisoning, such as fall or laceration. As shown in Table 2, we found that either E-coding by

BLANC ET AL n SURVEILLANCE

TABLE2. Poisoning Existing

OF POISONING

and Drug Overdose

17

Detection

by

Because we were particularly interested in the N-codes that these subjects did receive we studied them further. Of the 150 subjects, 124 (83%) received N-codes consistent with drug dependence or drug abuse (304.0-305.9) even though these codes are not applicable to acute toxic effects. We also examined the relative rank of the E-code among the assigned classifications. An E-code was listed among the first four diagnostic codes for 342 (64%) of the subjects; for 31 (6%) subjects an E-Code was listed but appeared fifth or greater among listed codes. In 160 (30%) of the subjects no E-code at all had been assigned, regardless of its rank. Also shown in Table 2, PCC reporting was the poorest case detection measure of those we examined, with only 123 subjects (23%; 95% confidence interval [CI] 20% to 27%) successfully identified. In contrast, DAWN reporting (studied at Hospital One only) demonstrated the greatest proportional case detection. Table 3 presents the results of multiple logistic regression analysis of potential predictors of “successful” case detection (defined as matching coding or a matching case report) in each of the four systems. For the purposes of this modelling, we considered N-code matching by either primary or secondary codes and E-Code matching by substance regardless of intent as “successful.” For N-code, E-code, and PCC the logistic model also included the hospital site studied among the predictor variables; for the DAWN system this was omitted because that analysis was limited to Hospital One only. Table 3 illustrates the varying relationship between the predictors studied and each outcome surveillance measure. In the multiple logistic model, sex was not statistically significantly related to surveillance success under any of the four systems. In contrast, age was significantly and inversely related to matching by either N-code or E-Code. Although hospital admission was not a statistically significant predictor of ICD-9 coding match, it was a positive predictor of poison center reporting (odds ratio [OR] = 3.0, 95% CI, 1.6 teria.

Surveillance Case Detection (95% Confidence

Surveillance

Measure

” (%)

Medical record N-code and E-code assignment match with study N-code match by principal cause of illness N-code match by principal or secondary cause Any poisoning or overdose N-code despite accuracy E-code match by intent E-code match by cause E-code match by cause or intent Any E-Code assigned regardless of match E-code match by cause or primary/ secondary N-code match Any E-code or any poisoning or overdose N-code Poison control center case report Drug Awareness Warning Network case report* l

DAWN eligible

IIlterval)

245 (46) (42% to 50%) 278 (52) (48% to 56%) 361 171 306 327

(68) (32) (57) (61)

(64% (28% (53% (57%

to to to to

72%) 36%) 63%) 66%)

373 (70) (66% to 74%) 317 (59) (55% to 63%)

383 (72) (68% to 76%) 123 (23) (20% to 27%) 399 (82) (79% to 85%)

cases, n = 487, Hospital

One only.

cause or N-coding matched for 317 (59%) of subjects. Further broadening potential surveillance criteria to include any poisoning or drug overdose N-code or any E-code (regardless of accuracy), 383 (72%) of the subjects could have been identified. .Analyzing overlap among these broadened dellnitions, E-coding captured 351 of 361 (97%) N-codeidentified subjects, while N-coding captured 351 of 373 (94%) E-coded subjects. However, stated in other terms, 150 (28%) of the subjects would not have been identified by either approach even applying such expanded surveillance cri-

TABLE 3.

Multiple

Logistic

Regression

Modeling

of Predictors

of Surveillance

Success

Surveillance E-code

N-code

Predictor Male Age* Admission Suicide Agent or cause Opiates Benzodiazepines Cocaine PCP Antidepressants Acetaminophen All others Mixed Overdose

POlSOil Center

OR

(95% Cl)

OR

0.8 0.8 1.2 1.6

(0.5, 1.2) (0.6, 0.9) (0.7, 1.5) (0.98, 2.7)

0.9 0.7 0.9 3.0

(0.6, 1.5) (0.6, 0.9)

0.9 1.5 0.3 0.5 4.7 0.4 1.0 0.8

(0.6, 1.6) (0.7, 3.1) (0.1, 0.7) (0.2, 1.2) (1.4, 15.2) (0.2, 1 .l) -

1 .o 0.9 0.3 0.3 4.1 1.3 1.0 0.7

(0.6, 1.7) (0.4, 1.8) (0.1, 0.8) (0.2, 0.9) (1.1,15.3) (0.4, 4.2) -

(0.5, 1.3)

NOTE. Except for DAWN, n = 519; model also includes ABBREVIATIONS: N-code, match by primary or secondary regardless of intention, OR, odds ratio. Risk expressed per 10 years of age.

l

Measure

hospital principal

(95% Cl)

(0.5, 116) (1.8, 5.1)

(0.4, 1.2)

OR

DAWN OR

(95% Cl)

1.1 0.9 3.0 4.1

(0.5, (0.7, (1.6, (2.1,

0.2 0.4 0.3 0.2 0.4 0.6 1 .o 1.8

(0.1, 0.6) (0.2, 0.8) (0.1) 1.2) (0.002, 1.3) (0.2, 1 .O) (0.2, 1.5)

1.4) 1.1) 5.5) 7.9)

(1 .l ,.o,

site studied. For DAWN, n = 473. ICD-9 code; E-code, match by external

1.5 0.9 0.2 1.9 3.3 1.0 2.2 7.0 2.5 11.5 1.0 0.8

(95% Cl)

(0.9, (0.7, (0.1, (0.9,

2.6) 1.2) 0.3) 3.7)

(1.5, 7.0) (0.4, 2.5) (0.8, 6.0) (1.4, 33.7) (0.8, 7.9) (1.4, 97.2) (0.4, 1.4)

cause of injury code by agent,

AMERICAN

18

JOURNAL

to 5.5) and, in contradistinction, a negative predictor of successful DAWN case detection (OR 0.2, 95% CI, 0.1 to 0.3). In each of the four systems, suicidal intent was positively associated with detection when compared with nonintentional cases (although we included in the nonintentional category the one case of assault in our series). This association was strongest and statistically significant for E-code match and for PCC reporting. When analyzed by principal drug cause as shown in Table 3, we observed a number of drug-specific reporting effects. Cocaine was significantly associated with match failure and antidepressants with matching success for both the N-code and E-Code measures. For PCC reporting, each of the common principal causes of drug overdose and poisoning relative to the reference comparison of “all other” (less common) causes was associated with risk of underreporting, although this achieved statistical significance at the 0.05 level for opiates and benzodiazepines only. Except for benzodiazepines, DAWN reporting demonstrated an opposite pattern. Mixed drug overdose was positively associated with detection success for poison control center reporting only. Despite the geographic proximity of the poison control center to the emergency department of Hospital One, this site was not associated with increased poison center reporting in the multiple logistic analysis (OR = 0.7; 95% CI, 0.3 to 1.6). To study case severity of illness as a predictor of reporting, we reanalyzed the logistic regression using the same predictors (except for hospitalization) in the model while adding intensive care unit admission as a potential predictor and limiting the analysis to hospitalized cases only. These data are presented in Table 4. Although intensive care treatment as compared with nonintensive care-treated hospitalized subjects were more likely to have a matching poisoning diagnostic E-code and be reported to the PCC or through the DAWN system, these associations did not achieve statistical significance. DISCUSSION Our findings suggest surveillance by standard measures is problem ridden. However, these data must be viewed within the context of our study limitations. First and foremost, caution is warranted in generalizing our findings, based as they are on only two hospital sites for PCC and ICD-9 coding and just one site alone for DAWN reporting. Our case mix is weighted towards exposure to drugs of abuse with an adult predominance in the study population, a clear potential for TABLE 4. Surveillance

Surveillance Measure N-code E-code PCC DAWN

Intensive Care Treatment Success

Subjects Model (n) 130 130 130 84

in

as a Predictor

Surveillance Success n (%) 90 96 61 47

(69) (74) (47) (56)

of

Intensive Care as a Predictor of Detection OR (95% Cl) 0.9 3.4 2.1 1.8

(0.3, (0.9, (0.9, (0.6,

2.9) 13.6) 5.0) 5.6)

NOTE. Multiple logistic regression model includes all predictors included in Table 3 except hospitalization. Hospitalized cases only studied, one case excluded due to missing data.

OF EMERGENCY

MEDICINE

n Volume

11, Number

1 n January

1993

error were these findings to be extrapolated to other sites with a different patient mix. Our regional PCC may be more closely tied to both ED operations and DAWN detection than at other sites. Moreover, the hospitals involved are large, university-affiliated urban institutions with patient referral patterns and illness severity distributions reflecting these attributes. We stated at the outset that there is no “gold standard” for the true incidence of poisoning and drug overdose and reiterate that this is critical methodologic issue. We would not claim that our review of medical records could serve as such a standard. To evaluate the success of manual medical record evaluation and E-code assignment we would have had to perform a different study, assessing cases reported to the PCC or identified in hospital discharge ICD-9 codes and then estimating case detection failures by our case-finding method. Our case finding would clearly fail to identify a variety of case types, including (as only two such examples) transfers that did not pass through the emergency department and cases with toxic effects but diagnosed as “abuse” rather than “overdose”. Even if it were a sensitive surveillance detection method, case finding by the manual methods we used for the purposes of this study is highly labor intensive and would be unlikely to apply in routine surveillance. Keeping the limitations of our study design in view, our findings are consistent with studies performed at other site locations using comparable methodologies, suggesting that our data may indeed be relevant to other settings. In a 1 year-study of toxic exposure/drug ingestion emergency department visits to a large urban hospital, using eligibility and exclusion criteria comparable with ours based on emergency department case finding, only 26% of 470 cases were reported to a regional poison control center of which that hospital was a member.’ Consistent with our findings, certain causes of overdose were even less likely to be reported to that PCC. such as only two out of 50 hospital-treated cases of heroin overdose. In a statewide study in Oregon based on hospital discharge ICD-9 coding of 2,486 poisoning and drug overdoses, the regional PCC was notified of 46% of the cases.” By comparison, among our 131 hospital-admitted cases we identified 61 (47%) with PCC reports. In a study of E-coding accuracy (3% of the cases were poisonings) in 323 independently reviewed medical records, 63% were found accurately coded by 3-digit E-code, also consistent with our findings even though using a more generous matching criterion than that we used.” In that study 98% of 366 E-coded poisonings received some type of injury N-code as compared with our finding of 94% using a similar comparison. Another recent study of E-code surveillance in injury (although specifically excluding poisoning) found that 45% of ICD-9 N-coded injury cases were not assigned any E-code, although approximately half of these were attributable to lack of coding space; increased age appeared to be associated with decreased likelihood of E-code assignment. I3 In neither of those studies was it possible to identify cases that did not receive any injury ICD-9 code at all (neither N-code nor E-code). Our data suggest that N-coding and E-coding may correlate such that they successfully identify overlapping poisoning and drug overdose cases. while both failing to capture a substantial number of others, in particular toxicity associated with drugs of abuse. Two other studies have independently assigned E-codes to

BLANC

ET AL n SURVEILLANCE

OF POISONING

emergency department-treated injuries including poisoning and drug overdose, allowing estimation of hospital admission rates for such cases. Among 848 emergency department pediatric poisoning cases from 23 hospitals, I25 (15%) were admitted.i4 In another study done at a single tertiary care center, 31 of 163 unintentional poisonings (19%) and 68 of 163 emergency department-treated intentional ingestions (42%) were admitted. l5 This compares with the 17% admission rate we observed at Hospital One (admissions only were studied at Hospital Two). The latter study also noted that 47% of 99 hospital-admitted cases (regardless of intentionality) received intensive care unit care in contrast to the 34% proportion we found. Although we assigned ICD-9 codes without systematic blinding to those assigned by medical records, in most instances match failures were due to the absence of any assigned E-code or poisoning or overdose N-code. Nonetheless, we have intentionally referred to ICD-9 N-code and E-code “match” failures to underscore that coding may be divergent but not necessarily “incorrect.” The purposes of public health surveillance are manifold, including estimating incidence, following temporal trends, and identifying sentinel events. Our findings should not be taken to suggest that existing poisoning and drug overdose data resources should not be used for surveillance. For example, consistent underreporting might not affect interpretations of incidence changes over time. Indeed, we have actively promoted the use of PCC data for surveillance and follow-up for variety of purposes, including in the reporting of occupational illness and in the study of inhalational injury.i6.” Our study, by analyzing multiple surveillance measures for the same group of independently identified subjects, represents a unique approach to assessing poisoning and drug overdose. We identified important sources of systematic underreporting. We found varying but substantial rates of case detection failure for all of the four potential surveillance approaches that we analyzed. These surveillance methods may be systematically effected by such factors as age, intentionality and type of poisoning. Indeed, Table 3 appears to tell a tale reminiscent of “Roshomon,” the Japanese story where each observer’s narration of the same episode appears to describe a markedly different event. Poisoning and drug overdose data derived from ICD-9 coding, PCC reporting, or DAWN case detection should be interpreted not only in the light of the story they have to tell, but the story that goes untold as well. The authors thank Michael Callaham, MD and Alan Tami, PharmD for their help in case identification and Elizabeth McLaughlin, PhD for comments and review of the manuscript.

19

REFERENCES 1. National Academy of Sciences: Injury in America. Washington, DC, National Academy, 1985. 2. Kozel NJ: Epidemiology of drug abuse in the United States: A summary of methods and findings. Bull Pan Am Health Organ 1990;24:53-62 3. Litovitz

TL, Bailey KM, Schmitz BF, et al: 1990 annual reof Poison Control Centers NaData Collection System. Ann Emerg Med 1991;9:461-509

port of the American Association tional

4. Sniezek JE, Finklea JF, Graitcer PL: Injury hospital discharge data. JAMA 1989;262:2270-2272

coding

and

5. Centers for Disease Control: External cause-of-injury coding in hospital discharge data-United States, 1992. MMWR 1992;41:249-257 5. Smith GS, Langlois JA, Buechner JS: Methodologic issues in using hospital discharge data to determine the incidence of hospitalized injuries. Am J Epidemiol 1991;134:1146-1158 7. Harchelroad F, Clark RF, Dean B, et al: Treated vs reported toxic exposures: Discrepancies between a poison control center and a member hospital. Vet Hum Toxicol 1990;32: 156-l 59 8. National Institute on Drug Abuse: Annual Data 1989. Data from the Drug Abuse Warning Network. Series 1; Number 9. (Department of Health and Human Services Publication) No. (ADM) 90-1717) Rockville, MD, US. Department of Health and Human Services, National Institute on Drug Abuse Statistical Series, 1990 9. World Health Organization: Manual of the International Classification of Diseases, Injuries and Causes of Death, ninth revision. Geneva, Switzerland, World Health Organization, 1977 10. Slinker BK, Glantz SA: Multiple regression data analysis: The problem of multicollinearity. 1985;249:Rl-R12 (suppl)

for physiologic Am J Physiol

Il. Griffin S, Steele P: Utilization of the uniform hospital charge data set (UHDDS) to evaluate poisoning incidence. Hum Toxicol 1991;33:376 (abstr)

disVet

12. Smith S, Cowell LS, Sniezek JE: An evaluation of external cause-of-injury codes using hospital records from the Indian Health Service. Am J Public Health 1990;80:279-281 13. Marganitt B, MacKenzie EJ, Smith GS, et al: Coding external causes of injury (E-Codes) in Maryland hospital discharges 1979-88: A statewide study to explore the uncoded population. Am J Public Health 1990;10:1463-1466 14. Guyer B, Berenholz G, Gallagher SS: Injury surveillance using hospital discharge abstracts coded by external cause of injury (E code). J Trauma 1990;30:470-473 15. Ribbeck BM, Runge JW, Thomason MH, et al: Injury surveillance: A model for recording E codes for injured emergency department patients. Ann Emerg Med 1992;21:37-40 16. Blanc PD, Rempel D, Maizlish N, et al: Occupational illness: Case detection by poison control surveillance. Ann Intern Med 1989;111:238-244 17. Blanc PD, Galbo acute irritant inhalation 1991;266:664-669

M, Hiatt P, et al: Morbidity following in a population-based study. JAMA