1 Alcohol Detection Services, Big Bend, WI and 2 Millennium Strategies, Madison, WI, USA
* Author to whom correspondence should be addressed at: Alcohol Detection Services, W236 S7050 Big Bend Drive, Big Bend, WI 53103, USA. Tel.: +262 783 7427; Fax: +262 783 7481; E-mail: edac{at}execpc.com
(Received 2 December 2003; first review notified 18 January 2004; in revised form 2 April 2004; accepted 3 April 2004)
![]() |
ABSTRACT |
---|
![]() ![]() ![]() ![]() ![]() ![]() ![]() |
---|
![]() |
INTRODUCTION |
---|
![]() ![]() ![]() ![]() ![]() ![]() ![]() |
---|
Efforts to utilize biochemical markers to identify harmful alcohol consumption have focused either on finding a single laboratory test or on examining combinations of several laboratory tests (Bean et al., 2002; Cambou 2002
; Reynaud et al., 2000
Allen, 2003
; Sillanaukee et al., 2003
; Soderberg et al., 2003
). Because single laboratory tests have usually lacked adequate performance as stand-alone tests, several models of multiple biomarkers have been sought. A comprehensive review of models that have used multiple biochemical tests for identifying heavy drinkers has been published previously (Hartz et al., 1997
). This review shows that the performances of these models have varied widely. One of the most successful approaches mentioned in Hartz's review is that pioneered by a team of researchers who used quadratic discriminant analysis of 25 routine tests to produce correct classification rates of 100% of non-alcoholics and 95% of alcoholics (Eckardt et al., 1981
, 1984
; Ryback et al., 1982
, 1983
, 1985
). Even though this early research was able to differentiate alcoholics from non-alcoholics with a high degree of accuracy, most investigators abandoned this approach because computing costs in the early 1980s made this type of statistical analysis impractical.
In the late 1990s, three similar approaches re-emerged to develop an alcohol screening method based on routine laboratory analytes: (1) a logistic regression equation using 40 routine laboratory tests to identify heavy drinking males (Hartz et al., 1997); (2) a discriminant function analysis (Welte and Chan, 1997
), and (3) a new linear discriminant function (LDF) model that culminated in what is now known as the early detection of alcohol consumption test, or EDAC. The main objective of this study is to describe the derivation, validation and performance of the EDAC test, which used a routine laboratory panel of 36 chemistry and haematology analytes in 1599 subjects reporting a wide range of alcohol intake. This article also compares the EDAC to two similar models developed previously. As awareness of the EDAC test has been increasing steadily in the US, Germany and Canada, a detailed description of the development and performance of this new LDF approach will ensure its proper interpretation and use by the medical community.
![]() |
MATERIALS AND METHODS |
---|
![]() ![]() ![]() ![]() ![]() ![]() ![]() |
---|
Heavy drinkers were recruited from substance misuse treatment centres (outpatients and inpatients) and from detoxification centres. Light drinkers and abstainers were recruited from churches that discouraged drinking, community recovery support groups and professional and social groups in which there was an opportunity for outside confirmation of the subjects' self report. The method used to confirm subjects' reports relied mainly on the settings. For instance, if the subject was recruited at a treatment centre we assumed there was a high probability the subject had been drinking. If the subject was recruited from certain religious institutions we assumed there was a high probability the subject had been abstinent. When there was doubt on self-report, we contacted a collateral including counsellors, ministers and work peers. We rejected 72 subjects during recruitment either because of doubts on self-report or due to the presence of extreme liver damage. The group demographics ranged from homeless public inebriates to upper-middle-class professionals, from daily heavy drinkers to infrequent consumers of alcohol, from subjects who had never drunk due to religious practices to recovering alcoholics who had been abstinent for a minimum of 10 weeks. As already stated, subjects with extreme elevations of liver enzymes were not included in this study.
Subjects signed informed consent. They were interviewed by the main author (J.H.) and by two trained counsellors for the purpose of gathering information regarding age, sex, race, chronic illness, concurrent drug use, medications, smoking habits, diet and alcohol consumption. Women were also asked about current pregnancy and exogenous hormone use. The instruments used to assess alcohol consumption were the Khavari alcohol test (Khavari and Farber, 1978) and a self-administered alcohol screening test (Swenson and Morse, 1975
). The original database contained 1701 subjects, of which 102 were left out because they were missing at least one discriminating variable. Therefore, the total number of subjects in the final analysis was 1599 (1019 men; 580 women) (Table 1).
|
Laboratory tests
Four tubes of blood were collected from each patient; two of these tubes were spun down and serum was separated within 4 h of sample collection. Refrigerated blood and serum specimens were shipped to the main laboratory overnight. A total of 36 laboratory tests results were used in the derivation of the EDAC model. Chemistry determinations were performed utilizing the Olympus AU5000 automated chemistry instrument and haematology parameters were determined utilizing the Argos Cobas instrument (LabCorp, Triangle Park, NC). We used a broad range of commonly available laboratory tests because almost every organ system in the body can be affected by large amounts of alcohol. The panel of analytes examined and their corresponding reference ranges are shown in Table 2. The results from the tests were forwarded to Alcohol Detection Services for analysis.
|
LDF essentially involves the development of a predictive equation that is a linear function of the independent variables (Polit, 1996; Gardiner, 1997
). To do this, it solves for the coefficients in the predictor set so as to maximize the variance between the two groups to the variance within the two groups. Based on the entire equation for the 36 laboratory test results, a D-score is calculated for each subject in the database and is used to classify subjects into groups. Actual group membership is obtained by self-report, whereas projected group membership is obtained by the LDF analysis. The sensitivity of the EDAC is defined as the proportion of all heavy drinkers who had a D-score greater than or equal to a chosen cut-off point and the specificity is defined as the proportion of all light drinkers who had a D-score less than that same cut-off point. In the LDF analysis, the assignment of this cut-off point can be manipulated by assigning what is called prior probabilities, which works by maximizing the predictive value of the classification process.
Significance testing, cross-validation and stepwise analysis
The initial significance test to evaluate the EDAC model was the test of the null hypothesis, which states that heavy drinkers and light drinkers cannot be reliably distinguished on the basis of the predictors in the analysis. In this study, the test of the null hypothesis was based on Wilks' lambda, and the significance level was based on a transformation that approximates a chi-squared distribution. As Wilks' lambda is the proportion of total variation in discriminant scores not explained by group differences, a lambda of 1 would be obtained if the mean of the discriminant scores were the same for both groups and there was no between groups variability. Lower values of Wilks' lambda reflect better performance of the model.
The validation of the EDAC equation was accomplished by splitting the data randomly in half and using different halves to develop and test the model. If the classification process in the validation set shows a high percentage of correct classifications, then the LDF equation can be used to classify new cases for which group membership is unknown. The results of the laboratory tests for a subject are inserted into the EDAC equation and this yields that subject's classification probability for each group. The case is assigned to the group for which it has the highest classification probability.
Stepwise analysis was performed in order to select those laboratory tests that contributed the most to the significance of the EDAC model. During stepwise analysis the 36 independent variables are stepped into the discriminant function in the order in which they meet certain statistical criteria. The first variable included is the one with the largest value for the selection criteria. If the selection criterion for entering predictors in the analysis is the minimization of Wilks' lambda, then at each step the value of the criterion is reassessed for each predictor not yet entered. When further variable entry results in a non-significant improvement, the analysis ceases. When selecting the variables that were the best predictors for identifying heavy drinkers using the EDAC model, we eliminated all variables not significant at the P < 0.001 level.
![]() |
RESULTS |
---|
![]() ![]() ![]() ![]() ![]() ![]() ![]() |
---|
Summary classification
The summary classification results (Table 3a) shows that of the 818 heavy drinkers, 660 (80.7%) were correctly classified and 158 (19.3%) were misclassified as light drinkers when the classification proceeded under the assumption of equal prior probability. ie. for when the classification proceeded under a cut-off point that maximizes sensitivity. Among the 781 subjects who were light drinkers, 659 (84.4%) were correctly classified as light drinkers and 122 (15.6%) were misclassified on the basis of the LDF. To be useful, the classification should be substantially better than what could be achieved by chance alone. As the overall correct classification rate (82.5%) obtained by the LDF is much greater than the chance probability of 50% the classification is useful. Next, the cut-off point was manipulated by assigning new prior probabilities. In theory, as the cut-off of the test is manipulated to approach 0% detection of heavy drinkers, then maximum specificity is achieved, and as the prior probability approaches 100% detection of heavy drinkers, then maximum sensitivity is obtained. When the cut-off was raised to maximize specificity, the results showed that specificity increased to 98%, whereas the sensitivity decreased to 33% for an overall rate of successful classification equal to 64.8% (Table 3b).
|
Significance testing and cross-validation
In the EDAC model, the value of Wilks' lambda is 0.573, which is statistically significant at the P < 0.0001 level. Thus, 42.7% (10.573) of the total variation in discriminant scores is accounted for by between-group differences in the 36 predictor variables. A significant Wilks' lambda supports the validity of the variables used to distinguish between heavy drinkers and light drinkers based on the EDAC model.
The results of the cross-validation analysis for the EDAC test are shown in Table 4. The entire population of 1599 subjects was split in two: 772 subjects were used as the training sample and 827 subjects were used to validate the LDF. As expected, the performance parameters for the training sample (82.1% sensitivity and 87% specificity rates) were slightly higher than the performance parameters for the validation sample (78.4% sensitivity and 82.5% specificity rates).
|
|
|
![]() |
DISCUSSION |
---|
![]() ![]() ![]() ![]() ![]() ![]() ![]() |
---|
The model developed by Welte and Chan (1997) is also similar to the EDAC in some respects. For instance, both models include subjects with a wide range of alcohol consumption histories and they also include men and women. In addition, three of the top four variables considered in the Welte and Chan model were also considered in the EDAC model. Thus the four main variables of Welte's model were GGT, MCV, monocyte counts and albumin, which rendered sensitivity and specificity rates in the 7075% range when prior probabilities were set at 0.6 (for the heavier drinking group) and 0.4. Of these, MCV was the only variable not selected by the stepwise analysis of the EDAC model. MCV was also not included in the final logistic regression equation of Hartz's model because it was less strongly associated with drinking status for African Americans.
Regarding discrepancies between the models, the EDAC approach identified 20 laboratory variables contributing to the identification of heavy drinkers whereas the model of Welte and Chan reported that the use of more than three or four variables did not improve their classification rates. The Wilks' lambda scores achieved by the four variables selected by Welte and Chan were always above 70 whereas the Wilks' lambda scores achieved by the EDAC model were always lower than 60, suggesting a better overall performance for the EDAC. This difference in the performance of these two models can be explained by the use of different variables, the use of different prior probabilities and/or the log transformations carried out by Welte and Chan to increase the correlation between alcohol consumption and blood variables.
When the EDAC was started, subjects were recruited from extreme groups with respect to alcohol consumption to achieve a big separation of the laboratory profiles identifying each group. After the EDAC equation was developed the next step was to apply the equation to less extreme groups. Harasymiw and Bean (2000) showed that in college students with known drinking histories and no elevation of liver enzymes, the EDAC identified 80% of young, heavy-drinking men and 100% of young, heavy-drinking women. These results, together with the fact that 80% of the subjects in our database show no elevation in AST and none shows elevated albumin, indicates that the EDAC can be used effectively not only to identify hazardous drinkers but also to detect alcohol misuse before alcohol-related liver damage occurs.
A strength of the EDAC is that the analytes included in the EDAC panel are routine tests usually available in most clinics. Thus, the EDAC constitutes a value-added tool to the medical community. The results of the panel can be forwarded to Alcohol Detection Services via fax or electronically and a research assistant will then enter the data in a personal PC. Using the widely available SPSS software, the EDAC calculations are done in less than 5 min for a cost of US$3.005.00 depending on the choice of data transmission and test volume. Alternatively, if there is a need to order the panel, these routine tests are performed daily by almost any laboratory in the US using a variety of automated analysers. These analysers have been programmed to measure several routine tests in the same run with minimal manual input (labour costs) for an average price of US$30.0035.00.
In summary, this study shows that LDF analysis in conjunction with new, more user-friendly computer packages such as SPSS, is a practical and cost-effective laboratory tool for detecting excessive drinking using blood constituents ordered routinely in a variety of clinical settings.
![]() |
ACKNOWLEDGEMENTS |
---|
![]() |
REFERENCES |
---|
![]() ![]() ![]() ![]() ![]() ![]() ![]() |
---|
Allen, J. P. (2003) Use of biomarkers of heavy drinking in health care practice. Military Medicine 168, 364367.[ISI][Medline]
Bean, P., De Bruin, T., Harasymiw, J. and Hallinan, P. (2002) New Applications of contemporary biomarkers of alcohol consumption. American Clinical Laboratory 20(3) 1921. Available at: www.iscpubs.com/articles/acl/c0203bea1.pdf (accessed 4 March 2004).
Cambou, J. P. (2002) A practical approach to the evaluation of excessive alcohol drinking. Annales de Cardiologie et d Angeiologie 51, 321326.[Medline]
Department of Health and Human Services (1999) National household survey on drug abuse series: H-10. Summary of findings from the 1998 national household survey on drug abuse. DHHS Publication No. SMA 99-3328. US Department of Health and Human Services, Substance Abuse and Mental Health Services Administration, Washington DC.
Eckardt, M. J., Rawlings, R. R., Ryback, R. S., Martin, P. R. and Gottschalk, L. A. (1984) Effects of abstinence on the ability of clinical laboratory tests to identify male alcoholics. American Journal of Clinical Pathology 82, 305310.[ISI][Medline]
Eckardt, M. J., Ryback, R. S., Rawlings, R. R. and Graubard, B. I. (1981) Biochemical diagnosis of alcoholism. A test of the discriminating capabilities of gamma-glutamyl transpeptidase and mean corpuscular volume. Journal of the American Medical Association 246, 27072710.[Abstract]
Gardiner, W. P. (1997) Statistical discriminant analysis. In Statistical Analysis Methods for Chemists: A Software-Based Approach, pp: 313325. The Royal Society of Chemistry, Cambridge, UK.
Harasymiw, J. W. and Bean, P. (2001) Identification of heavy drinkers using the Early Detection of Alcohol Consumption score. Alcohol: Clinical Experimental Research 25, 228235.[CrossRef][ISI][Medline]
Harasymiw, J. W., Vinson, D. and Bean, P. (2000) The Early Detection of Alcohol Consumption (EDAC) score in the identification of heavy and at-risk drinkers from routine blood tests. Journal of Addictive Diseases 19, 4358.
Hartz, A. J., Guse, C. and Kajdacsy-Balla, A. (1997) Identification of heavy drinkers using a combination of laboratory tests. Journal of Clinical Epidemiology 50, 13571368.[CrossRef][ISI][Medline]
Javors, M. A., Bean, P., King, T. S. and Anton, R. F. (2003). Biochemical markers of alcohol consumption. In Handbook of Clinical Alcoholism Treatment, Johnson, B. A., Ruiz, P. and Galanter, M., eds, pp. 6279. Lippincott, Williams & Wilkins, Baltimore.
Khavari, K. and Farber, P. (1978) A profile instrument for the quantification and assessment of alcohol consumption. Journal of Studies in Alcohol 39, 15251539.[ISI][Medline]
Kleck, W. R. (1994) Discriminant Analysis. Sage Publications, Newbery Park, NJ.
National Institutes of Health (2000) 10th Special Report in the US Congress on alcohol and Health. NIH Publication No. 001583 NIAAA. National Institutes of Health, Bethesda, MD.
Olfson, M., Tobin, J. N., Cassells, A. and Weissman, M. (2003) Improving the detection of drug misuse, alcohol misuse, and depression in community health centers. Journal of Health Care of the Poor and Underserved 14, 386402.
Polit, D. F. (1996) Discriminant analysis and logistic regression. In Data Analysis and Statistics for Nursing Research, pp. 381412. Appleton and Lange, Stamford, CT, USA.
Reynaud, M., Schellenberg, F., Loisequx-Meunier, M. N., Schwan, R., Maradeix, B., Planche, F. and Gillet, C. (2000) Objective diagnosis of alcohol misuse: compared values of carbohydrate-deficient transferrin (CDT), gamma-glutamyl transferase (GGT), and mean corpuscular volume (MCV). Alcohol: Clinical and Experimental Research 24, 14141419.[ISI][Medline]
Ryback, R. S., Eckardt, M. J., Felsher, B. and Rawlings, R. R. (1982) Biochemical and hematologic correlates of alcoholism and liver disease. Journal of the American Medical Association 248, 22612265.[Abstract]
Ryback, R. S., Eckardt, M. J., Negron, G. L. and Rawlings, R.R. (1983) The search for a biochemical marker in alcoholism. Substance and Alcohol Actions Misuse 4, 217224
Ryback, R. S., Rawlings, R., Faden, V. and Negron, G. L. (1985) Laboratory test changes in young abstinent male alcoholics. American Journal of Clinical Pathology 4, 474479.
Sillanaukee, P., van der Gaag, M. S., Sierksma, A., Hendriks, H. F., Strid, N., Ponnio, M. and Nikkari, S. T. (2003) Effect of type of alcoholic beverages on carbohydrate-deficient transferrin, sialic acid, and liver enzymes. Alcohol: Clinical and Experimental Research 27, 5760.[ISI][Medline]
Soderberg, B. L., Salem, R. O., Best, C. A., Cluette-Brown, J. E. and Laposata, M. (2003) Fatty acid ethyl esters. Ethanol metabolites that reflect ethanol intake. American Journal of Clinical Pathology 119, S94S99.[Medline]
Swenson, W. M., and Morse, R. M. (1975) The use of a self-administered alcoholism screening test (SAAST) in a medical center. Mayo Clinical Proceedings 50, 204208.[ISI][Medline]
Tabachnick, B. G. and Fidell, L. S. (1996) In Using Multivariate Statistics, 3rd edn. Harper Collins, New York.
Welte, J. W. and Chan, A. W. (1997) Factors affecting the discriminant function analysis of blood chemistry profiles. Alcohol 14, 161166.[CrossRef][ISI][Medline]