Injection drug use is commonly recognized for its ongoing impact on the transmission and maintenance of the HIV epidemic in several developed countries.1,2 Recently, a series of epidemics among injection drug users (IDUs) in Russia, other parts of eastern Europe, and central Asia has led to HIV incidence rates reportedly increasing faster than anywhere else in the world.3-5
Although the Health Canada Surveillance Report published in December 2003 showed a relative decrease of positive HIV tests among IDUs in 2002 compared with previous years, the prevalence of risk-taking behaviors in this population remains high despite harm reduction efforts brought forward to expand syringe coverage, outreach, and drug treatment options.6 The majority of IDUs living in Canadian cities report needle or equipment sharing (73.5% in Montreal and 66% in Vancouver),7,8 and 1 of 5 seropositive IDUs also reports sharing needles or equipment (20.2%).9
These facts highlight the importance of developing preventive vaccines to curtail HIV epidemics among IDUs worldwide. Programs aimed at implementing phase 3 vaccine trials should be concerned with issues related to the selection of IDU participants. One of the potential pitfalls of conducting a vaccine trial in the heterogeneous population of IDUs is the difficulty of appraising baseline risks of the participants and their ensuing probability of seroconversion during the study period. Even large trials could founder if the randomization maldistributed high-risk individuals between the experimental vaccine and the control arms of the study.
In modern medicine, clinicians and researchers frequently use patient risk characteristics to stage,10-12 prognosticate,13-15 and sometimes to diagnose diseases.16 As yet, there is little published work as to generate risk scores among the very heterogeneous population of IDUs with regard to their risk of contracting blood-borne infection.17 The development of an HIV risk score among IDUs is challenged by the need to use variables different from the typical set of clinical indicators, such as factors related to behaviors and drug use patterns. Likewise, scoring systems need to possess many sterling qualities: utility, ease of use, universal applicability, and durability.18 Among IDUs, a valid scoring system of HIV risk of infection could conceivably be composed by including factors that are consistently associated with HIV seroconversion, regardless of particular environmental, cultural, and social settings. To link epidemiologic findings to preparedness for vaccine trials among IDUs, adequate risk assessment through the development of risk scores could improve design, proper matching, and stratification.
Factors meeting these criteria can be selected from a large body of literature on risk factors associated with HIV seroconversion. “Efficient mixing,” whereby there is a continued exposure to a number of HIV-infected individuals and a cumulative number of sharing episodes,19 has been consistently associated with seroconversion in several settings, both in high and low prevalence areas.20-24 The exchange of blood through sharing of needles/syringes with an HIV-infected person is the most common cause of HIV transmission among IDUs.25 Other factors found consistently in the literature to be independently associated with HIV seroconversion include marginalization indicators such as homelessness,26 intensity of cocaine injection,27-29 and having unprotected intercourse with an HIV-positive partner,30,31 especially among women.32,33
Conversely, several risk factors, such as age34 and gender,35 have been found to be context-dependent since their relationship with HIV seroconversion is greatly influenced by environmental, cultural, and social settings.
Herein, we examined the relationship between HIV seroconversion and a number of variables related to injection behavior, drug use pattern, and lifestyle in a large cohort of male IDUs, with the intention of designing a scoring system and developing a risk score of seroconversion in this population. This risk score has been developed to provide a basis for inclusion criteria and stratification of patients prior to randomization in therapeutic or clinical trials. This tool might further be used in the clinical management of drug users, allowing IDUs to quantitatively appreciate their level of HIV risk based on their own unique composition of risk behaviors.
Injection drug users older than 14 years who had injected drugs in the past 6 months and were residing in the greater Montreal area were eligible to participate in the St. Luc Cohort prospective study. The study was approved by the Ethical Review Board of the Saint-Luc Hospital of the Centre Hospitalier de l'Université de Montréal and informed consent was obtained from all participants. Participants were recruited through self-referral (51%), from the St. Luc Detoxification Unit (22%), and from collaborating institutions, health care centers, and private physicians (22%). The remaining 5% were recruited from various sources.
At enrollment and semiannually, participants completed an interviewer-administered questionnaire to elicit data on sociodemographic characteristics, drug use patterns, injection and sexual behaviors, health issues, and service utilization. After pretest counseling, a blood sample was obtained by venipuncture for antibody testing of HIV. Serologic screening for HIV antibodies was performed with an enzyme-linked immunosorbent assay. A stipend of $10.00CDN was provided at each visit, at which time referrals were made for universal medical care, HIV/AIDS care, available drug and alcohol treatment, and counseling as requested. The first follow-up visit was scheduled after 3 months, with subsequent visits every 6 months thereafter.
Sample for Analysis
Male participants who were HIV negative at enrollment and completed at least 1 follow-up visit were included in the statistical analysis. Female IDUs represented only 20% of all participants in this cohort and were not included. We herein report results of analyses on 1602 male participants recruited between September 15, 1992 and October 1, 2001. The event of interest in this study was HIV seroconversion. The date of seroconversion was estimated using the midpoint between the last seronegative and the first seropositive antibody test results. Participants who were still seronegative at the time of analysis were censored on the date of their last visit.
To generate a risk score that will be useful in several settings, only factors consistently and independently associated with HIV seroconversion in several studies, without being negatively associated in any report, were included as potential predictors in the present analysis. Data were retrieved from the baseline questionnaire for all participants. Sociodemographic and marginalization indicators included employment status, lifetime incarceration, unstable housing (single room occupancy hotels, transitional living arrangements), and income (percent income from illegal sources). Variables pertaining to injection behaviors included sharing injection equipment with >1 person in the past month, having ever shared injection equipment with a known HIV-positive individual, and having practiced “booting” (a practice more likely to contaminate the injection equipment with blood36) in the past 6 months. To capture the intensity of IV cocaine use, we divided IV cocaine users in 3 categories according to the average number of injections per injecting day in the past month. To explore sexual risk, a composite variable “unsafe sex” was computed as follows: subjects who reported not always using condoms and also reported and answered “Yes” to at least 1 of the following statements: having an HIV-positive or an IDU sexual partner in the past month, having >1 sexual partner in the past 6 months, practicing commercial sex activities, or having had a homosexual or bisexual relationship in the past 6 months. Control variables consisted of age, education, marital status, and language (French, English, or other).
Model Building and Calculating Risk Scores
Cox proportional hazards regression modeling was used to identify independent predictors of HIV seropositivity.37 Univariate survival analyses were conducted with all sociodemographic and other variables of interest, and variables that were found to be either significantly or marginally (P < 0.25) associated with HIV seroconversion were considered for inclusion in a large multivariable Cox regression model. Reduction to a simpler model was done using a backward step-by-step procedure (retained in model if P < 0.05). Several multivariate predictive models were tested and the final model was selected according to the selection criteria outlined above. Simple forms of departure from the proportional hazard model were investigated by using a graphical technique.38 The risk score value for each participant was calculated by adding the products of each of the variable scores with the corresponding coefficient. The sample was divided into 3 risk groups by the K-means clustering iterative algorithm available under the statement FASTCLUS in SAS v.8 (SAS Institute, Cary, NC). Adequate separation of the survival curves among the 3 groups was measured with the log-rank test for homogeneity. The predictive power of the model was then tested following a procedure of data splitting described by Schlichting et al.39 This procedure aimed at validating the score in a set of patients (validation set) who are different from the ones from whom the score was devised (development set). The score was considered useful only if it accurately predicted outcome in this group of patients. To validate the final model, regression coefficients for the variables included in this model were estimated using a randomly drawn development set (n = 1213) representing 75% of the 1602 participants. The validation set consisted of 389 participants divided in 3 groups by k-mean clustering, and the average estimated survivorship function in each group was compared with empirical survivorship function. The samples for validation were drawn in such a way as to obtain approximately equal seronegative-to-seropositive ratios in both groups. The average estimated survivorship functions in the 3 groups obtained from this clustering were compared with the Kaplan-Meier plots. In each group, the difference between observed and estimated survivorship function was tested using the Wilcoxon rank sum test.
A total of 2444 male IDUs completed the enrollment questionnaire between 1992 and 2001. Two hundred sixty-eight IDUs were excluded based on their positive serostatus at intake. Of the remaining 2176 seronegative IDUs, 1602 (73.6%) completed at least 1 follow-up visit and were included in the investigation. When compared with subjects included in this analysis, subjects lost to follow-up were younger (32.8 vs. 34.4 years of age) and more likely to report living in unstable housing (43% vs. 38%), but no significant difference was found relative to drug use patterns or injection or sex risk behaviors.
We observed 178 HIV seroconversions for an incidence rate of 3.35 (95% CI, 2.89-3.88) per 100 person-years. The baseline survival function for this cohort of male IDUs is depicted in Figure 1.
Prognostic Variables and Prognostic Index
The variables that provide significant prognostic information as well as control variables are shown in Table 1 with their regression coefficients. Among sociodemographic and marginalization indicators, lifetime incarceration history, unemployment, and an unstable housing condition were associated with HIV seroconversion. With regards to drug use, the intensity and pattern of IV cocaine use were correlated with HIV infection. Injection practices measured by the number of injections in the previous month and high-risk practices such as booting and sharing syringes were also found to be associated with seroconversion. Reporting sexual relations with an HIV-positive partner was the sole sexual-related behavior associated with HIV acquisition. Moreover, we found no effect modification of sexual risk by injection practices (data not shown).
The final prognostic model presented in Table 2 included 5 highly significant prognostic variables under the following regression equations:
Cluster analysis divided the cohort in 3 sets of low (n = 420), intermediate (n = 812), and high (n = 370) risk. Risk scores values ranged from 0-2.80 with a mean of 1.04 and a median of 0.83. Figure 2 illustrates the survival functions for the 3 groups of IDUs according to their risk scores. Incidence rates for low-, intermediate-, and high-risk groups were 0.91 (95% CI, 0.55-1.52) per 100 person-years, 3.10 (95% CI, 2.49-3.84) per 100 person-years, and 7.82 (95% CI, 6.30-9.73) per 100 person-years, respectively. The clustering algorithm led to a reasonable separation of survival curves over strata (P < 0.0001 for the log-rank test).
Validity of the Prognostic Model
As shown in Figure 3, no statistically significant difference was found between the estimated and the observed functions for the 3 groups using the Wilcoxon statistical test (χ2 = P = 0.084, 0.57, and 0.70 for low-, medium-, and high-risk groups, respectively).
As more efforts go into the development of preventive vaccines and other preventive measures to prevent the spread of the HIV epidemic, issues relating to the proper conduction of clinical trials among the very heterogeneous population of IDUs are arising, especially with regard to the selection of participants.
In the present study, a risk assessment tool was developed to assist stratification of IDU participants in clinical trials. In the analysis, we have favored variables having the potential to predict the occurrence of HIV seroconversion in several settings. This model was not generated to identify the determinants of HIV seroconversion. Instead, by including only 3 key risk indicators, the risk score captured the prognostic information required to identify 3 different subgroups of IDUs in relation to subsequent HIV seroconversion rates.
In clinical medicine, prognostic model building using this statistical approach is well documented.12-15,39,40 In contrast to most studies, our model did not use clinical data but called for variables that best portrayed IDUs' risk behavior. The final multivariate model was constructed with fixed measures of self-reported risk behaviors taken from baseline survey data. This strategy was selected to imitate the situation encountered during the randomization process, whereby information is available from 1 single data collection. Variables selected for the prognostic model included drug use patterns,27-29 injection and sexual behaviors,19,25 and marginalization indicators,26 all of which were found to be independently associated with HIV seroconversion. The final model was validated by measuring its reliability. “Reliability” or “calibration,” which refers to the amount of agreement between predicted and observed outcomes, was evaluated comparing the observed and expected probability of remaining seronegative. Our validation results suggest that the proposed final model is reliable. This analysis demonstrates the robust prognostic performance of the risk score in a large group of IDUs representative of the heterogeneous population of drug users in Montreal.
Our scoring system was generated with male IDUs exclusively to ensure optimal internal validity and power. Women in the St. Luc Cohort only account for 20% of the total population, and only 17 seroconversions (8% of the cases) were observed in the study period, generating unstable models. Our decision to exclude women was based on the general acceptance that HIV infection determinants are largely gender specific.35,41,42 For instance, unsafe heterosexual intercourse has been identified as an independent predictor of HIV seroconversion among women in many settings, but not among men.32,43,44 These gender-specific differences in risk factors for HIV infection should be taken into consideration when planning prevention initiatives by using distinct explanatory models for each gender.32
There are other limitations in this study. As in other observational studies, self-reported data may be subject to bias, although several studies have reported only moderate degrees of inconsistency in test-retest designs.45,46 Even if we used multiple recruiting sites and methods to reduce sampling biases, the sample was not randomly selected. We found few differences between subjects followed and those lost to follow-up, yet we cannot rule out that differential loss to follow-up may have biased the study findings in an unforeseeable way. Conversely, characteristics of individuals lost to follow-up in this study may well be similar to those lost in the course of a vaccine trial, making the findings of this study more generalizable to that purpose.
In our study, baseline characteristics of the participants suggest an overrepresentation of male cocaine-using, high-risk IDUs relative to the injecting population in Quebec,47 which may affect the external validity. Nevertheless, we believe that this is precisely 1 group of IDUs who are at high risk for contracting HIV and are hence ideal candidates for vaccine trials. This longitudinal study provided the ideal setting to assess the impact of several predictive factors of HIV seroconversion.
Finally, the extent to which our model can be generalized to other male IDU populations and to female IDUs is yet unknown. Our risk score was computed by using variables consistently associated with HIV seroconversion in a context-independent fashion, and in areas with various seroprevalences, increasing the likelihood of being adapted to other settings. Validation of this scoring system will have to be performed in other cities with different HIV seroprevalences20 and among women.
In view of new interventions, including therapeutic and vaccine trials targeted at IDUs, the development of a simple and effective tool to identify individuals highly susceptible to HIV has been carried out. This study highlights prognostic factors that may serve to stratify risk across a spectrum of IDUs. The risk score developed may be useful in triage to different risk groups prior to vaccine or other therapeutic trials. Overall, the prognostic models designed using these methodologic tools help identify which patients will benefit from specific treatments and provide accurate stratification for vaccine or other therapeutic trials.
The authors thank the staff and participants of the St. Luc Cohort.
1. Holmberg SD. The estimated prevalence and incidence of HIV in 96 large US metropolitan areas. Am J Public Health
2. Alcabes P, Friedland G. Injection drug use
and human immunodeficiency virus infection. Clin Infect Dis
3. Rhodes T, Lowndes C, Judd A, et al. Explosive spread and high prevalence of HIV infection among injecting drug users in Togliatti City, Russia. AIDS
4. Hamers FF, Batter V, Downs AM, et al. The HIV epidemic associated with injecting drug use in Europe: geographic and time trends. AIDS
5. Sarkar S, Das N, Panda S, et al. Rapid spread of HIV among injecting drug users in north-eastern states of India. Bull Narc
6. Health Canada. HIV and AIDS in Canada, surveillance report to December 31, 2003. Division of HIV/AIDS Epidemiology and Surveillance, Center of Infection Diseases Prevention and Control; Ottawa, Ontario, Canada; 2004.
7. Bruneau J, Lamothe F, Franco E, et al. High rates of HIV infection among injection drug users participating in needle exchange programs in Montreal: results of a cohort study. Am J Epidemiol
8. Strathdee SA, Patrick DM, Archibald CP, et al. Social determinants predict needle-sharing behaviour among injection drug users in Vancouver, Canada. Addiction
9. Wood E, Tyndall MW, Spittal PM, et al. Factors associated with persistent high-risk syringe sharing in the presence of an established needle exchange programme. AIDS
10. Fisher RI. Overview of non-Hodgkin's lymphoma: biology, staging, and treatment. Semin Oncol
11. Haase GM, Perez C, Atkinson JB. Current aspects of biology, risk assessment, and treatment of neuroblastoma. Semin Surg Oncol
12. Mountain CF. Staging classification of lung cancer. A critical evaluation. Clin Chest Med
13. Kaneko M, Kanda Y, Oshima K, et al. Simple prognostic model
for patients with multiple myeloma: a single-center study in Japan. Ann Hematol
14. Resnic FS, Ohno-Machado L, Selwyn A, et al. Simplified risk score
models accurately predict the risk of major in-hospital complications following percutaneous coronary intervention. Am J Cardiol
15. Solomon DH, Ganz DA, Avorn J, et al. Which patients with unstable angina or non-Q-wave myocardial infarction should have immediate cardiac catheterization? A clinical decision rule for predicting who will fail medical therapy. J Clin Epidemiol
16. Mistry FP, Karnad DR, Abraham P, et al. A scoring system to differentiate cirrhotic from non-cirrhotic portal hypertension. Indian J Gastroenterol
17. Cheng FK, Ford WL, Weber MD, et al. A probability-based approach for predicting HIV infection in a low prevalent population of injection drug users. Ann Epidemiol
18. Shah S, Karnad DR. What's the score? On the development and use of prognostic scores in clinical practice. Indian J Gastroenterol
19. Greenhalgh D, Lewis F. The general mixing of addicts and needles in a variable-infectivity needle-sharing environment. J Math Biol
20. Friedman SR, Jose B, Deren S, et al. Risk factors for human immunodeficiency virus seroconversion among out-of-treatment drug injectors in high and low seroprevalence cities. The National AIDS Research Consortium. Am J Epidemiol
21. Neaigus A, Friedman SR, Jose B, et al. High-risk personal networks and syringe sharing as risk factors for HIV infection among new drug injectors. J Acquir Immune Defic Syndr Hum Retrovirol
22. Jackson LA, Bailey DL, Fraser JR, et al. Safer and unsafe injection drug use
and sex practices among injection drug users in Halifax, Nova Scotia. An exploratory look at community and interpersonal influences. Can J Public Health
23. Thorpe LE, Bailey SL, Huo D, et al. Injection-related risk behaviors in young urban and suburban injection drug users in Chicago (1997-1999). J Acquir Immune Defic Syndr
24. Van Ameijden EJ, Langendam MW, Notenboom J, et al. Continuing injecting risk behaviour: results from the Amsterdam Cohort Study of drug users. Addiction
25. Grund JP, Friedman SR, Stern LS, et al. Syringe-mediated drug sharing among injecting drug users: patterns, social context and implications for transmission of blood-borne pathogens. Soc Sci Med
26. Song JY, Safaeian M, Strathdee SA, et al. The prevalence of homelessness among injection drug users with and without HIV infection. J Urban Health
27. Alary M, Hankins C, Parent R, et al. SurvIDU Working Group. Risk factors for HIV acquisition among IDUs in the SurvIDU Network: long-term cocaine injectors are those most at risk. Can J Infect Dis
. 2000;11(Suppl B):Abstract No. 302.
28. Anthony JC, Vlahov D, Nelson KE, et al. New evidence on intravenous cocaine use and the risk of infection with human immunodeficiency virus type 1. Am J Epidemiol
29. Perron S, Raynault MF, Bruneau J. Frequency of intravenous cocaine bingeing and risk factors associated with HIV transmission. Can J Infect Dis
. 2002;13(Suppl A):Abstract NB 336P.
30. Kral AH, Bluthenthal RN, Lorvick J, et al. Sexual transmission of HIV-1 among injection drug users in San Francisco, USA: risk-factor analysis. Lancet
31. Chitwood DD, Griffin DK, Comerford M, et al. Risk factors for HIV-1 seroconversion among injection drug users: a case-control study. Am J Public Health
32. Spittal PM, Craib KJ, Wood E, et al. Risk factors for elevated HIV incidence rates among female injection drug users in Vancouver. Can Med Assoc J
33. Estebanez PE, Russell NK, Aguilar MD, et al. Women, drugs and HIV/AIDS: results of a multicentre European study. Int J Epidemiol
34. Kral Ah, Lorvick J, Edlin Br. Sex and drug-related risk among populations of younger and older injection drug users in adjacent neighborhoods in San Francisco. J Acquir Immune Defic Syndr
35. Bruneau J, Lamothe F, Soto J, et al. Sex-specific determinants of HIV infection among injection drug users in Montreal. Can Med Assoc J
36. Kral AH, Bluthenthal RN, Erringer EA, et al. Risk factors among IDUs who give injections to or receive injections from other drug users. Addiction
37. Cox DR. Regression models and life tables. J R Stat Soc
38. Hess KR. Graphical methods for assessing violations of the proportional hazards assumption in Cox regression. Stat Med
39. Schlichting P, Christensen E, Andersen PK, et al. Prognostic factors in cirrhosis identified by Cox's regression model. Hepatology
40. Rex DK. Screening for colon cancer and evaluation of chemoprevention with coxibs. J Pain Symptom Manage
41. Evans JL, Hahn JA, Page-Shafer K, et al. Gender differences in sexual and injection risk behavior among active young injection drug users in San Francisco (the UFO Study). J Urban Health
42. Latkin CA, Mandell W, Knowlton AR, et al. Gender differences in injection-related behaviors among injection drug users in Baltimore, Maryland. AIDS Educ Prev
43. Dwyer R, Richardson D, Ross MW, et al. A comparison of HIV risk between women and men who inject drugs. AIDS Educ Prev
44. Strathdee SA, Galai N, Safaiean M, et al. Sex differences in risk factors for HIV seroconversion among injection drug users: a 10-year perspective. Arch Intern Med
45. De Irala J, Bigelow C, McCusker J, et al. Reliability of self-reported human immunodeficiency virus risk behaviors in a residential drug treatment population. Am J Epidemiol
46. McElrath K, Chitwood DD, Griffin DK, et al. The consistency of self-reported HIV risk behavior among injection drug users. Am J Public Health
47. Comité Permanent de Lutte à la Toxicomanie. Le Point sur la Situation de la Toxicomanie au Québec en L'An. 2000. Gouvernement du Québec, Ministère de La Santé et des Services Sociaux. Montreal, Québec, Canada: 2001.
Keywords:© 2005 Lippincott Williams & Wilkins, Inc.
prognostic model; risk score; survival study; HIV behavioral risk factors; injection drug use