Background: The current French hepatitis C virus infection screening programme is not yet reaching all populations at risk. In order to better identify individuals that would benefit from a screening test, we investigated an expanded combination of personal characteristics as potential screening criteria for this infection. Methods: We constructed two multiple-regression models predicting hepatitis C antibody seropositivity using the population sample from the 2004 French national hepatitis C antibody seroprevalence survey (SPS) (n = 14 416): one representing current screening guidelines and another constructed from personal characteristics collected for the SPS. Performance of the two predictive models was statistically compared and we internally validated the better performing model. Results: The expanded screening criteria model better discriminated seropositive and seronegative individuals [area under the ROC curve (AUC) 0.869 (95% CI 0.861–0.873)] than the current screening guidelines model [AUC 0.821 (95% CI 0.810–0.824)]. This performance difference was statistically significant (P < 0.00001). The expanded criteria model contains the variables age, sex, pre-1992 blood transfusion, intra-venous drug use, receipt of medical welfare for precarious individuals, previous surgeries, illicit nasal drug use, previous hepatitis C screening, tattoo, raised alanine aminotransferase level and birth in a hepatitis C high/moderate-prevalence country. Conclusion: Results indicate that an expanded combination of screening criteria better predicted hepatitis C antibody status and thus individuals needing screening than the current French-screening guidelines. The proposed combination of screening criteria could more effectively target hepatitis C risk-populations in France and could serve as the basis for a decision-making screening tool for the general population.
A first national seroprevalence survey for hepatitis C virus (HCV) carried out in France in 1994, estimated a prevalence of antibodies to HCV (anti-HCV) of 1.1%, corresponding to between 500 000 and 650 000 seropositive individuals.1 Subsequent to this study, national hepatitis C prevention programmes were put in place in 1999, 2002 and 2009.
Screening high-risk populations is considered an essential component of HCV prevention programmes in France and recommendations concerning populations to screen have been published.2 In 2001, the French Agency for Accreditation and Evaluation in Health (ANAES) produced a list of risk factor-based HCV screening criteria for identification of populations that would benefit from a screening test.3 These criteria continue to be used by medical practitioners prescribing screening tests for hepatitis C in France.
A second anti-HCV seroprevalence survey carried out in 2004 estimated a national prevalence of 0.8% (95% CI 0.7–1.1%) and identified additional factors independently associated with seropositivity for anti-HCV that are not cited in the ANAES screening guidelines.4 These factors included illicit intranasal drug use and low socioeconomic status as defined by receipt of the state-funded medical welfare for precarious individuals (MWPI). Additionally, data from the laboratory-based surveillance of HCV screening activity in France provide evidence of a recent decrease in effective targeting of risk populations by the national screening programme. The number of reported anti-HCV screening tests carried out between 2000 and 2006 increased by 30% while an 11% decrease in the number of positive anti-HCV screening tests was reported during the same period, although the representativeness of this voluntary surveillance system must be considered when interpreting these data.5
The 2004 prevalence survey provided unique access to a large national population sample that had been interrogated about a range of personal characteristics potentially associated with hepatitis C that is wider than those classically cited in screening guidelines. The study objective was to identify a combination of screening criteria for hepatitis C that would better identify individuals at risk for this infection than the current French screening guidelines.
Study population and data collection
The population sample of the 2004 national anti-HCV seroprevalence survey as described previously by Meffre et al.4 was used for this study. This sample comprised 14 416 individuals living in France aged between 18 and 81 years who are subscribed to the national health insurance system, which represents 80% of the French general population. A complex sampling design was used with stratification by five geographic inter-regions, socioeconomic status (beneficiaries or non-beneficiaries of the MWPI) and by age (<65 and ≥65 years).
The study population had a global anti-HCV prevalence of 0.8% with women more affected (1.0%) than men (0.7%).4 The age group with highest prevalence was 45–49 years (2.3%); 18–24-year-olds had the lowest prevalence (0.04%).
Individuals participating in the 2004 seroprevalence survey had been invited by letter to a health check-up and to participate in a health study.4 A questionnaire collecting socio-demographic characteristics and information on potential risk factors for acquiring hepatitis C was administered by a medical professional along with a blood test.4
Statistical analyses were preformed in Stata® (v8) accounting for the complex sampling design through use of the ‘svylogit’ command and in the software R.
Logistic regression modelling
Two predictive logistic regression models were constructed from the 2004 seroprevalence survey sample: one representing the current French screening guidelines issued by the ANAES in 2001 (ANAES model) and a second model based on characteristics collected during the 2004 national anti-HCV seroprevalence survey (2004 SPS model). Univariate logistic regression analyses were performed using seropositivity for anti-HCV as the dependant variable. Independent variables for the 2004 SPS model were selected from the self-reported characteristics of the survey (N = 36). For the ANAES model, variables of the 2004 seroprevalence survey were matched as closely as possible or were transformed to match criteria of the ANAES guidelines (N = 12). Odds ratios (ORs) were calculated with 95% confidence intervals (95% CI) and P-values.
Several variables had to be constructed from data of the 2004 seroprevalence survey. The variable representing an elevated level of the liver enzyme alanine aminotransferase (ALT), a marker of liver disease and damage, was constructed using the log normal distribution of the ALT readings taken by laboratories during the 2004 seroprevalence survey. Age and sex-adjusted threshold values for an elevated ALT level were determined from this distribution according to the equation 10(mean + 1.96 SD).6 The variable representing HCV prevalence in birth region was constructed according to the ANAES guidelines for the ANAES model (high prevalence = Southeast Asia, Middle-East, Africa and South America) and for the 2004 SPS model based on the World Health Organisation (WHO) centralized national hepatitis C prevalence estimates (high-moderate prevalence defined as ≥2.5% = North Africa, Middle-East, Sub-Saharan Africa, Indian subcontinent, Asia and the Pacific).3,7
Variables showing an association of P ≤ 0.25 were included in the multivariable analysis. Backwards stepwise selection was performed following the Hosmer and Lemeshow procedure and variables were allowed to remain in the model with a Chi-squared test considered significant at P ≤ 0.05.8 Interactions in the final multivariable models were examined for statistical significance and biological plausibility.
Performance evaluation of multivariable models
Predicted probabilities of being seropositive for anti-HCV were calculated for study population members using the two logistic regression equations. The ability of the models to differentiate between individuals seropositive and seronegative for anti-HCV (discrimination) was quantified using the area under the curve (AUC) of the receiver operating characteristic (ROC) curves.9–11 AUC values can range from 0.5 (purely random discrimination) to 1.0 (perfect discrimination). Calibration, an assessment of how closely predicted outcomes agree with actual outcomes, was graphically assessed by plotting the observed frequencies of anti-HCV seropositivity against predicted probabilities.12,13 The ratio of the means of predicted and observed anti-HCV seroprevalence was also calculated.13
Evaluation of the performance of the two predictive models was complemented by a statistical comparison of the AUC values to assess the better discriminating model. A bootstrap approach was used to generate the statistical variance needed for the comparison. Two-hundred samples were generated by sampling with replacement of the 2004 seroprevalence survey sample. Both predictive models were applied to the bootstrap samples generating two-hundred ROC curves and associated AUCs for each model. The distributions of the AUCs for both models were compared and the model producing AUC values closest to 1.0 was identified using the Mann–Whitney non-parametric test. A statistically significant difference was defined as a P-value ≤0.05.
Validation of the better performing multivariable model
The better performing of the two prognostic models was then internally validated to adjust for the statistical phenomenon of over-optimism, where a predictive model's performance in the population from which it was developed often over-represents its performance in other populations.13,14 A bootstrap validation technique was applied to quantify the over-optimism of the final model and to adjust its predictive accuracy.15,16 Another 200 random samples of 14 416 were drawn with replacement from the original sample. A prediction model was fitted to each bootstrap sample using the same criteria for selection as in the original model. The predictive accuracy of each bootstrap model was estimated in the bootstrap and then in the original sample to quantify the difference. The 200 differences were then averaged to a single estimate of the over-optimism in the model's predictive accuracy, which was then subtracted from the initial AUC of the model. To account for over-optimism in future populations, the fitted model was then recalibrated by estimation of a shrinkage factor applied to the regression coefficients of each predictor using the heuristic method of van Houwelingen et al.17
Results of the univariate analysis are presented in table 1. The following variables were associated with seropositivity for anti-HCV and significant at the 5% level: lifetime illicit intravenous (IV) drug use, an elevated ALT level, stay in prison, transfusion of blood or blood products <1992, having a tattoo, being born in a high or moderate/high HCV prevalence region, being 30 years or older, receipt of MWPI, having undergone at least one surgical procedure, lifetime illicit intranasal drug use, a previous screening test for HCV, French region of residence, possession of private health insurance, the duration of unemployment, having attained a high school diploma or equivalent, having undergone a detoxification for alcohol, a stay in a psychiatric institution, professional needle-stick injury (needle contaminated by human body fluids), having a partner that uses IV drugs, having at least 10 sexual partners in lifetime.
Professional needle-stick injury (needle contaminated by human body fluids)
Professional spill of human body fluids onto damaged skin or other mucosal surface
Partner using IV drugs
Number of sexual partners in lifetime
↵a: Observed prevalence (%) in the study sample. Values not extrapolated to the French population
↵b: French Agency for Accreditation and Evaluation in Health (ANAES) guidelines definition of high HCV prevalence regions: South-east Asia, Middle-East, Africa and South America3
↵c: Based on WHO centralised data: moderate/high HCV prevalence regions (defined as ≥2.5%): North Africa, Middle-East, Sub-Saharan Africa, Indian subcontinent, Asia and the Pacific)6
↵d: MWPI (Medical Welfare for Precarious Individuals): an indicator of low socio-economic status
↵e: Post-study classification of alcohol consummation as ‘moderate/excessive’ on average number of alcohol drinks consumed per week during the 3 months prior to questionnaire administration. Excessive: average ≥21 glasses for women and ≥28 for men
Multivariable predictive models
Table 2 shows the results of the multivariable logistic regression analysis performed on all 14 416 individuals. In the ANAES screening guidelines (2001) model, seropositivity for anti-HCV was independently associated with lifetime illicit IV drug use, an elevated ALT level, receipt of a transfusion of blood or blood products pre-1992, having a tattoo, acupuncture and birth in a country with a high anti-HCV prevalence (table 2). No statistically significant interaction was identified.
↵a: Adjusted OR for predictors of the final model accounting for over-optimism calculated by the heuristic method of van Houwelingen et al.17
↵b: ANAES guidelines definition of high HCV prevalence regions: South-east Asia, Middle-East, Africa and South America3
↵c: Based on WHO centralised data: moderate/high HCV prevalence regions (defined as ≥2.5%): = North Africa, Middle-East, Sub-Saharan Africa, Indian subcontinent, Asia and the Pacific)6
↵d: MWPI (Medical Welfare for Precarious Individuals): an indicator of low socio-economic status
The 2004 SPS model showed that seropositivity for anti-HCV was independently associated with lifetime illicit IV drug use, an elevated ALT level, receipt of a transfusion of blood or blood products pre-1992, having a tattoo, birth in a country with a moderate-high anti-HCV prevalence, age, receipt of MWPI, underwent at least one surgical procedure, lifetime illicit intranasal drug use and a previous HCV screening test (table 2). The variable sex was forced in this multivariable model due to its known interactions with risk factors such as IV drug use and blood transfusion. A statistically significant interaction was observed between sex and IV drug use (OR of interaction term = 7.1, P = 0.041); results have been adjusted for this interaction in table 2.
Performance of predictive models
Both models had a good discrimination capacity with AUC values above 0.80. At 0.869 (95% CI 0.861–0.873), the AUC value of the 2004 SPS model is closer to the ideal value of 1.0 than the ANAES model value of 0.821 (95% CI 0.810–0.824). The calibration plots for the study sample show good calibration of the 2004 SPS model for anti-HCV seroprevalence between 1% and 3% (figure 1a) and poorer calibration of the ANAES model (figure 1b). The ratio between mean predicted and observed prevalence of the 2004 SPS model was 1.07 and thus close to the optimal value of 1.0, whereas the ratio of 0.67 for the ANAES model indicated that the predictions from this model were systematically too low.
Calibration plot of the actual prevalence of antibodies to HCV against predicted probabilities for the 2004 seroprevalence survey model (a) and the ANAES guidelines 2001 model (b)
The statistical comparison of the bootstrapped AUC distributions of both models indicated that the 2004 SPS model produced AUC values closer to 1.0 and thus better identifies anti-HCV seropositive individuals (P < 0.00001) (figure 2).
Distribution of the areas under the receiver operating characteristic curves from the 200 bootstrapped samples for the multivariable 2004 seroprevalence survey model (grey) and the ANAES 2001 guidelines model (black)
Internal validation of the better performing 2004 SPS model quantified over-optimism in the AUC of 0.01. The AUC was adjusted from 0.87 to 0.86. The adjusted regression coefficients for predictors of the final model that were obtained after application of the shrinkage factor and adjusted ORs are presented in table 2.
This study has identified a combination of personal characteristics that better identifies individuals seropositive for anti-HCV than the ANAES screening guidelines (2001) currently used in France. The transmission routes for HCV are well described in the literature.18–20 The often cited transmission routes of lifetime illicit IV drug use, transfusion of blood or blood products pre-1992, having a tattoo, biochemical evidence of an abnormal liver function and birth in a country with a high hepatitis C prevalence remained independently associated to seropositivity for anti-HCV in both predictive models in this study.
The combination of characteristics identified from the 2004 anti-HCV seroprevalence survey contained several additional characteristics not included in the ANAES guidelines: intranasal illicit drug use, receipt of state-funded MWPI, having undergone at least one surgical procedure and having had a previous screening test for hepatitis C.
While illicit IV drug use is the current most efficient transmission route for hepatitis C infection, illicit intranasal drug use is less frequently reported as a risk factor.18,19 A prevalent case–control study carried out in France between 1997 and 2001 using 450 individuals seropositive for anti-HCV with no known history of blood transfusions nor having ever used IV drugs, found intranasal drug use to be a risk factor for HCV infection.21 This finding was not, however, repeated in a French incident case-control study of individuals sero-converting between 1998 and 2001.22 Two possible explanations for the observed independent association with anti-HCV seropositivity could be a deliberate down-grading of IV drug use to nasal drug use due to social desirability concerns, or transmission of the virus through sharing of blood-contaminated straws or other devices.23,24 A study estimating the prevalence of hepatitis C antibodies in the USA between 1999 and 2002 also found an independent association for non-injection drug use (excluding marijuana).25 The authors reported a correlation between IV drug use and non-injection drug use and suggested that unacknowledged IV drug use could account partially for the observed association. Twenty percent of nasal drug users in this study reported additional IV drug use and so it also remains possible that the observed association between nasal drug use and anti-HCV seropositivity is to some degree attributable to concomitant or previous IV drug use. It was unfortunately not possible to differentiate concomitant or previous IV drug use from the study data-set.
Socio-economic hardship, or poverty, is a factor associated with hepatitis C infection for which screening is not formally recommended despite its demonstrated association with anti-HCV seropositivity in multiple cross-sectional studies.4,25–27 In this study, a low socioeconomic status, as defined by receipt of the state-funded French MWPI, was independently associated with anti-HCV seropositivity. Receipt of this welfare was used as an indicator of socio-economic hardship due to the fact that 75% of French people living under the poverty threshold are benefiting from this financial support.28 The French national survey of 2004 demonstrated an anti-HCV seroprevalence that was 3.5 times higher among the population in receipt of this welfare than among the population receiving the standard national health insurance.4 Low socioeconomic status could be considered an indirect marker of situations at risk for transmission of the virus and to mask other risk behaviours. However, receipt of the MWPI retains its association with seropositivity after adjustment for other risk factors, suggesting that the observed association is not uniquely explained by indirect association to these risk behaviours.
The independent association between having undergone at least one surgical procedure and anti-HCV seropositivity is a result contrary to the published literature. The use of this non-specific risk factor for hepatitis C screening is not generally recommended, as unless clearly defined it amounts to a generalized population screening.3 A French study has additionally shown that a screening strategy including this risk factor is not cost effective in general practice.29 Previous case-control studies carried out in France identified certain nosocomial procedures such as a digestive endoscopy and invasive radiological examination as risk factors for HCV transmission.21,22 In our study, these two risk factors were not associated with anti-HCV seropositivity, even at univariate level, while the broader question of previous surgeries remained in the final model. Because of the cross sectional design of the 2004 survey, it is likely that many prevalent cases of HCV identified in the survey had been infected in the distant past when surgery is thought to have played a greater role in transmission than in recent years.
The fact that having undergone a surgical procedure remains associated to anti-HCV seropositivity after adjustment for the commonly associated risk factors of age and blood transfusion could be explained by HCV transmission via health-care procedures not involving a transfusion of blood or blood products. Nosocomial HCV transmission via unsafe injection practices, including the use of multi-dose vials, has previously been documented.30,31 It also remains possible that non-specific question concerning undergoing a surgical procedure is in fact a confounding factor for another risk factor not included in the 2004 anti-HCV seroprevalence survey questionnaire.
The performance of both multivariable predictive models was evaluated using parameters with implications for a practical screening context. The 2004 SPS model demonstrated better accuracy as defined by better discrimination and calibration than the model based on the ANAES guidelines (2001). While the AUC for the 2004 SPS model is closer to the ideal value of 1.0, the ANAES model also attains a good, if less optimal level, of discrimination. The better performance of the 2004 SPS model was then confirmed by a statistical comparison of the two AUC values.
The ability of a predictive model to differentiate high and low risk patients for the outcome under consideration, discrimination, is a key attribute for a predictive model to be used in screening. The use of ROC curves and the associated AUCs to measure discrimination is well described in the literature.9,10,32 ROC curves are conventionally used with continuous biological markers in order to define an optimal threshold for classifying a screening test as positive in terms of sensitivity and specificity. In this study, ROC curves have been used for a combination of characteristics that are principally dichotomous in nature. In order to be useful in deciding who to screen, a threshold of positivity would need to be determined for the retained predictive model based on pre-defined desired sensitivity and specificity values.
To adjust maximally for the over-optimistic performance of a predictive model, an external validation is recommended.13,14 However identification of a pertinent external population for validation is difficult and not always feasible. The use of a boot-strapping procedure for an internal validation enables evaluation of the model's predictive performance on multiple data sets generated from the original study sample and is the optimal form of internal validation.13,14 This procedure carried out on the better performing 2004 SPS model demonstrated the presence of a low level of such over-optimism. When adjusted for this phenomenon, the discriminative power of the 2004 SPS model remained high at 0.86.
This study is subject to limitations, principally concerning the representative nature of the ANAES guidelines model. The 2004 seroprevalence survey questionnaire was not designed to evaluate the ANAES guidelines of 2001 and so does not cover all risk factors cited in the guidelines. Certain risk factors could not directly be matched to variables in the 2004 survey, for example the notion of tattooing and piercing with non-‘single-use’ instruments, and having an elevated ALT level of ‘unknown cause’. The best approximation to these risk factors was used. Other ANAES risk factors such as being born of a hepatitis C infected mother, having received health-care in a country reputed to have a high hepatitis C prevalence and being seropositive for HIV could not be matched to the 2004 questionnaire and were not included in the ANAES model.
Public health perspectives
The practical outcome of identifying a combination of screening criteria that better identifies individuals seropositive for anti-HCV than the current guidelines, would be the use of the associated predictive rule for development of a decision-making screening tool for the French general population. This approach has previously been described for hepatitis C in the USA and for Chlamydia trachomatis in the Netherlands.33,34 Such a tool would consist of a hierarchy of questions based on the identified combination of personal characteristics in the predictive model. The provided answers would lead to a decision on whether or not to propose a test.
Before development of this screening tool could take place, several practical issues would have to be addressed. An evaluation of this model in terms of sensitivity, specificity, positive and negative predictive values would be required to define the optimal threshold for proposition of a screening test. Considering the potential clinical severity of undetected chronic hepatitis C infection (cirrhosis of the liver and hepatocellular carcinoma35), the authors would favour attributing a high sensitivity to the model. The impact of a principally sensitive tool for detection of anti-HCV seropositive individuals on the associated positive predictive value and on the absolute number of screening tests to be administered in a low prevalence country such as France would also require assessment. A cost-effectiveness analysis of the screening tool would be necessary.
This study has identified a combination of screening criteria for hepatitis C that could facilitate better targeting of at risk individuals. The proposed combination includes additional criteria that have never figured in the French screening guidelines and that may help identify individuals missed by the current screening programme. Several practical issues would have to be addressed before practical implementation of the proposed combination of screening criteria was feasible.
The authors wish to acknowledge Drs Margaret Pepe, Todd Alonzo and Ewout Steyerberg for providing technical advice during development of the protocol. This work was presented orally as a 10-minute presentation at the 2007 European Scientific Conference on Applied Infectious Disease Epidemiology (ESCAIDE) in Stockholm on 20 October 2007.
Conflicts of interest: None declared.
The current French HCV-screening programme is not yet reaching all populations at risk.
This study aimed to identify a combination of screening criteria that would better identify at risk individuals than the current screening guidelines.
We identified a broader combination of personal characteristics that better identifies individuals seropositive for hepatitis C antibodies which includes illicit intranasal drug use, low socioeconomic status and a previous HCV screening test.
The proposed combination of HCV screening criteria could be used to develop a decision-making screening tool for the general French population.
Agence Nationale d’Accréditation et d’Evaluation en Santé, Association Française du Foie, Société Nationale Française de Gastro-entérologie. Dépistage et traitement de l’hépatite C. Conférence de consensus des 16 et 17 janvier 1997 (Screening and treatment of hepatitis C. Consensus conference of the 16 and 17 January 1997). Gastroenterol Clin Biol 1997;21:45-55.
Agence Nationale d’Accréditation et d’Evaluation en Santé. Screening for hepatitis C – populations to screen and screening methods. Dépistage de l’hépatite C – populations à dépister et modalités du dépistage. Recommandations du comité d’experts uni par l’ANAES (Screening for hepatitis C – populations to screen and screening methods Recommendations of the expert committee united by the ANAES). Agence Nationale d’Accréditation et d’Evaluation en Santé. 2001 [(4 March 2009, date last accessed)]. http://www.snfge.asso.fr/01-Bibliotheque/0D-Pratiques-cliniques/HAS/depistage_vhc_2001_rapport.pdf.
Institut de Veille Sanitaire. Surveillance de l'activité de dépistage de l'hépatite C en France, réseau de laboratoires Rena-VHC (Surveillance of the screening activity for hepatitis C in France, the Rena-VHC laboratory network: epidemiological data 2000-2006). [(4 March 2009, date last accessed)]. http://www.invs.sante.fr/surveillance/hepatite_c/default.htm.
. Mandell, Douglas and Bennett's principle and practice of infectious diseases. Vol. 2, Microbes and problems. 5th rev edn. MandellGL, BennettJE, DolinR, editors. Philadelphia, PA: Churchill Livingstone; 2000. p. 1736-59.
. Enquête auprès des bénéficiaires de la CMU (novembre 2000), principaux résultats [Study of those receiving the medical welfare for precarious individuals (November 2000), principle results]. Série statistiques, n°41. Document de travail. Paris: Ministère des affaires sociales, du travail et de la solidarité, Ministère de la santé, de la famille et des personnes handicapées. 2002 [(4 March 2009, date last accessed)]. http://www.sante.gouv.fr/drees/seriestat/pdf/seriestat41.pdf.
. Viability of the area under the receiver operating characteristic curve in the diagnostic evaluation of liver fibrosis markers: impact of biopsy length and fragmentation. Aliment Pharmacol Ther 2007;25:733-9.
Lisa A.King, YannLe Strat, ChristineMeffre, ElisabethDelarocque-Astagneau, Jean-ClaudeDesenclosEur J Public Health(2009)19 (5):
527-533DOI: http://dx.doi.org/10.1093/eurpub/ckp112First published online: 10 August 2009 (7 pages)