Risk prediction models for targeted testing of HIV, hepatitis B and hepatitis C: a systematic review and meta-analysis

Abstract Background Diagnosing human immunodeficiency virus (HIV), hepatitis B virus (HBV), and hepatitis C virus (HCV) infections in general population settings is challenging. We conducted a systematic review and meta-analysis of prediction tools designed to help identify individuals at risk of these blood-borne viruses. Methods We included studies on individuals of any age at risk of blood-borne viruses from healthcare, community settings, and national databases. We searched the Web of Science, MEDLINE, EMBASE, and CENTRAL databases (from database inception to 2023) and used the Prediction model Risk Of Bias ASsessment Tool (PROBAST) to evaluate the quality and systematic risk of bias of these studies. We extracted model accuracy using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. A mixed-effects model (for AUC) and bivariate random-effects model (for sensitivity/specificity) were used to generate pooled values for these studies. Results Of the 41, 585 records, 71 were included, covering over 31 million participants and more than 65, 000 cases of blood-borne viruses. We examined 67 models: 47 for HIV, 13 for HCV, 5 for HBV, and 2 from studies that assessed multiple viruses separately. The studies were conducted in 41 low- and middle-income and 30 high-income countries. They covered 11 different populations (including men who have sex with men, the general population, and women), 8 types of settings (including sexual health, secondary care, and primary care) and 7 types of risk factors (behavioural, clinical, and demographic). The methods comprised traditional regression (n = 50), machine-learning models (n = 17), and others (n = 4). The risk of bias was high in 64 studies and low in seven. Among 33 studies reporting mean and 95% CI, pooled AUC values were 0. 73 (95% CI: 0. 67–0. 80, I² = 74%) across HIV studies (including 8 machine-learning models), 0. 80 (0. 73–0. 86, I² = 86%) for HCV (including 2 machine-learning models) and 0. 79 (0. 76–0. 81, I² = 93%) for HBV (including 3 machine-learning models). Conclusions Significant heterogeneity exists in blood-borne virus prediction accuracy across diverse settings and populations, with a high risk of bias. Contributions from primary care were limited, and evaluation and reporting were inconsistent. Developing and evaluating effective models for the combined risk assessment of HIV, HBV, and HCV in general population settings remains a priority.