Establishing a sample-to cut-off ratio for lab-diagnosis of hepatitis C virus in Indian context

A potential new way to facilitate HCV elimination: The prediction of viremia in anti-HCV seropositive patients using machine learning algorithms

BACKGROUND AND STUDY AIMS

The present study was undertaken to design a new machine learning (ML) model that can predict the presence of viremia in hepatitis C virus (HCV) antibody (anti-HCV) seropositive cases.

PATIENTS AND METHODS

This retrospective study was conducted between January 2012-January 2022 with 812 patients who were referred for anti-HCV positivity and were examined for HCV ribonucleic acid (HCV RNA). Models were constructed with 11 features with a predictor (presence and absence of viremia) to predict HCV viremia. To build an optimal model, this current study also examined and compared the three classifier data mining approaches: RF, SVM and XGBoost.

RESULTS

The highest performance was achieved with XGBoost (90%), which was followed by RF (89%), SVM Linear (85%) and SVM Radial (83%) algorithms, respectively. The four most important key features contributing to the models were: alanine aminotransferase (ALT), aspartate aminotransferase (AST), albumin (ALB) and anti-HCV levels, respectively, while "ALB" was replaced by the "AGE" only in the XGBoost model.

CONCLUSION

This study has shown that XGBoost and RF based ML models, incorporating anti-HCV levels and routine laboratory tests (ALT, AST, ALB), and age are capable of providing HCV viremia diagnosis with 90% and 89% accuracy, respectively. These findings highlight the potential of ML models in the early diagnosis of HCV viremia, which may be helpful in optimizing HCV elimination programs.

The signal-to-cutoff ratios to predict HCV infection among people who inject drugs

People who inject drugs (PWIDs) are primarily the high-risk population for HCV infection. This study aims to determine the optimal cut-off values for predicting HCV infection status based on the Signal-to-Cutoff (S/CO) ratio. In this study, a total of 719 PWIDs’ samples were collected, and performed for screening test by ELISA assay, and followed by RIBA assay and NAT assay to detect HCV antibody and HCV RNA levels, respectively. The findings revealed that the prevalence of HCV infection among PWIDs was 54.66% (393/719), and the false-positive rate of HCV antibody detection by ELISA assay among PWIDs was only 3.85% (16/416). In addition, when the optimal cut-off value for S/CO ratio was 2.0, the sensitivity and specificity of HCV antibody were 100.00% and 93.55%, respectively. And when the optimal cut-off value for S/CO ratio was 21.36, the sensitivity and specificity of HCV RNA positive were 89.90% and 72.73%, respectively. In conclusion, the status of HCV infection can be predicted based on the S/CO ratios of the ELISA assay, which can improve diagnosis and facilitate timely treatment to effectively prevent the spread of HCV infection.

Signal-to-cutoff ratios of current anti-HCV assays and a suggestion of new algorithm of supplementary testing

BACKGROUND

The detection of hepatitis C virus antibody (anti-HCV) is known to have high false-positive rates. Using signal-to-cutoff (S/Co) ratios in reflex supplemental testing, however, could reduce false-positive rates. Here, we analyzed the 2-year data of an anti-HCV assay to understand the significance of the S/Co ratio and make a new algorithm by confirming with a second anti-HCV assay.

METHODS

We reviewed 32,573 samples of the Architect assay (Abbott Diagnostics) from a tertiary hospital. Retests with the Elecsys (Roche Diagnostics) and Vitros (Ortho Clinical Diagnostics) assays were performed in 346 anti-HCV-positive samples. HCV RNA PCR and recombinant immunoblot assay (RIBA) were performed in 147 and 11 anti-HCV-positive samples, respectively.

RESULTS

Among 32,573 samples, 446 (1.37%) yielded positive results and 32,127 (98.6%) yielded negative results. Concordance rates in low S/Co samples (0.9-10.0) were 35.2%, 43.8%, and 81.9% for the Architect-Elecsys, Architect-Vitros, and Elecsys-Vitros comparisons, respectively. Correlation coefficients of S/Co ratios were as follows: Architect-Elecsys, 0.20; Architect-Vitros, 0.42; and Elecsys-Vitros, 0.46. In logistic regression, the S/Co value for predicting positivity with 95% probability was 3.13, while that for predicting 50% probability was 8.85. S/Co ratios of 1.70-3.34 showed one reactive antigen out of five antigens, and S/Co ratios of 13.54-17.72 showed three to five positive reactions out of five antigens used in the RIBA.

CONCLUSIONS

Supplementary testing of anti-HCV screening results is necessary to distinguish between true positivity and biological false positivity for anti-HCV. In this study, we presented an algorithm of supplementary testing by a retest with a second reagent, which could be useful in clinical laboratories.