Exploration and time-serial validation of logistic regression models composed of multiple laboratory tests for early detection of HCV-associated hepatocellular carcinoma

Abstract

Background: We developed a laboratory test-based regression model for early detection of hepatocellular carcinoma (HCC) associated with HCV in its surveillance.

Methods: This matched case-control study was conducted by enrolling 452 patients with chronic hepatitis and/or cirrhosis, including 129 patients complicated with HCC. One-to-one propensity score matching was performed by referring to sex, age, and fibrosis-4 index, which resulted in 102 patients each in HCC and non-HCC groups. Logistic regression models (LRM) for distinguishing the two groups were explored from variable combinations of laboratory tests. The model was validated by our new scheme of applying it retroactively to trimonthly previous datasets.

Results: Models with a practical level of diagnostic accuracy (C-statistic) were α-fetoprotein (AFP) alone (0.810), LRM3 comprising AFP, AST, and ALT (0.850), and LRM4 comprising AFP, AFP/(AST × ALT), and AST (0.862). After retroactive application of each model, LRM4 showed the highest distinction of the two groups at -12M, -6M, -3M with C-statistics of 0.654, 0.786, 0.834, respectively. LRM4 was accurate even after limiting cases to early-stage HCC.

Conclusions: LRM4 was proved useful in prompting clinicians to perform timely image study in the surveillance. The retroactive validation scheme is applicable to assess diagnostic models of other neoplastic diseases.

Keywords: HCC surveillance; Laboratory informatics; Multivariate logistic regression; ROC analysis; α-fetoprotein.