The Doylestown Algorithm: A Test to Improve the Performance of AFP in the Detection of Hepatocellular Carcinoma

Abstract

Biomarkers for the early diagnosis of hepatocellular carcinoma (HCC) are needed to decrease mortality from this cancer. However, as new biomarkers have been slow to be brought to clinical practice, we have developed a diagnostic algorithm that utilizes commonly used clinical measurements in those at risk of developing HCC. Briefly, as α-fetoprotein (AFP) is routinely used, an algorithm that incorporated AFP values along with four other clinical factors was developed. Discovery analysis was performed on electronic data from patients who had liver disease (cirrhosis) alone or HCC in the background of cirrhosis. The discovery set consisted of 360 patients from two independent locations. A logistic regression algorithm was developed that incorporated log-transformed AFP values with age, gender, alkaline phosphatase, and alanine aminotransferase levels. We define this as the Doylestown algorithm. In the discovery set, the Doylestown algorithm improved the overall performance of AFP by 10%. In subsequent external validation in over 2,700 patients from three independent sites, the Doylestown algorithm improved detection of HCC as compared with AFP alone by 4% to 20%. In addition, at a fixed specificity of 95%, the Doylestown algorithm improved the detection of HCC as compared with AFP alone by 2% to 20%. In conclusion, the Doylestown algorithm consolidates clinical laboratory values, with age and gender, which are each individually associated with HCC risk, into a single value that can be used for HCC risk assessment. As such, it should be applicable and useful to the medical community that manages those at risk for developing HCC.

Figure 1. Study Design

Model development utilized 360 samples with HBV, HCV and non viral liver disease. After model development and internal validation, external validation was performed by independent analysis of the Doylestown algorithm in three sample sets consisting of over 2,700 patient samples. Samples consisted of those with HBV, HCV and non viral liver disease.

Figure 2. Development of an AFP based algorithm for the detection of HCC

A) The algorithm as developed. B) AUROC for either AFP or the Doylestown algorithm from just the samples from UM. C) AUROC for either AFP or the Doylestown algorithm from patients in the HALT-C set. Dotted line is the line of 95% specificity.

Figure 3. Validation of the Doylestown algorithm in the NCI EDRN sample set and in an HBV infected sample set from Thomas Jefferson University

A) AUROC of AFP or the Doylestown algorithm in the NCI EDRN validation set in all patients. B) AUROC of AFP or the Doylestown algorithm in the NCI EDRN validation set in patients with early stage HCC. C) AUROC of AFP versus the Doylestown algorithm in the Thomas Jefferson University validation group. Dotted line is the line of 95% specificity.

Figure 4. Validation of the Doylestown Algorithm in the UTSW patient set

A) AUROC of AFP or the Doylestown algorithm in the UTSW sample set. B) AUROC of either AFP alone or the Doylestown algorithm in patients with varying ranges of AFP. For graph, Y-axis is the AUC for either AFP or the Doylestown algorithm in the specified patients. For X axis, group 0 are patients with AFP<10; group 1 are patients with 10

Figure 4. Validation of the Doylestown Algorithm in the UTSW patient set

A) AUROC of AFP or the Doylestown algorithm in the UTSW sample set. B) AUROC of either AFP alone or the Doylestown algorithm in patients with varying ranges of AFP. For graph, Y-axis is the AUC for either AFP or the Doylestown algorithm in the specified patients. For X axis, group 0 are patients with AFP<10; group 1 are patients with 10