Page 9 - Read Online

Page 4 of 13 Tong et al. Hepatoma Res 2019;5:36 I http://dx.doi.org/10.20517/2394-5079.2019.005

Predictors of tumor growth rate
Bivariate analysis - The bivariate analysis for assessing each categorical predictor vs. log TGR was computed
using t-tests/analysis of variance. The correlation between log TGR and continuous variables was computed
10
via the Spearman correlation (r ).
s
Multivariate analysis - The multivariable regression tree (binary partition) analysis was used to determine
the simultaneous association between log TGR and 19 potential predictors, including age, gender, ethnicity,
10
HCC surveillance, serum albumin, serum AFP, platelet count, cirrhosis, diabetes, initial tumor size, HBV
or HCV infection, and antiviral treatment. For hepatitis B patients, HBV genotype, HBV DNA, precore
mutation, basal core promoter mutation, and HBeAg values were included. For hepatitis C patients, HCV
RNA and genotype were included. In this tree model, every value of each predictor variable was considered.
Patients with slow vs. fast TGRs were separated via a progression of binary splits (partitions). The best split
was determined by the impurity criterion, a reduction of the residual sum of squares due to the binary split
(GINI criterion). Missing values were allowed. Each split resulted in one parent node and two child nodes.
Child nodes, in turn, were split until further splits did not significantly improve the predicted TGR. The final
[15]
result was an intuitive and interpretable decision tree . A P < 0.07 was considered statistically significant.
Predictors of recurrence-free survival
Predictors of HCC recurrence-free survival were analysed. The outcome (event) was HCC recurrence or
death. The primary predictor was log TGR. The other 9 potential predictors were age, gender, ethnicity,
10
HCV or HBV, diabetes, cirrhosis, macrovascular invasion, HCC surveillance, and the most definitive
treatments (OLT, surgical resection, RFA, PEI, TACE, chemotherapy, or supportive) for a total of 10 potential
predictors. There were 125 events, 39 HCC recurrences and 85 deaths with no recurrence.

Bivariate analysis - Hazard ratios (HR) for each potential predictor, ignoring the other 9 predictors, were
computed along with its 95% confidence bounds and P-values. Restricted cubic splines were used to
determine if the relation between a continuous predictor vs. the log hazard ratio was linear.

Multivariate analysis - The 10 potential predictors simultaneous to the event rate were assessed using a Cox
proportional hazard model. A backwards minimal AIC search was used to determine which of the potential
predictors were significant, with the restriction that log TGR was included in all models. For the final
10
model, all possible two-way interactions were evaluated. Statistical significance was taken as P < 0.07. Model
accuracy was assessed using Harrell’s C concordance statistic with values of C ranging from 0.50 (worse) to
1.0 (best).

RESULTS
The baseline characteristics of 164 HCC patients who had two consecutive imaging studies with either
MRI or CT scans prior to treatments are listed in Table 1. The average age was 64.48 ± 10.38 years, 64.6%
were male, and the majority were Asian (64.0%), followed by white (18.3%), Hispanic (14.0%), and African
American (3.70%). Hepatitis B infection was detected in 39.6% of patients, Hepatitis C infection in 59.8%,
and the remaining patients were co-infected with both viruses. In the HBV infected HCC patients with
measurable tests, 21.5% were HBeAg positive, 29.2% were genotype C, 30.8% had basal core promoter
6
mutations, 23.1% had precore mutations, and the mean HBV DNA level was 2.41 × 10 IU/mL (IQR: 1.00-1.23
5
× 10 ). In the HCV infected HCC patients with measurable tests, 45.9% had genotype 1 and the mean HCV
6
6
RNA was 1.44 × 10 IU/mL (IQR: 594.5 - 1.27 × 10 ). The mean albumin level was 3.80 ± 0.66 g/dL, platelet
3
count was 138,000 ± 75,600 mm , and AFP level was 45.2 ± 11.8 ng/mL. Of 164 HCC patients, 68.9% were
detected by surveillance. 19.5% had diabetes, 78.7% had cirrhosis, and 5.50% had macrovascular invasion.