Professional Documents
Culture Documents
1263-1269 (1996)
We applied a multivariate analysis to a large series of serum in the evolution of chronic liver diseases from cirrhosis to
biochemical tests in an attempt to identi1y a function that HC.
could efficiently discriminate cirrhosis from hepatocellular
carcinoma (HG). We analyzed two successive temporal LNDENG TERMS: a-fetoprotein#{149}alkaline phosphatase . lactate
cohorts (1987-90; 1991-94) of HC and cirrhotic patients, dehydrogenase . ‘y-gbutamyltransferase . aspartate aminotrans-
all histologically classified (first cohort: 69 cirrhosis and 39 ferase #{149}
isoenzymes copper #{149}
#{149} statistics
HG; second cohort: 66 cirrhosis and 38 HC). Using data
from the first temporal cohort of patients, we obtained a
Hepatocebbubar carcinoma (I-IC), one of the most frequent
discriininant function based on seven serum analytes: a-fe-
human tumors [1], commonly (in 95% of cases) evolves from
toprotein, the hepatic isoenzyme of alkaline phosphatase,
cirrhosis, particularly in areas such as Southern Italy that have a
lactate dehydrogenase isoenzyme 5, total ‘y-glutamyltrans-
high incidence of hepatitis C and B viruses [2]? Hepatitis B virus
ferase (GGT), GGT isoforms complexed with low-density
is carcinogenic [3], and the C virus has been implicated in the
lipoprotein, aspartate aminotransferase, and copper. The
pathogenesis of liver cancer [3, 4]. Because of the evolution from
same panel of anabytes emerged when the second cohort
chronic hepatitis of various origins to, in most cases, cirrhosis
was tested and also when both cohorts were tested to- and then liver cancer, the need often arises to differentiate liver
gether. In the two successive cohorts (total, 212 patients) cirrhosisfrom HC. Furthermore, because HC is frequently
with a prevalence of cirrhosis vs HC of -2:1, the discrimi- superimposed on preexisting cirrhosis, monitoring cirrhotic
nant function correctly classified 93% of cases, the highest patients for early recognition of neoplasia is essential for opti-
percentage of correct classification of the two diseases mum treatment of HG. Several diagnostic procedures used thus
obtained so far by laboratory approaches. Validation with far have proved to be less than satisfactory. Computed tomog-
the jackknife reallocation statistical algorithm confirmed raphy has a low sensitivity in detecting neoplastic evolution in
these results. In addition, of six patients with liver cirrhosis cirrhotic patients [5]; similarly, magnetic resonance imaging has
for whom we had the opportunity of following up and bow diagnostic efficiency for correctly classifying malignant vs
observing the evolution to HG, five were classified as HC at nonmalignant liver diseases [6, 7]. In other cases, the instrumen-
diagnosis by the multivariate discnminant analysis; i.e., tal approaches may be more sensitive but are based on invasive
discriminant analysis provided a diagnostic bead time of procedures, such as ultrasound scanning-guided fine-needle
6-12 months over histology. This discriminant function, aspiration [8]. Finally, the clotting disorders typical of advanced
based on easy-to-perform serum biochemical tests, may cirrhosis often preclude liver biopsy via baparoscopy.
help solve a fundamental problem of differential diagnosis Within the realm of biochemical signals, high serum con-
centrations of cs-fetoprotein (AFP) are claimed to be strongly
predictive of HG [9], but AFP has a low diagnostic sensitivity
(-50%) and is thus not a great aid in differential and early
Dipartimento di Biochimica e Biotecnologie Mediche, Universit3 di Napoli
“Federico II,” via S.Pansini 5,1-80131 Naples, Italy, and CEINGE Biotecnologie
Avanzate, Naples, Italy.
2 Cattedra di Gasnoenterologia, Facoltl di Medicina e Chirurgia, Universith
di Napoli “Federico II.” Nonstandard abbreviations: HG, hepatocellular carcinoma; AFP, a-fetopro-
“Author for correspondence. Fax + 39 81 746 36 50. rein; MDA, multivariate discrimmant analysis; LD, lactate dehydrogenase; AP,
Received December 22, 1995; accepted April 1, 1996. alkaline phosphatase; and GGT, y-glutamyltransferase.
1263
1264 Gastaldo et al.: Discriminant function differentiating between hepatocellubar carcinoma and cirrhosis
diagnosis. Similar conclusions can be drawn for isoforms of AFP ANALYTICAL METHODS
[10, ii] and for various other biochemical signals [12-15]. Serum from each patient was collected and processed within 2 h
After a bong-term study, we recently obtained good results in for the analysisof lactatedehydrogenase (LD) and LD isoen-
differentiating between malignant and nonmalignant ascitesin zyrnes, AFP, total alkaline phosphatase (AP), AP isoenzymes
patients with ascites of unknown origin [16] and between (bone, hepatic, biliary, and intestinal), total y-glutamyltrans-
primary and secondary liver cancer [17]. In both cases we used ferase (GGT), GGT isoenzymes, 5 ‘-nucleotidase, leucine ami-
associations of biochemical analytes selected via multivariate nopepudase, cholinesterase, iron, ferritin, and copper. In addi-
discriminantanalysis(MDA). tion, a panel of other hematochemical indices was also
Well before undertaking these studies, we started to analyze examined: aspartate aminotransferase, abanine aminotransferase,
a large variety of biochemical analytes and subjected the results total bilirubin, cholesterol, triglycerides, glucose, and urea. Iron,
to MDA to select the variables that could best differentiate total serum enzyme activities (U/L), and the other hemato-
between cirrhosis and HG. We wanted to avoid recourse to liver chemical indiceswere evaluated at 37 #{176}G with a Hitachi 737
Table 1. Serum analytes with a significantly different distribution (P <0.01, Mann-WhItney U test) between patients
affected by cirrhosis and those affected by hepatocarcinoma (HC) In the first cohort (1987-90).
Cirrhosis (n = 69) HC
The first cohort of patients was used to select, among all the results of the MDA on the first cohort of patients. First, we used
variables, the linear combination that best discriminated HG the jackknife procedure, in which one patient at a time is
from cirrhosis. To avoid introducing into the model any vari- excluded, the rule is rederived from the remaining data, and the
ables that were correlated among themselves [27], we checked rederived rube is used to classify the excluded patient [29]. This
the correlation between all pairs of anabytes by calculating the algorithm removes much of the bias of the simple reallocation
Pearson r coefficient. Of the highly correlated variables method [27]. Second, we tested the validity of the discriminant
(r >0.50), only the most significant one was used for further function prospectively with 104 patients (38 HG and 66 cirrho-
analysis[28]. The MDA was carried out stepwise with use of the sis) recruited later (199 1-1994).
minimum Wilks’ lambda (ratio between within-groups sum of After the positive check of the validity of the first discrimi-
squares and the total sum of squares) to evaluate the group nant function, we calculated a second function to cover the
discrimination. At the first step, the variable with the greatest overall population, i.e., a pool of the two cohorts studied. The
discriminant power was selected; at each of the following steps, validity of this second MDA was also verified by the jackknife
the selection criteria were reevaluated for all variables not in the reallocation method, which confirmed the correct classification
model, and the one with the largest acceptable criterion value rate of the MDA.
was included. The variables previously selected were reevaluated
to check if they met the removal criterion. Any variable that met Results
this criterion was removed, and the variable selection was In this long-term study of cirrhotic and HG cases we measured
concluded when no more variables met entry or removal a large variety of biochemical tests, including all the major
criteria. markers used so far, to differentiate between the two phases of
The discriminant score, calculated for each patient on the chronic liver disease. Table I shows the analytes that were
basis of a linear combination of variables selected by MDA, was significantly different between the two disorders (P <0.01,
used to classify the patients into one of two groups, according to nonparametric Mann-Whitney U-test) for the first cohort of
Bayes’ rule [29], which takes into account not only the proba- patients. Notwithstanding the high significance of the differ-
bility that a case belongs to one of two groups on the basis of the ences in the mean values for all these analytes, there was always
discriminant score (conditional probability) but also the inher- some degree of overlap between the cases of the two clinical
ent probability that the case belongs to one of the two groups in situations, so that no cutoff value efficiently discriminated HG
the absence of any information (prior probability). In our case, from cirrhotic patients. To enhance the discrimination power,
the observed proportions of cases in each group were used as we entered all of the anabytes into the SPSS computer program
estimates of the prior probabilities because we considered our for MDA, after having bog-transformed their measured values;
sample to be representative of the patient population. The we alsospecifieda prevalenceof 2:1 for cirrhosisvs HG. Total
diagnosticsensitivity and specificity of the discriminationfunc- protein, bilirubin, alanine aminotransferase, cholinesterase,
tion were evaluated by taking into account the rate of misclas- iron,LD isoenzymes 2-4, cholesterol,and triglycerides were
sification of this reallocation method. excluded from the Wibks analysis of the MDA because they were
We used two statistical techniques to cross-validate the correlated to other, more significant, variables [28].
1266 Gastaldo et ab. : Discriminant function differentiating between hepatocellular carcinoma and cirrhosis
Table 2. Differential dIagnosis between cIrrhosis and HC by serum bIochemical Indices selected by MDA In two consecutive
cohorts of patients.
1987-90 cohort 1991-94 cohort
Cases correctly classified, no. (%) Cases correctly classified, no. (%)
Cirrhosis HC n” Cirrhosis HC
Among the variables inserted, the statistical procedure se- Table 3), 9 cases of HG and 4 cases of cirrhosis were incorrectly
lected serum AFP (gfL), liver AP isoenzyme, LD5, GGT, classified. To evaluate the comparative power of the discrimi-
GGT isoforms complexed with low- and very-low-density ii- nant function in differentiating between HG and cirrhosis, we
poprotein, aspartate aminotransferase, and copper (ji.molIL) for plotted a ROG curve for the discriminant function and for other
a discrilninant function that correctly classified 90.3% of the analytes that had previously been found to be significantly
HG and 98.5% of cirrhotic patients (96.0% diagnostic efficien- different between HC and cirrhosis (see Fig. 2). The best curve
cy). Table 2 shows the numerical values for the discriminant in terms of diagnostic efficiency is clearly the one calculated with
equation, the cutoff selected, and the cases correctly classified the discriminant function reported above. The difference be-
for the two diseases in the first temporal cohort of patients (left tween the area under the curve of this ROC plot and the values
panel). Applying this discriminant function to the second cohort obtained with all the other analytes listed in Table 1 was always
(right panel) of patients admitted to the study yielded a diag- highly significant (P <0.001), demonstrating that the MDA is
nostic efficiency of 92.3%, confirming the results obtained in more efficient than any univariate anabyte.
the first cohort. Finally, we pooled the data of both cohorts, and The diagnostic sensitivity of the MDA score for HG detec-
again confirmed that the overall MDA selected with very high tion is 100% in the HG subgroup with >2 lesions and in the HG
diagnostic efficiency the same analytes previously chosen in the subgroup with total lesion diameter >10 cm. However, the
two separate cohorts of patients. The final equation is reported diagnostic sensitivity for single-lesion HG was 69.2%, and that
in Table 3. for HG with a totallesiondiameter <5 cm was 75%.
To further validate the final discriminant equation, we Figure 3 shows the probability for each patient of being
checked by the jackknife reallocation
it algorithm and again
confirmed its high discriminatory power: diagnostic sensitivity
85.5%; diagnostic specificity 96.9%; and overall diagnostic :
efficiency 93.2%.
The scattergram in Fig. I shows the numerical MDA score ‘
for each patient, classified as cirrhosis or HG on the basis of ...
histology. At the cutoff chosen by the MDA (0.70 as reported in ...
2 #{149}....
dIagnosIs between cirrhosis
Table 3. DIfferential and HC by #{149}:#{149}:::
60
Table 4 shows the values of the MDA in these six patients at
ZU) diagnosis, together with the probability each patient at that time
0.7
/ the diagnostic efficiency of the variables in our panel. Fluctua-
tions of LD concentration reflect the impaired glycobytic me-
a
0.6 tabolism in cancer cells [32]; LD isoenzyme 5 is produced
0.5 predominantly by muscle and liverand isincreasedin the serum
a
of HG patients [32]. The fraction of GGT complexed to low-
0.4
and very-low-density lipoprotein (an adjunctive tool with which
0.3 to discriminate HG from cirrhosis) is a signal of the cholestasis
.0
a
.0
0
0.2 associated with HG [22]. The hepatic isoenzyme of AP is
Classified as CirTb,,,/
0. overproduced by neoplastic liver cells [33], and serum copper is
-. #{149},
#{149} increased in HG because of the reduced biliary excretion of the
-4 -3 .2 -1 0 1 2 3 4 ion and the increased production by neoplastic cells of fetal
Dlscriminant Patient Score copper-binding proteins [14]. Finally, the increase in serum AFP
Fig. 3. Relation between the probability of being affected by HC and the in HG patients is due to the reexpression of the related gene,
score obtained from the MDA for each patient. which is usually repressed in adult subjects.
1268 Gastabdo et ab.: Discriminant function differentiating between hepatoceblubar carcinoma and cirrhosis
The panel selected via MDA showed a diagnostic sensitivity patients with cirrhosis: CT sensitivity and specificity in 200
for HG of 85%, higher than or comparable with that of other, consecutive transplant patients. Radiology 1994:193:645-50.
sometimes invasive tools, including instrumental approaches. 6. Kamba M, Suto Y, Kato 1. Chondroitin sulfate iron colloid-en-
hanced MR imaging in patients with hepatocellular carcinoma.
For example, for approximately the same number or fewer cases,
Comparison with CT during arterial portography. Acta Radiol
the diagnostic sensitivity of computed tomography for HG was 1994:35:570-5.
70% [5], that of magnetic resonance imaging ranged between 7. Ferrucci iT. Liver tumor imaging. Cancer 1991:67:1189-95.
80% and 90% [6], and that of cytology by fine-needle aspiration 8. Borzio M, Borzio F, Macchi R, Croce AM, Bruno S, Ferrari A, Servida
biopsy for HG was <85% [34]. E. The evaluation of fine-needle procedures for the diagnosis of
An additional result of our study is the probability plot, focal liver lesions in cirrhosis. i Hepatol 1994:20:117-21.
which allows calculation on the basis of the MDA score of each 9. Fu-Sun V. Kong-Nien S. Epidemiology and early diagnosis of
primary liver cancer in China. Adv Cancer Res 1986:46:85-8.
patient’s probability for being affected by HG. This can help
10. Hirai H, Taketa K. Lectin affinity electrophoresis of alpha-fetopro-
physicians to plan the best therapeutic strategy for each patient,
incidence. In: Galen RS, Gambino SR, editors. Beyond normality: 31. Ikeda K, Saitoh 5, Koida I, Tsubota A, Arese Y, Chayama K.
the predictive value and efficiency of helical diagnoses. New York: Kumada H. Imaging diagnosis of small hepatocellular carcinoma.
Wiley Biomedical Publishers, 1977:10-4. Hepatology 1994;2O:82-7.
26. Zweig MH, Campbell G. Receiver-operating characteristic (ROC) 32. CastaldoG, OrianiG, Cimino L, Topa M, Budillon G, Salvatore F,
plots: a fundamental evaluation tool in clinical medicine [Review]. Sacchetti L. Serum lactate dehydrogenase isoenzyme 4/5 dis-
ClinChem 1993;39:561-77. criminates between hepatocarcinoma and secondary liver neopla-
27. Norusis Mi, ed. SPSS/PC. Advanced statistics manual. Chicago: sia. Clin Chem 1991;37:1419-23.
SPSS, 1990. 33. Wolf PL. Clinical significance of serum high-molecular mass
28. Han P, Fung KP. Discriminant analysis of iron deficiency anaemia alkaline phosphatase, alkaline phosphatase-lipoprotein X com-
and heterozygous thalassaemia traits: a 3-dimensional selection plex, and intestinal variant alkaline phosphatase. i Clin Lab Anal
of red cellindices.ClinLab Haematol 1991;13:351-62. 1994;8:172-6.
29. Wasson JH, Sox NC, Neff RK, Goldman L. Clinical prediction rules: 34. Angeles-Angeles A, Gamboa-Dominguez A, Velazquez-Fernandez
application and methodological standards.N Engl i Med 1985; D, Munoz-FernandezL. Hepatic fine needle aspiration biopsy.