Discriminant analysis in diagnosing carcinoma of the pancreas and of the papilla of Vater GIUSEPPE MONTALTO1, LEONARDO FICANO2, ANTONIO CARROCCIO1, FRANCESCOD'ARPA2, LUIGI GRECO3, MAURIZlO SORESI1, and PASQUALE SALVO2 1Istituto di Medicina Interna e Geriatria, via del Vespro 143; 90127 Palermo, Italy 2Cattedra di Patologia Medica IIe Cattedra di Patologia Chirurgica B, Universit~ di Palermo, Italy 3Dipartimento di Pediatria, Universit/t di Napoli, Italy
Abstract: The clinical and biochemical presentation of car-
cinoma of the pancreas (PC) and of the papilla of Vater (CPV) are very similar, and, consequently, detailed investigations are required to correctly distinguish between them. The aim of the present study was to select the clinical and biochemical variables that would most efficiently discriminate the precise site of tumor origin. The study group consisted of 72 patients with PC and 22 patients with CPV consecutively hospitalized in our department. The following clinical parameters were considered: age, asthenia, anorexia, vomiting, weight loss, pain, fever, pruritis, and constipation; the biochemical parameters considered were total, direct, and indirect bilirubin, glucose, alkaline phosphatase, gamma glutamy transferase, transaminase, total protein, amylase, and occult blood in stools. The results indicated that in the initial phase of PC the most frequent clinical parameters were weight loss (P < 0.0001), anorexia (P < 0.02), constipation (P < 0.001), and pruritus (P < 0.01). In contrast, in CPV, fever (P < 0.003) was most frequent in the same phase. There was a statistically significant difference in occult blood in stools (P < 0.0001), total (P < 0.03) and direct bilirubin (P < 0.02), alkaline phosphatase (P < 0.05), and transaminase (P < 0.002) values in the two groups. On discriminant analysis, weight loss, constipation, pruritus, nausea, anorexia, and fever were the variables which best discriminated between the two types of tumors. In fact, the presence of weight loss, anorexia, asthenia, constipation, and pruritus correctly classified 87.5% of the patients with PC, while the presence of fever and nausea correctly classified 72.7% of the patients with CPV.
The clinical diagnosis of periampullary cancer (carcinoma of the head of the pancreas, terminal bile duct, duodenum, or of the ampulla) is still based on the classic triad: jaundice, abdominal pain, and weight loss. Unfortunately, however, when these three symptoms are overt, the disease has already reached too advanced a stage to be treated surgically, with chemotherapy, and/or with effective support therapy, and, as a consequence, the diagnosis has no effect on prognosis. A hypothetical diagnosis of these carcinomas at a relatively early stage could be made on the basis of other symptoms and conditions (epigastric or right hypochondrial distension, bloating, nausea, belching, asthenia, anorexia, diabetes mellitus) which are generally associated with, but also precede, the classic triad. Such symptoms and conditions are often ignored by both the patient and the doctor, as they are considered too vague and aspecific to suggest the presence of cancer. The most common location of these tumors is the head of the pancreas, followed by those arising from the mucosa of the ampulla of Vater. 1'2 The clinical presentation of both types is very similar, but they differ in terns of the course of the illness and prognosis. 2-5 The aim of the present study was to select those clinical and biochemical variables that would most efficiently discriminate the precise site of tumor origin.
Patients and methods
Offprint requests to: G. Montalto Received for publication on April 8, 1994; accepted on May 6, 1994
The study group consisted of 22 patients (13 male, 9 female) with carcinoma of the ampulla of Vater (CPV) and 72 patients (38 male, 34 female) with carcinoma of the head of the pancreas (PC) treated consecutively at our department during the period from April 1988 to October 1992. The diagnosis of CPV was made on the
G. Montatto et al.: Periampullary carcinomas
basis of surgical and histological findings in l0 patients and on the basis of endoscopy/biopsy findings in the remaining 12. The diagnosis of PC was established histologically in 42 patients, by surgery in 17, and in the remaining 13 it was made by means of endoscopic retrograde cholangiopancreatography (ERCP), computed tomography (CT), ultrasonography, and laboratory findings. The following parameters were considered: (a) Clinical: age at diagnosis, age at the onset of symptoms, asthenia, anorexia, vomiting, weight loss, pruritus, abdominal pain, fever, jaundice, constipation or diarrhea; (b) Laboratory data: glucose, total protein, total, direct, and indirect bilirubin, alkaline phosphatase, gamma glutamyl transferase, transaminase, amylase, occult blood in stools. The data obtained were analyzed with a SPSS PC+ statistical package. Continuous variables were examined for normality in distribution: since the deviation from normality was always significant, a decimal log transformation was performed before we performed further parametric analysis. Student's t-test was used to evaluate the differences between the means of log transformed data. Pooled or separate variance estimates were used according to the result of the F test for homogeneity of variances. First degree error was set at P = 0.05. Chi square (with Yates' correction, when required) was used to compare the frequencies of dichotomous variables. Discriminant analysis was performed using the stepwise method to enter variables with a significant contribution to discriminate between the two groups, in a multivariate fashion. F-to-enter and F-to-remove were set at 1.0. Variables which did not contribute higher F levels after the inclusion of the other significant variables were excluded.
The highest probability of classifying each individual case in a defined group was obtained by the discriminant equation computed with the contribution of the significant variables.
Table 1 shows the differences between the two groups in the continuous variables. The mean age of CPV patients was 67.7 + 11 years, while the mean age of PC patients was 63.2 _+ 9 years; the difference between the two groups was not statistically significant. The time lapse between onset of symptoms and diagnosis was similar in CPV patients (3.9 months) and in PC patients (4.4 months). Mean values of biochemical parameters were elevated in both groups, but were higher in the PC patients. In fact, there were statistically significant differences in total bilirubin (P < 0.03), direct bilirubin (P < 0.02), alkaline phosphatase (P < 0.05), alamine aminotransferase (ALT) (P < 0.0001), and aspartate aminotransferase (AST) (P < 0.002). Moreover, occult blood in stools was positive in 14 of the 22 patients with CPV and in only 11 of the 72 patients with PC (chi-square 17.8; P < 0.0001). Data are expressed as raw, untransformed values, while t-test and significance levels refer to log transformed values. Table 2 shows the frequency of the main clinical symptoms in the two types of tumor. Of the classic triad symptoms (jaundice, abdominal pain, and weight toss) only weight loss was significantly higher in the PC patients (P < 0.0001). As regards the other symptoms studied, there was a statistically significant difference in fever (P < 0.01), anorexia (P < 0.02), constipation (P < 0.001), and pruritus (P < 0.01).
Table 1. Differences between the two groups in continuous variables Variable Age (years) Latency (months) T. Bil (mg) D. Bil (rag) I. Bil (rag) Alk Phos (U/l) ALT (U/l) AST (U/l) T. Prot (g/l) Amylase (U/l)
*Significant,P < 0.05 CPV, Carcinomaof the papilla of Vater; PC, carcinomaof the pancreas; T. Bil, total bilirubin; D. Bil, direct bilirubin;L Bil, indirectbilirubin;Alk Phos, alkalinephosphatase;ALT, alanine aminotransferase;AST, aspartate aminotransferase;T. Prot, total protein
G. Montalto et al.: Periampullary carcinomas
Table 2. Frequency of clinical variables in the two groups (percentages)
Female sex Weight loss Fever Jaundice Pain Epigastric distension Anorexia Asthenia Diarrhea Constipation Pruritus Nausea Vomiting
Discriminant analysis In order to select those variables more likely to differentiate the CPV from the PC patients, a stepwise discriminant analysis was performed. A first run with all the available variables (clinical as well as laboratory data) suggested that the inclusion of the laboratory variables (raw data or after log transformation) did not add any significant discriminant power to the clinical variables. Since the clinical variables were all expressed in a dichotomous mode, a discriminant equation with only clinical variables was performed. Table 3 shows the results of the discriminant analysis (variables are ordered according to their discriminant power). Wilks' lambda is a measure of the discriminating power between the two groups: Lambda is the ratio of the within-groups sum of squares to the total sum of squares. A value of 1 suggests a complete overlap between groups and values close to 0 a complete separation. The Table shows the cumulative decrease of Wilk's lambda from 0.798 to 0.611 by the stepwise addition of the significant variables. Weight loss was the first (most important) discriminating variable, followed by constipation, pruritus, nausea, anorexia, and fever. The variables not included in the discriminant equation because of a lack of significant
contribution to the differentiation between the two groups (F-to-enter less than 1.0) were: jaundice, pain, epigastric distension, asthenia, diarrhea, vomiting. All but nausea had different frequencies in the two groups, as shown in Table 2. Weight loss, anorexia, asthenia, constipation, and pruritus were more common in the PC group, while fever and, to a lesser extent, nausea, were more common in the CPV group.
Classification analysis Using those clinical variables selected because they significantly contributed to the differentiation between the two groups, we computed a discriminant equation with its corresponding standardized canonical discriminant function coefficients, in order to calculate a discriminant score (D-score) for each patient in the two groups. Patients were then classified by the equation without considering their original group (PC or CPV). Table 4 shows the results of the classification analysis; 84% of the patients were correcly classified by simple clinical variables. More PC patients were correctly classified (87.5%) than CPV patients (72.7%). Obviously, individual patients all had different "strengths" of group membership, according to their individual D-scores. It is possible to plot individual D-scores vs the highest probability of being members of a defined group. Figure 1 shows the plot of D-scores vs group membership probability. Once the D-score is obtained, it may then be wise to consider a strong diagnostic suggestion for a D-score associated with a high (above 80%) probability of being classified in one of the two groups, and then submit to further diagnostic procedures those patients who score lower diagnostic probabilities on the basis of clinical information only, although diagnostic procedures are performed for all patients. This limit would be marked by a D-score below - 1 . 2 for the CPV group and above 0.3 for the PC group. Table 5 gives the canonical discriminant function coefficients of the equation. The standardized
Table 4. Classification of patients
0.798 0.710 0.674 0.640 0.624 0.611
0.00001 0.00001 0.00001 0.00001 0.00001 0.00001
Predicted group membership Actual group
CPV 22 16 (72.7%) 6 (27.3%) PC 72 9 (12.5%) 63 (87.5%) Percentage of patients correctly classified, 84.04%
G. Montalto et al.: Periampullary carcinomas
Table 5. Discriminant function coefficients
Standardized Weight loss Fever Anorexia Constipation Pruritus Nausea
Fig. 1. Plot of D-score vs group membership probability
-1.93 (constant) = -2.54; this would be associated with a 97.8% probability of this patient being in the group of CPV patients. As previously suggested, values between 0.3 and 1.2 should be considered with caution, since they are associated with probabilities of group membership of lower than 80%.
coefficients give an estimate of the relative importance of each variable in the discriminant capability of the function, but they have to be considered in a multivariate fashion, since the variables are intercorrelated. Unstandardized coefficients were readily usable to calculate individual D-scores and to classify patients, plotting the score obtained in Fig. 1. The simple equation is: D-score = (loss of weight 1.74) - (fever 0.61) + (anorexia 0.65) + (constipation 0.58) + (pruritus 0.74) - (nausea 0.92) = 4.84 Calculations are simple, since variables are dichotomous where 0 = absent and 1 = present. So it is a matter of considering only those variables which are present (= 1), and adding to the constant the corresponding coefficient. A patient with loss of weight, anorexia, constipation, and pruritus would have a Dscore of: 1.74 (loss of weight) + 0.65 (anorexia) + 1.29 (constipation) + 0.74 (pruritus) - 1.93 (constant) = 2.50; this patient would thus have a 99.6% probability of being in the group of PC patients. On the other hand, a patient with only fever and no other symptoms would have a D-score of -0.61 (fever)
In the last few decades there has been a marked increase in the frequency of periampullary carcinomas, possibly as a result of improvements in diagnostic technique, but also due to a real increase in frequency, dependent on various changes in life-style over this period. 6-8 The clinical presentation of these types of malignant neoplasm, i.e., PC and CPV, is very similar, creating problems of differential diagnosis; these problems can be solved by invasive examinations, but sometimes only surgical intervention can clearly establish the correct location of the carcinoma. Despite both the clinical and therapeutic similarities, prognosis is very different, as PC has a very rapid evolution, whereas CPV has a slower and more benign evolution. 2'9 The aim of our study was to examine the clinical and biochemical presentations of the two types of carcinoma, in order to differentiate between them and to determine, by using discriminant analysis, which characteristic features, if any, could help to diagnose the correct location of a carcinoma before invasive examinations are resorted to. The time lapse between the onset of symptoms and diagnosis in the two groups was, unfortunately still quite long, although it was shorter than in other studies. 1~ Many of the dyspeptic symptoms presented in the early
G. Montalto et aI.: Periampullary carcinomas stages of the illness were disregarded because they were considered to be aspecific and also because they are very frequent in the healthy population. A more detailed investigation of these symptoms would probably have shortened the time lapse before diagnosis, although it is clear from our study that the location of the carcinoma has a greater effect on prognosis, in terms of the agressiveness and invasiveness of the malignant neoplasm. Among the symptoms we studied there were statistically significant differences between the PC and CPV groups in fever, anorexia, constipation, and pruritus, with a higher frequency of these symptoms being seen in carcinomas of the head of the pancreas. However, the most important symptom associated with pancreatic carcinoma was weight loss (P < 0.000001). In this tumor, in fact, there is an early impairment in digestive function, both in terms of the drop in pancreatic secretion and in terms of the reduction in bile secretion, as suggested by the high number of patients with pruritus and the marked increase in cholestatic parameters. In fact, once the common bile duct is impaired (by compression or infiltration), the jaundice in PC is progressive, if it is not treated surgically, while in Vater carcinomas, the obstruction of the common bile duct is often only partial and intermittent, 1a'12 thus allowing a normal digestive phase. As regards the laboratory data, there was a significant association of the presence of occult blood in stools and CPV; this can easily be explained by the frequent tendency of these tumors to ulcerate and bleed; the blood lost being, by necessity, discharged via the intestine. However, the presence of this sign is rather aspecific, as it may be associated with benign or malignant pathologies anywhere in the digestive system. There was a statistically significant difference between transaminase levels in PC and CPV. The higher levels in PC may be secondary either to an epiphenomenon resulting from the biliary stasis or to the presence of metastatic hepatic lesions. However, in general, the results suggest that in both carcinomas of the head of the pancreas and in those of the more specific area of the ampulla of Vater there are quite early signs and symptoms of the disease which may suggest the existence of a tumor, but, as they are non-specific and also quite frequent in the general population as a whole, they are often disregarded by both doctors and patients. Diagnosis is partially based on the presence of a series of more
577 severe symptoms in PC - - signs of cholestasis, hepatic cytolysis, and weight loss - - which often indicate invasion of the neoplasia and the impossibility of effective treatment. In conclusion, our results, based on the D-score, do offer significant help in suggesting the correct diagnosis; their use is, however, not reliable in lifethreatening conditions. The classification analysis results apply only to the set of patients in the present study; their application to different sets, although advisable, should be made with caution. However, some initial signs, such as anorexia, constipation, pruritus, and loss of weight, if correctly taken into consideration, could indicate the diagnosis of a tumor of the pancreas, whereas other symptoms, such as fever and nausea, especially if associated with occult blood in stools, could suggest a tumor of the papilla of Vater at an earlier stage.
References 1. Nakase A, Matsumoto Y, Uchida K, Honjo I (1977) Surgical treatment of cancer of the pancreas and the periampullary region. Ann Surg 185:52-57 2. Michelassi F, Erroi F, Dawson PJ, Pietrabissa A, Noda S, Handcock M, Block GE (1989) Experience with 647 consecutive tumors of the duodenum, ampulla, head of the pancreas and distal common bile duct. Ann Surg 210:544-556 3. Warren KW, Veidenheimer MC, Pratt HS (1967) Pancreatoduodenectomy for periampullarycancer. Surg Clin North Am 47:639 4. Wise L, PizzimbonoC, Dehner LP (1976) Periampullarycancer. A clinicopathologicstudy of 62 patients. Am J Surg 131:141-148 5. Jones BA, Longer B, Taylor B, Girotti M (1985) Periampullary tumors: Which one should be resected? Am J Surg 149:46-52 6. Hayes DH, BoRon JS, Willis GW, Bowen JC (1987) Carcinoma of the ampulla of Vater. Ann Surg 206:572-77 7. Forrest JF, Longmire WP (1979) Carcinoma of the pancreas and periampullary region. A study of 279 patients. Ann Surg 189:129-138 8. Levin DL, Connelly RR, Devesa SS (1981) Demographic characteristics of cancer of the pancreas: Mortality, incidence, and survival. Cancer 47:1456 9. Herter FP, Cooperman AM, Ahlborn TN, Antinori C (1982) Surgical experience with pancreatic and periampulIary cancer. Ann Surg 195:274-281 10. Balasegaram M (1976) Carcinoma of the periampullary region: A review of a personal series of 87 patients. Br J Surg 63:532537 11. Mong6 JJ, Judd ES, Gage RP (1964) Radical pancreatoduodenectomy: A 22-year experience with the complications, mortality rate, and survival rate. Ann Surg 160:711-715 12. Buckwalter JA, Lawton RL, Tidrick RT (1964) Pancreatoduodenectomy. Arch Surg 89:331-334