RIGHT LAB
News Archive
GENOME INSTITUTE OF SINGAPORE
New options using fuzzy logic-based tumour marker profiles
J. Cancer Res. Clin. Oncol. - received: 14 April 1998/Accepted: 14 May 1998
Thomas Keller#, Norman Bitterlich#, Stefanie Hilfenhaus#, Heidrun Bigl=, Thomas L�ser# , Peter Leonhardt
Prof. Dr. med. habil. P. Leonhardt, St�dt. Klinik Leipzig West, Robert-Koch-Klinik, Nikolai-Rumjanzew-Str. 100, 04207 Leipzig
#pe Diagnostik GmbH, Fichtestr. 9, 04275 Leipzig
Deceased
Abstract: The diagnosis of lung cancer and early knowledge of its histological type are very important; however, this is still a difficult subject for the physician. The aim of this study was to improve the diagnostic efficiency of tumour markers in the diagnosis of bronchial carcinoma by mathematical evaluation of a tumour marker profile employing fuzzy logic modelling. A panel of five tumour markers including CYFRA21-1, CEA, NSE, and three additional parameters was determined in 281 patients with confirmed primary diagnosis of bronchial carcinoma of different histology and stage. A further 131 persons, who had acute and chronic benign lung diseases, served as a control group. A classificator was developed using a fuzzy-logic rule based system. The diagnostic value of the combined tumour markers was significantly better than that of the individual markers and of a combination of CYFRA21-1, CEA, and NSE. The discrimination of malignant versus benign diseases was realized with a sensitivity of 87.5% and specificity of 85.5%. The rate of correct classification of small-cell vs non-small-cell lung carcinoma was 90.6% and 91.1%, respectively; for Squamous cell carcinoma vs adenocarcinoma it was 76.8% and 78.8%, respectively. Our detailed analysis has shown that the fuzzy logic system improves diagnostic accuracy up to a rate of 20%, especially in early stages and in patients with all marker levels in the grey area. Our concept proved to be more powerful than measurement of single markers or the combination of CEA, CYFRA21-1, and NSE. Its use may help in distinguishin between malignant and benign disease and make it possible to define different subgroups of patients earlier in the course of the disease.
KEYWORDS: Bronchogenic carcinoma, Lung cancer, Tumour markers, Fuzzy logic, Small cell carcinoma, Non small cell carcinoma
Introduction
At present, the treatment of lung cancer is one of the greatest challenges in medical oncology, because of its high incidence in both men and women and its poor prognosis [Everone 1997]. The latter is the result of both late detection and the limited effectiveness of therapeutic regimens available for advanced lung cancers. Early detection is often the result of an incidental finding, because all diagnostic methods have their limits at small tumour sizes. Expectations were high that tumour markers, when first described, could fill this diagnostic gap. Up to now over 20 tumour markers have been described with varying tumour and organ specificity and sensitivity. However, clinical application of tumour markers is far from ideal, owing to the lack of sensitivity/specificity from the diagnostic point of view. For this reason their use is recommended mainly for monitoring of therapy and postoperative surveillance. The usefulness for the detection of tumours, differential diagnosis of the histological subtypes and staging is viewed skeptical.
Hence, the development of highly sensitive and specific tumour marker-based tests might help physicians to detect cancer at an early stage and enhance the success of therapy. An obvious way of increasing sensitivity is to combine tumour markers, but the gain in sensitivity is counterbalanced by a loss of specificity. Neuron specific enolase (NSE), cytokeratin 19 antibody (CYFRA21-1) and carcinoembryonic antigen (CEA) can be considered established tumour markers for lung cancer. Their overall sensitivity for the detection of lung cancer of any histology is below 50%, but their sensitivity for certain histologies is higher [Ebert et al., 1997, see also references therein]. Serum levels of these three markers are often analysed at the time of the primary diagnosis. Less well known is the worth of the cancer antigens CA72-4, CA15-3 and CA19-9 for the diagnosis of lung cancer. These markers detect malignant processes in other parts of the body, but retain a certain sensitivity for lung cancer. In addition to that, some laboratory markers that are sensitive to malignant processes in general are involved in the malignant - benign decision. Among these markers are CRP - also used to indicate inflammatory processes often associated with neoplasias - and ferritin; an increase in the ferritin concentrations correlates with malignancy and tumour size irrespective of histology [Yang et al., 1995].
We developed a new principle of tumour marker use in diagnosis based on the analysis of a panel of tumour markers and their mathematical processing by means of fuzzy logic. Fuzzy logic modelling takes into account that the description of a decision in terms of �more...less� is more adequate to real problems than by a sharp �yes - no� judgement used by most computer programs. It was first described by Zadeh (1965) and is related primarily to the human way of thinking. Therefore, it is not surprising that the first reported applications addressed medical reasoning [e.g., Kalmanson and Stegall 1975]. Of continued interest is the modelling of decision making on the basis of clinical signs and/or laboratory data [e.g. Sanchez 1977, 1998, Rau et al., 1995]. Numerous groups have performed investigations using fuzzy logic methods in image processing [Pathak and Pal 1986; Hemdon et al. 1996]. However, computer assisted decision support has been a preference in fuzzy logic approaches, because it can include the physician�s knowledge in a manner not achieved by other mathematical methods.
Fuzzy set approaches were found to be superior to the ordinary clear cut definition of normalcy in clinical evaluation of laboratory data [Nishimura et al. 1991]. They should become established in analysis of laboratory data, not with the intention of replacing physicians� knowledge by a computer analysis but to help in objectifying and validating the physicians� decision-making on a basis of high diagnostic efficiency. Here we present a system employing fuzzy logic for the analysis of laboratory data in the differential diagnosis of lung cancer. In this study we set out to determine the diagnostic value of this new tool (fuzzy classificator) for the classification of 1) malignant - benign lung disease 2) non small cell lung carcinoma (NSCLC) - small cell lung carcinoma (SCLC) 3) squamous cell carcinoma (SCC) - adenocarcinoma (AC).
Materials and Methods
Patients
The study includes 281 patients with a primary diagnosis of lung cancer and 131 patients with benign lung diseases, all of whom were treated at a hospital that specializes in lung diseases (St�dtisches Klinikum Leipzig West: Robert-Koch-Klinik). Serum samples were obtained from patients with histologically confirmed lung cancer during the first staging before treatment. Staging was carried out according to the latest TNM classification [UICC IV, Hermanek and Sobin 1992]. The distribution by histological subtype is shown in Table 1. Only patients with known histology of a primary lung cancer, confirmed by bronchoscopy, cytology, and histology were included in the study. The gold standard was set by an extensive morphology (cytology / histology). In the case of mixed histology, the patient was grouped according to the dominant histology. Three cases of SCC - AC mixed histology, three cases of large cell lung carcinoma (LCLC) and two cases of less well defined NSCLC were grouped together under �others�. One patient had mixed SCLC-NSCLC histology and is therefore listed separately. Exclusion criteria were cancer therapy and relapse. Pulmonary metastases of extrapulmonary tumours, mesthelioma, sarcoma and lymphoma were not included in the study. Benign lung diseases included tuberculosis, sarcoidosis, benign tumours, chronic ischaemic heart diseases, pneumonia, asthma, pneumoconiosis, fibrosis, chronic inflammatory lung diseases, diseases of the respiratory tract, and variuos allergically and immunologically determined diseases. Patients with benign pulmonary diseases but with a case history of malignant disease were excluded from the study.
The patient population was divided by random selection into two subpopulations: one subpopulation to develop the classification system (development data) and the other to test it (test data). The condition for the random selection was a homogeneous distribution of patients over stages and histology. The mathematical model was developed and tested with the development data including 170 persons with confirmed malignant lung cancer and 80 persons with a benign lung disease (see Table 1). The data on the remaining 163 patients (111 with malignant / 52 with benign disease), were used to test whether the fuzzy classificator would reach the same performance as with the development data (Table 4).
Patient population by histology and stages. Bold numbers represent the distribution in the whole patient population, standard print represents the patient population of the development period (AC adenocarcinoma, SCC Squamous cell carcinoma, SCLC small-cell lung carcinoma, NSCLC non-small-cell lung carcinoma).
Patients in the benign disease group (75 men and 56 women) were aged 15-96 years with a mean age of 54 years (SD 18.0), patients with confirmed primary lung carcinoma (238 men, 42 women) were aged 24-85 with a mean age of 64 years (SD 9.3). There were no differences in age between the histology groups (NSCLC: 64.4 �9.3 years; SCLC: 64.0�9.3 years; AC: 63.4�11.3 years; SCC: 65.3�7.9 years).
Tumour markers
Blood samples were processed within 60 min. Sera were kept frozen at -18�C, and within one week the analysis was carried out with the respective kit assays (Enzymun�-Test�) from Boehringer Mannheim (Germany) or Abott (Imx�-SCC) on the ES300 multianalyser or Hitachi 704 according to the manufacturer�s instructions. Error! Reference source not found. reports the cut-off values for persons with benign lung diseases.
Parameters analysed. Manufacturer�s cut-off values are understood at a specificity of 95% versus healthy persons, except for CYFRA21-1. (n.d.: a cut off for these parameters was not relevant, Ca cancer antigen, CEA carcinoembryonal antigen, CYFRA21-1 cytokeratin 19 fragments, NSE neuron-specific enolase)
Mathematical analysis
The panel of lung cancer-associated markers were evaluated by fuzzy logic modelling, first described by Zadeh in 1965. Applications of fuzzy logic to problems in diverse fields have been realized employing pattern- or rule-based approaches as described by Zimmermann (1991) and Bocklisch and Bitterlich (1994). In the case of our fuzzy classificator, the rule-based procedure is the method of choice. An important characteristic is the substitution of a graded (�fuzzy�) function for the sharp threshold value of a given �yes-no� decision: thus, the co-ordination of a tumour marker level to the criterion e.g. �malignant� would be described in terms of �more-less...� and not a sharp cut-off value. So-called membership functions of triangular shape are used in our model to describe the relation of each single marker to the term �malignant�, for example. For the mathematical definition of the rules, inference-algorithms of the MIN-MAX-type were used. The defuzzyfication employed the centre of gravity (COG) method to yield an output variable quantifying the distinctness of malignancy. The membership functions and the rules were defined with reference to the development data. The result is a multidimensional calculation in the form of specially adapted computer software. The system was confirmed by the test data. Evaluating the whole patient population with the fuzzy classificator gave similar results (see Table 4). By means of this fuzzy logic modelling the complex information of the tumour marker panel is processed to generate an indicator for the existence of a malignant process and further information about the histology of this malignancy.
The Chi-square test was used to assess the statistical significance of differences between observed ratios. Because the serum levels of the markers did not follow gaussian distribution, the significance of differences between the groups was calculated by means of a non-parametric test (Mann Whitney�s U-test). Values of P<0.01 were considered significant.
Results
Choice of parameters
The concept we employed is based on the measurement of a panel of markers related to lung cancer and its evaluation by fuzzy logic mathematical modelling. During a pilot phase the levels of several tumour markers were determined. The markers under investigation were either strongly associated with tumours of the lung (CYFRA21-1, CEA, NSE, SCC-Tm) or they reflected malignant processes in several organs including the lung (CA15-3, CA19-9, CA72‑4, CA125). In addition, ferritin, CRP and IgE were chosen because of their indicative value for tumour associated processes. The markers were evaluated with reference to their diagnostic value, redundancy of information and diagnostic gain by multivariate methods (data not shown). At the end of the pilot phase a set of markers was defined for used in the development period.
Performance of CEA, NSE, CYFRA21-1, CA-15-3, CA72-4, CA19-9 and SCC-Tm related to histology and tumour stage: percentages of elevated marker concentrations are given (cut-off values for two given specificities versus benign pulmonary disease group: upper part); numbers referred to in the text are on grey ground.
Table 3 shows sensitivity and specificity of the selected tumour markers according to histology and stage of the tumour. CYFRA21-1 has got the highest sensitivity for malignancy (49%). The most significant difference in sensitivity between NSCLC and SCLC is seen for NSE (NSCLC: 19.1%, SCLC: 86.7%; P<0.0001). This difference is diagnostically important. CA72-4 is less suiTable for this differentiation (P< 0.07), but adds to the information gained by NSE Whilsts the high sensitivity of NSE for SCLC has been described before, it is also interesting to note the low sensitivity of CA72-4 for SCLC. Ca15-3 marker levels are significantly more sensitive for aC than for SCC (AC: 37.3%, SCC: 10.3%; P<0.0002). The difference in sensitivity for CEA (SCC: 43.4%, AC: 64%), however, is not significant (P<0.12) and neither is the tumour marker SCC-Tm (SCC: 37.5%, AC: 20%, P<0.33).
fuzzy modelling is able to integrate the complex information contained in such a tumour marker profile into the classification system. This prompted us to evaluate a panel containing the tumour markers CYFRA21-1, CEA, NSE, CA15-3, and CA72-4 by the fuzzy logic approach. CA19-9 and SCC were left aside, since they did not add additional information; for this reason the two markers were not determined in every patient after the pilot phase. The value of the parameters ferritin, CRP and IgE was estimated by a multivariate analysis using data from all patients with main tumour markers (Cyfra-21-1, CEA, NSE) in the grey area (data not shown). It was found that CRP and IgE were nessesary to distinguish between SCC and AD, whereas ferritin gave additional information useful in the malignant-benign discrimination.
Performance of classification and analysis by histology groups
The fuzzy classificator formulated on the basis of the development data enabled correct classification of 91.9% of the malignant diseases and 88.2% of the benign diseases of the test data (Table 4; �test�).
Sensitivity and specificity of the fuzzy classificator for the classification malignant vs benign. Performance of the classification system on different databases.
A classification of all patients (�all� in Table 4) gives with 87.5% sensitivity and 85.5% specificity virtually the same results when the confidence intervals are considered. In data given below will refer to classification of the whole patient population if not stated otherwise because these data refer to a larger number of patients. Error! Reference source not found. lists sensitivities of the classification in relation to stage and histology and the specificities in relation to the subgroups of benign diseases. A sensitivity of more than 80% is seen regardless of histology, sensitivity for SCLC being the highest with 89.7%. No subtype is disadvantaged in the classification. Hence the �malignant-benign� classification can be applied regardless of the subtype of a given lung cancer.
Sensitivity and specificity of the fuzzy classificator for the classification malignant vs benign. The data are broken down into subtypes.
Analysis by ROC curves
The following Table 6 and the receiver-operating characteristics (ROC) show the advantage of using the fuzzy classificator rather than the best single tumour marker (CYFRA21-1) or a combination of CEA, CYFRA21-1 and NSE (Fig. 1 and Table 6). To obtain the ROC curve for the fuzzy classificator, for each data point at a given specificity a new classification procedure has to be done. Another approach was taken to generate the ROC curve evaluating the combination of CEA, NSE and CYFRA21-1: to generate one point on the ROC curve, the cut-off values of these three markers were varied to find the best sensitivity for the �OR�-combination at a given specificity. This gives the curve called �mathematical combination� which is different from a simple combination based on the manufacturer�s cut-off for healthy persons, as is often routinely used.
Figure 1 Receiver-operating characteristics (ROC) for CYFRA21-1, the mathematic combination of CYFRA21-1, NSE and CEA and the fuzzy classificator.
At a high specificity of over 90%, the best single marker, CYFRA21-1, shows a rather low sensitivity of around 60%. The combination of the three main markers results in an increase in sensitivity which can be further increased by the fuzzy classificator. Data points at lower specificity (eg- 70%) show a high sensitivity, and the improvement gained by combining the three markers rather than CYFRA21-1 alone narrows substantially. Some values from the ROC curve are presented in Table 6 to exemplify the differences. The diagnostic performance is compared either at a given specificity or at a given sensitivity (Table 6, comparable values on grey background).
For comparison, the specificity of 85.5% is chosen, since it is comparable to the overall specificity of the combination of the three markers CEA, CYFRA21-1 and NSE with a cut-off of 95% each (0.95)�=0.857. At this specificity, the fuzzy classificator has a sensitivity of 87.5%, which exceeds that of the best performing single marker CYFRA21-1 by nearly 20% (Table 6, upper part). CEA, CYFRA21-1 and NSE together reach 78.8% being nearly 10% less sensitive than the fuzzy classificator over the combination.
Individual values from the ROC curves for the comparison of sensitivity and specificity. The grey background indicates the values used as basis for comparison (same specificity or same sensitivity, respectively).
If 90% of malignant diseases are correctly classified (sensitivity), the specificity against benign lung diseases is 15% higher for the fuzzy classificator than for the optimal mathematical combination and over 20% higher than for CYFRA21-1 alone (Table 6, middle part).
At a chosen specificity of around 95%, ROC curves generally approach each other because of their asymptotic character. With 75.4% sensitivity the fuzzy classificator is 25% better than the best single marker (Table 6, lower part). Compared to 70.6% sensitivity of the mathematical combination of three markers, there is still a gain of 5%. In case of the mathematical combination it should be mentioned that the specificity for each of the single markers is 98-99%, giving 95% specificity alltogether. However, if compared at the same sensitivity of 75%, which is reached by the fuzzy classificator at 95% specificity, the other methods do not get as far as 95% specificity (Table 6, lower part).
Analysis by stages
In an analysis broken down into stages it becomes clear that the gain in sensitivity with use of the fuzzy classificator is due mainly to a substantial increase in the classification of patients in stage I-IIIa (Fig. 2). For comparison the cut-offs for the single marker and the combination have been chosen to meet the specificity of the fuzzy classificator of 86%. CYFRA21-1 (cut-off 3.4 ng/ml; see also Table 3) alone detects only 30% of malignant diseases in stages I and II correctly. In higher stages its sensitivity rises above 60% (Fig. 2). The advantage of the mathematic combination over the best single marker is sometimes small or nonexistent (Fig. 2: stage II, IIIb, IV, LD and ED). Only in stages I and IIIa is a substantial gain in sensitivity observed when the combination is used.
Figure 2 Comparison of the performance of the best tumour marker (NSE or CYFRA21-1) and of the combination of CEA, CYFRA21-1 and NSE with the fuzzy classificator grouped by stages. For the best single marker (CYFRA21-1 for NSCLC and NSE for SCLC), the 86%-Cut-off value is given in Table 3. For the combination, the 95%-Cut-Offs are taken; the combined specificity CEA, CYFRA21-1 and NSE is 86.3% versus benign diseases. The specificity of the fuzzy classificator is 85.5% (see figure 1 and Table 5).
The fuzzy mclassificator evaluates five tumour markers, which results in a large increase of sensitivity especially at low stages (stageI: 58.6%, stage II: 53.3%, stage IIIa: 86.0%).For example, in stage II the increase is 20%. The difference from CYFRA21-1 is still over 20% in stage IIIa, and 10% in IIIb. In stage LD the fuzzy classificator correctly detects 83.3% of cases, over 15% more than NSE. Generally, since NSE is a good marker for small cell lung carcinoma, the combination of three markers is not better than NSE alone. At the cut-off for 86% specificity (14ng/ml), NSE even detects all cases with extensive disease. The combination of three markers does not find one malignant case with NSE concentrations between 14ng/ml (86%-cut-off) and 18ng/ml (95%-cut-off), because each single marker in the combination has the 95%-cut-off. The fuzzy-classificator classifies 94% of EDs correctly. The �missed cases� are those with rather low NSE concentrations.
Histological subclassification
Histological subclassification of malignant tumours by the fuzzy classificator is correct in over 90% of SCLCs and NSCLCs.
Table Error! Bookmark not defined.. Subclassification of malignant tumours into SCLC and NSCLC by the fuzzy classificator as percentage of correctly detected cases. For comparison, the values for NSE (SCLC) in the same patient group are given. The cut-off for NSE in this case is c = 18 ng/ml (also see Table 3). The grey background indicate to be used as basis for comparison.
Error! Reference source not found. shows the performance of the fuzzy classificator in the discrimination between SCLC and NSCLC; the result is compared with that yielded by NSE, which is considered the most sensitive for SCLC. If the cut-off is set for NSE to reach the same sensitivity for SCLCs, NSCLCs are correctly detected in 78% of cases.. Moreover, a subclassification of NSCLC into squamous carcinoma (SCC) and adenocarcinoma (AC) is possible with 76.8% and 78.8% sensitivity respectively (Table 8).
Subclassification of NSCLCs into SCC and AC by the fuzzy classificator as percentages of correctly classified cases. For comparison, the values for CYFRA21-1 and CEA in the same patient group are given. Cut-off values are 4.4 ng/ml for CEA and 3.6 ng/ml for CYFRA21-1. The grey background indicates the values have to be used as basis for comparison.
For comparison, CEA and CYFRA21-1 as single markers have been chosen and cut-off values set so that sensitivities are comparable. In this case and compared with CEA, the fuzzy classificator shows more than 30% improvement when it comes to detecting SCC at comparable sensitivity to that for AC (Table 8). High CYFRA21-1 concentrations are more indicative of SCC, but at a sensitivity of 76.5% for SCC its sensitivity for AC is only 26.4%. The tumour marker SCC-Tm is considered specific for SCC tumours. In our study it was only evaluated during the pilot phase, among 42 patients with AC or SCC of the lung it reached 50.0% sensitivity for SCC and 68.6% sensitivity for AC at the manufacturer�s cut-off (c = 1.5 ng/ml, data not shown). We found that SCC-Tm was not as good as CEA for this differentiation. It was therefore not included in our parameter profile.
Classification results in the �grey area�
The grey area describes tumour marker levels close to the cut-off. In this study the grey area ranges form 0.9 times to twice the 95%cut-off value (as given by the manufacturer). If for one of the three main markers CEA, CYFRA21-1 or NSE the concentration was within these limits and for none of them above the double cut-off value, the patient concerned was grouped in this grey area. In fact, 35% of the patients met these criteria. fuzzy logic modelling is very efficient in classification of these cases, because it evaluates the significance of such small differences and their coincidence. In Table 9 the classification results are listed, comparing the fuzzy classificator, the combination of NSE/CEA/CYFRA21-1, and CEA or CYFRA21-1 alone. On the basis of 87.5% sensitivity and 85.5% specificity for all patients, the fuzzy classificator detects 75.6% of malignant and 73.1% of benign diseases in the grey area correctly (Table 9).
Correctly classified cases by different classification methods in the grey area (range 0.9 x - 2 x 95% Cut-off as specified by the manufacturer for CEA, NSE or CYFRA21-1). Patient population in this area: 90 with malignant and 53 with benign diseases.
There are different possible approaches comparing these figures with the usual results recorded in the grey area. The specificity either of single markers or of the combination (CEA, CYFRA21-1, NSE) is set to give the same rate of correct classification of benign diseases, which means determining new cut-off values especially for the grey area (Table 9, upper part). In this case, the fuzzy classificator is 20% more sensitive than CEA or the combination, and 30% more sensitive than CYFRA21-1 alone for the detection of malignant cases. Since the definition of this grey area is itself based on normal cut-off values, however, a comparison with the results reached on the basis of the manufacturer�s cut-off values makes more sense (Table 9, middle part). Here, CEA, CYFRA21-1 and NSE together (combination) give a higher sensitivity (85% of malignant diseases) than the fuzzy classificator but with an unacceptably low specificity of only 21% correctly classified benign diseases. CYFRA21-1 keeps a good balance between detection of malignant (46.7%) and benign diseases (67%) but this is still much lower than for the fuzzy classificator. CEA, however, shows a specificity of more than 80%, but it detects only every third patient with lung cancer correctly.
Finally, the use of the 95% cut-off values for benign lung diseases can be compared with use of the fuzzy classificator (Table 9, lower part). Sensitivity of the combination is 20% lower than of the fuzzy classificator at the same specificity. For CYFRA21-1 and CEA the detection rate for benign diseases is higher (88.5% and 90.4%), but for malignant diseases it is clearly much too low (18.9% and 27.8%). These results reflect the common experience in the grey area: elevated marker levels that are just barely elevated are very difficult to interpret. Here the result of around 75% sensitivity and specificity for the fuzzy classificator is a very useful advantage.
Analysis of the malignancy indicator
The output variable of the fuzzy classificator termed �malignancy indicator� can take any value between 0 (no signs of malignancy) and 1 (definite malignancy). Any value below 0.5 is considered indicative of a benign process. Around 0.5 there are weak signs of a malignant process. If the malignancy indicator approaches 1.0, existence of a malignant tumour is increasingly certain, being 100% certain at the value 1.0. This output variable gives a quantitative result for the judgement of a suspected malignancy.
Figure 3 Value of the �malignancy indicator� for patients with (M1) and without (M0) metastases compared with CYFRA21-1. The differences are significant (Mann-Whitney�s U-test).
When tumour patients with and without metastases are compared by means of a box plot (Fig. 3) a correlation becomes evident between the value of the �malignancy indicator� and the presence of metastases. CYFRA21-1 concentrations also correlate with the presence of metastases (Fig. 3), but there is a larger overlap between the two groups. The fuzzy classificator is therefore more informative in this context.
Discussion
In the unique mathematical approach to a diagnostic problem presented here, the information from a panel of tumour markers and other parameters is processed by fuzzy logic modelling for the evaluation of a given lung disease. The fuzzy classificator provides information concerning the presence of a malignant bronchial tumour, its histological subclassification into SCLC or NSCLC and the differentiation into AC and SCC (Fig. 4).
For the discrimination of malignant from benign disease the fuzzy classificator performs over 15% better in sensitivity than CYFRA21-1 (the best single marker) and surpasses the sensitivity of the combination of CEA, CYFRA21-1 and NSE by nearly 10%. The sensitivities reported for single tumour markers in the literature are concordant with our results [Stieber et al. 1997, Ebert et al. 1997 see also references therein]. There are few studies in which evaluation of a combination of several markers has been attempted. A fuzzy logic approach to this or a similar classification problem has not so far been attempted, however.
Figure 4 Classification steps with the fuzzy classificator. Sensitivities are given for the appropriate branch (i.e. correctly classified cases).
Combinations of two markers, such as NSE and CYFRA21-1, increased the sensitivity above that of either of the single markers [Giovanella et al., 1997: 80% SCLC, 78% NSCLC] or did not make a difference [Stieber et al. 1993a], depending on the study. This held also true when SCLC and NSCLC were considered separately [Stieber et al. 1993a]. In a different study, the sensitivity of CYFRA21-1 alone could not be markedly increased by any combination with another lung cancer specific marker [Plebani et al. 1995].
In a recent study Stieber et al. (1997) tested many possible combinations of three markers: CEA, CYFRA21-1, NSE. Only the combination of CYFRA21-1 and CEA led to a small increase in sensitivity (to 47%) for all histological types. Other authors find that a combination of several markers can increase sensitivity, in some cases up to 90% [Muraki et al. 1996; Plebani et al. 1995]. But fairly often it passes unnoticed that because of the multiplication of the specificities this increase in sensitivity is achieved at the expense of the overall specificity, which falls to 70% [Muraki et al. 1996] or even 41% [Plebani et al. 1995]. This of course counterbalances the gain in sensitivity, and thorough analysis on the basis of ROC curves could not reveal any net advantages. To compare the fuzzy approach with these alternatives on the basis of our data we evaluated a mathematic combination of the three markers CEA, CYFRA21-1 and NSE: For a given specificity the respective cut-off values were varied by trial and error until the combination of the three cut-off values with the best sensitivity was determined. Such an approach is unlikely to take place in everyday routine, but represents the best result which could be reached without employing advanced mathematical methods. This method gave better results than the use of single markers, but the fuzzy approach was still superior. The ROC curve (Fig. 1) underscores the superiority of this system over the simple combination of three markers and CYFRA21-1 at any chosen specificity. (The area under the curve is largest for the fuzzy classificator.)
One of the main drawbacks of serum tumour markers is that the highest levels are usually found when the disease is at an advanced stage (TNM IIIb or IV). The fuzzy classificator substantially improves the sensitivity of tumour markers at stages I-IIIa. It correlates better with the presence of metastases (Fig. 3) than CYFRA21-1 or CEA. It also parallels the tumour mass (data not shown) but has proved to be more sensitive than single tumour markers notably at early stages where tumour mass is small and thus the fuzzy classificator gives a hint of the severity of the malignant process. These characteristics may prove useful in monitoring and prognosis. A fuzzy logic approach to the evaluation of prognostic factors is currently under way.
For the differentiation between NSCLC and SCLC only some single tumour markers can be employed. NSE is most sensitive for SCLC among all histological subtypes (Stieber et al., 1997: 45%; Spinazzi et al., 1994: 88%) and CYFRA21-1 is considered the best overall marker for NSCLC [Molina et al. 1994, Stieber et al. 1997, Paone et al. 1997, Spinazzi et al. 1997]. Its discriminating power for SCLC and NSCLC is widely accepted and more or less obvious in different studies, whereas CYFRA21-1 does not discriminate well between the two [Molina et al. 1994, Huang et al. 1997]. Basically, a limit for the evaluation of the histology classification analysis is set by the common cytological and histology methods, which are bound to an error rate.
More recently Paone et al. (1997) used a discrimination analysis as a method to optimize the combination of CYFRA21-1 and NSE for the differentiation of SCLC and NSCLC. They report correct classification of 88.5% of SCLC and 94.3% of NSCLC in a patient population with 261 malignant diseases, but no benign diseases were included. Hence, these values have to be compared with our sensitivity of 91% for the correct classification of SCLC and of 91% for NSCLC among malignant diseases. We tested this type of analysis with our data and found sensitivities slightly lower than with the fuzzy method: 86.7% for SCLC and 86.3% for NSCLC. This is not surprising, since the fuzzy approach takes into account additional information from other parameters in an advanced mathematical manner. The diagnostic use of the discrimination analysis, however, is limited exclusively to cases of malignancy since the analysis of Paone et al. has not been evaluated for a benign disease group: it therefore does not include an a priori statement on the malignancy of a given lung disease. The fuzzy system, however, also provides information about the malignancy.
Further subclassification into SCC and AC can be done with 76.9% and 78.8% sensitivity respectively by the fuzzy classificator. Tumour markers are normally not employed for this differentiation. CYFRA21-1, CEA and SCC-Tm are considered only theoretically, if at all. Several authors have found, that single markers cannot be used for this purpose [Molina et al. 1994, Huang et al. 1997, Margolis et al. 1993 see also references therein]. In our study, CEA, CYFRA21-1 and SCC-Tm alone were not very useful for this decision. CA15-3, CRP and IgE contributed helpful information in this context (see Table 3) even though its overall sensitivity for bronchial carcinoma is low. Nevertheless, the fuzzy classificator takes even small differences into account and gives the right classification in 76.8 % and 78.8% of cases, respectively. With this sensitivity it is as good as classic cytological and histological techniques: The success of forceps biopsy, bronchial washing and brushing is 70-80% for SCC and AC with central lesions; sensitivities not reached for small (< 2 cm) or peripheral lesions [Arroliga and Matthay 1993, see also references therein].
The differentiation between lung cancer and other malignant histologies, such as mesothelioma, sarcoma, lymphoma and embryonal carcinoma, was not aim of this study. The number of cases with these more rare histologies was too low for a check of the classification system; e.g. three cases of mesothelioma were correctly classified as �malignant�. A study with large groups of such diseases or patients with pulmonary metastases of other tumours would be interesting.
In the grey area (0.9x - 2.0x cut-off) 40% of benign diseases and 25% of malignant diseases show barely elevated marker levels, and therefore these cases are equivocal in customary tumour marker analysis. The fuzzy logic modelling is notably able to integrate such unsharp settings making them interpreTable. It classifies 76% of malignant diseases and 73% of benign diseases correctly, whereas the reliability of single markers or simple combination of tumour markers is very low in this area (see Table 9).
Conclusions
The use of tumour markers in the diagnosis of lung cancer is very limited at the present. Owing to their rather low positivity detection rates, they are clearly inferior to histological and cytopathological procedures. The new approach presented here enhances the diagnostic performance of tumour markers substantially up to 89% sensitivity and may therefore justify their use in more cases in the process of differential diagnosis when lung cancer is suspected. The primacy of tumour morphology in diagnosis will not be put into question by the fuzzy classificator, but this new approach may prove a useful contribution particularly if morphological characterization cannot be done. It should also be helpful in advanced stages, in which patients are very often not eligible for extensive investigations owing to their physical condition. Complications, which occur in 10% of cases overall [Arroliga et al. 1993], could be prevented in this way. Besides being a noninvasive and uncomplicated analytical method, the fuzzy classification system provides the results quickly.
Acknowledgements:
We wish to thank H. Kleindienst, E. Richter, M. Fritsche and H. Wenzel for excellent technical assistance. We are grateful to W. Rotzsch for fruitful discussions. This work was supported in part by a grant # TOU 250 from the BMBF.
Abbreviations:
SCC Squamous cell carcinoma
AC Adenocarcinoma
SCLC Small cell lung carcinoma
NSCLC Non-small cell lung carcinoma
BCa Bronchial Carcinoma
CEA Carcino-embryonal Antigen
CA15-3 Cancer Antigen 15-3
CA72-4 Cancer Antigen 72-4
CYFRA21-1 Cytokeratin 19 fragments
NSE Neuronspecific enolase
SCC-Tm the tumour marker named �SCC�
REFERENCES:
Arroliga AC, Matthay RA (1993) The role of bronchoscopy in lung cancer. Lung Cancer, 14: 87-98
Bocklisch SF, Bitterlich N (1994) Fuzzy pattern classification - methodology and application. In: Kruse R, Gebhardt J, Palm R (eds.) Fuzzy systems in computer science, Vieweg, Wiesbaden.
Ebert W, Muley T, Drings P (1996) Does the assessment of serummarkers in patients with lung cancer aid in the clinical decision making process? Anticancer Res, 16: 2161-2168
Ebert W, Hoppe M, Muley T, Drings P (1997) Monitoring of therapy in inoperable lung cancer patients by measurement of CYFRA 21-1, TPA- TP, CEA, and NSE. Anticancer Res,17: 2875-2878
Everone CA (ed.), Dec 1997, Lung Cancer - Base Report;; available at URL http://www.fis.org/lungc.html
Giovanella L, Piantanida R, Ceriani L, Bandera M, Novario R, Bianchi L, Roncari G (1997) Immunoassay of neuron-specific enolase (NSE) and serum fragments of cytokeratin 19 (CYFRA21-1) as tumor markers in small cell lung carcinoma: clinical evaluation and biological hypothesis, Int J Biol Markers, 12: 22-26
Hemdon RC, Lancaster JL, Toga AW, Fox PT (1996) Quantification of white and gray matter volumes from T1 parametric images using fuzzy classifiers, J Magn Reson Imaging 6: 425-435
Hermanek P, Sobin LH (eds) (1992) International Union Against Cancer (UICC) TNM classification of malignant tumours, 4th ed. (1987,revised 1992) Springer Berlin Heidelberg NewYork.
Huang MS, Jong SB, Tsai MS, Lin MS, Chong IW, Lin HC, Hwang JJ (1997) Comparison of cytokeratin fragment 19 (CYFRA21-1), tissue polypeptide antigen (TPA) and carcinoembryonic antigen (CEA) as tumour markers in bronchogenic carcinoma. Respir Med 91: 135-142
Kalmanson D, Stegall F (1975) Cardiovascular investigation and fuzzy concepts. Am J Cardiol 35: 30-34
Margolis ML, Hyzy JB, Schenken LL, Schepart BS (1993) Serum tumor markers in non-small cell lung cancer. Cancer 73: 605-609
Molina R, Agusti C, Mane JM, Filella X, Jo J, Joseph J, Gimenez N, Estape J, Ballesta AM (1994) CYFRA21-1 in lung cancer: comparision with CEA, CEA 125, SCC and NSE serum levels, Int J Biol Markers 9: 96-101
Muraki M, Tohda Y, Iwanaga T, Uejima H, Nagasaka Y, Nakayima S (1996) Assessment of serum CYFRA21-1 in lung cancer. Cancer 77: 1274-1277
Nishimura H, Kambe M, Futagami K, Morishita K, Tsubokura T (1991) Fuzzy realization in clinical test database system. Int J Biomed Comput 28: 289-296
Paone G, De Angelis G, Portalone L, Greco S, Giosu� S, Taglienti A, Bisetti A, Ameglio F, (1997) Validation of an algorithm able to differentiate small-cell lung cancer (SCLC) from non-small-cell lung cancer (NSCLC) patients by means of a tumour marker panel: analysis of the errors, Br J Cancer 75: 448-450
Pathak A, Pal SK (1986) Fuzzy grammars in syntactic recognition of skeletal maturity from X-rays. IEEE Trans Systems Man Cybern 16: 657-667
Plebani M, Basso C, Navaglia F, De Paoli M, Tommasini A, Cipriani A (1995) Clinical evaluation of seven tumour markders in lung cancer diagnosis: can any combination improve the results? Br J Cancer 72: 170-173
Rau G, Becker K, Kaufmann R, Zimmermann HJ (1995) Fuzzy logic and control: principal approach and potential applications in medicine. Artif Organs 19: 105-112
Sanchez E (1977) Solutions in composite fuzzy relation equations. Application to medical diagnosis in Brouwerian logic. In: Gupta MM, Saridis GN, Gaines BR (eds) Fuzzy automata and decision processes. North Holland, New York, pp 221-234
Sanchez E (1998) Fuzzy-logic and inflammatory protein variations. Clin Chim Acta 270: 31-42
Spinazzi A, Soresi E, Boffi R, Nonis A, Noseda A, Cobelli S, Scanni A (1994) Evaluation of neuron-specific enolase, tissue polypeptide antigen as markers for staging and monitoring response to therapy of lung cancer. Cancer Detect Prev 18: 209-220
Stieber P, Dienemann H, Hasholzner U, M�ller C, Poley S, Hofmann K, Fateh-Moghadam A (1993a) Comparison of cytokeratin fragment 19 (CYFRA21-1), tissue polypeptide antigen (TPA) and tissue polypeptide specific antigen (TPS) as tumour markers in lung cancer, Eur J Clin Chem Clin Biochem 31: 689-694.
Stieber P, Bodenm�ller H, Banauch D, Hasholzner U, Dessauer A, Ofenloch-H�hnle B, Jaworek D, Fateh-Moghadam A (1993b) Cytokeratin 19 Fragments: a new marker for non-small-cell lung cancer. Clin Biochem 26: 301-304
Stieber P, Dienemann H, Schmitt UM, Reinmiedl J, Hasholzner U, Reiter W, Hofmann K, Schalhorn A, Yamaguchi K (1997) Pro-Gastrin-Releasing Peptide (ProGRP) - a new marker in small-cell lung carcinomas. J Lab Med 21: 336-344
Yang H-B, Hsu P-I, Lee J-C, Chan C-H, Lin X-Z, Chow N-H (1995) Adenoma-carcinoma sequence: a reappraisal with immunohistochemical expression of ferritin. J Surg Oncol 60: 35-40
Zadeh LA (1965) Fuzzy sets. Information and Control 8: 338-353
Zimmermann HJ (1991) Fuzzy set theory and its application 2nd edn. Kluwer, Boston.