Our reference: ESWA 7195

P-authorquery-v10

AUTHOR QUERY FORM

Journal: ESWA Please e-mail or fax your responses and any corrections to: E-mail: corrections.esch@elsevier.sps.co.in Article Number: 7195 Fax: +31 2048 52799

Dear Author, Please check your proof carefully and mark all corrections at the appropriate place in the proof (e.g., by using on-screen annotation in the PDF le) or compile them in a separate list. Note: if you opt to annotate the le with software other than Adobe Reader then please also highlight the appropriate place in the PDF le. To ensure fast publication of your paper please return your corrections within 48 hours. For correction or revision of any artwork, please consult http://www.elsevier.com/artworkinstructions. Any queries or remarks that have arisen during the processing of your manuscript are listed below and highlighted by ags in the proof. Click on the Q link to go to the location in the proof. Location in article Query / Remark: click on the Q link to go Please insert your reply or correction at the corresponding line in the proof

Q1 Q2

Please confirm that given names and surnames have been identified correctly. Please provide significance of bold italic values in Tables 3, 5, 6, 7, and 9. Kindly check.

Thank you for your assistance.

ESWA 7195 5 December 2011

Expert Systems with Applications xxx (2011) xxxxxx 1

No. of Pages 9, Model 5G

Contents lists available at SciVerse ScienceDirect

Expert Systems with Applications

journal homepage: www.elsevier.com/locate/eswa

2 3 4 Q1 5 6 7

An algorithm designed for improving diagnostic efciency by setting multi-cutoff values of multiple tumor markers
Qiang Su a, Jinghua Shi a, Ping Gu b, Gang Huang b,, Yan Zhu c
a

Department of Industrial Engineering and Logistics Management, Shanghai Jiao Tong University, Shanghai 200240, China Department of Nuclear Medicine, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai 200127, China c School of Economics and Management, Tsinghua University, Beijing 100084, China
b

8 1 0 2 0 11 12 13 14 15 16 17 18 19

a r t i c l e

i n f o

a b s t r a c t
Currently, tumor markers have been effectively applied for colorectal cancer (CRC) diagnosis. In order to decrease the information loss caused by single cutoff value and improve diagnosis efciency (DE), we explore the integrative application of multiple tumor markers with multiple cutoff values systematically by developing an optimization algorithm named MVMTM. The effectiveness of the MVMTM is experimentally studied based on a real medical dataset. With MVMTM, the united use of three tumor markers can enhance DE from 0.78 to 0.86. Furthermore, MVMTM has been proved to be better than other baseline machine learning algorithms signicantly. 2011 Published by Elsevier Ltd.
21 22 23 24 25 26 27 28 29

Keywords: Colorectal cancer (CRC) Diagnostic efciency (DE) Tumor markers Cutoff values Rough set theory (RST) Genetic algorithm (GA)

30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55

1. Introduction Colorectal cancer (CRC) is one of the most common cancers whose mortality is ranked third in the world. Colonoscopy and fecal occult blood test (FOBT) are the frequently used screening measures for CRC. Although large-scale clinical trials demonstrate that colonoscopy is effective for the diagnosis of CRC, the comparatively high cost and the painful process seriously limit its application (Lewis, 2000). In terms of FOBT, due to the low diagnostic accuracy, it is not an effective method for CRC diagnosis. In this circumstance, a number of tumor markers have been set forth and utilized in CRC diagnosis. For example, in Renji Hospital, a teaching hospital in Shanghai, China, nine different tumor markers, viz. AFP, CEA, CA 19-9, CA 125, CA 153, CA 50, CA 724, CA 211 and CA 242 can be employed for cancer diagnosis. Among them, the most commonly used tumor markers for colorectal cancer diagnosis are CEA, CA 19-9 and CA 50. Usually, single cutoff value is applied to separate the value range of the tumor marker into two segments as normal and abnormal. Some studies demonstrate that diagnostic result can varied widely along with the adjustment of the cutoff value (Krner, Sreide, Stokkeland, & Sreide, 2007). Up to now, many efforts have been dedicated to searching for an appropriate cutoff value so as to achieve the highest diagnosis efciency (DE) (Armitage, Davidson, Tsikos, & Wood, 1984; Carriquiry & Pineyro, 1999; Wichmann, Lau-Werner, Muller, Stierber, & Schildberg, 2000).

Corresponding author.
E-mail address: cherry_shi44@163.com (G. Huang). 0957-4174/$ - see front matter 2011 Published by Elsevier Ltd. doi:10.1016/j.eswa.2011.11.089

The commonly used approach is to numerate the possible cutoff values and choose the value at which the highest DE can be achieved (Wan & Zhang, 2007; Weiss, Niwas, Grizzle, & Piyathilake, Space inconsistent 2004). For instance, Krner et al. suggested that the optimal cutoff value for CEA was 4 lg/L (Krner et al., 2007). Carpelan-Holmstrom check global et al. suggested a cutoff value of 5 lg/L for CEA and 20 U/ml for CA 242 (Carpelan-Holmstrom, Haglund, Kuusela, Jarvinen, & Roberts, 1995). Some researchers found that the combination of different tumor markers can improve DE. Lucarotti et al. stated that the combination use of CEA, CA 19-9, and CA 50 could improve DE signicantly in differentiating benign tumor from malignant tumor for the pancreatic cancer (Lucarotti et al., 1991). Moghimi and Ghodosi gured out that the combination use of CEA and CA19-9 could achieve a higher DE compared with individual usage (Moghimi & Ghodosi, 2007). These ndings imply that, instead of being used individually, the different markers should be applied simultaneously and the check results should be considered synthetically. In almost all existing studies, only a single cutoff value is set for each tumor marker (Duffy, 2001; Moertel et al., 1993; Persijn & Hart, 1981; Wood, Ratcliffe, Burt, Malcolm, & Blumgart, 1980). In this way, some important information obtained from the diagnosis test will be neglected. Taking CEA as an example, its value can range from 0 to 550 lg/L. If its cutoff value is set to 4 lg/L, the test results of 5 lg/L and 500 lg/L will be simply judged as the same (abnormal) although they are very different from each other. From this point of view, it is too rough to simply separate the broad value range of a tumor marker into two segments of normal and abnormal. Hence, multi-cutoff values should be employed to take more advantages of the test result so as to improve DE.

56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85

Please cite this article in press as: Su, Q., et al. An algorithm designed for improving diagnostic efciency by setting multi-cutoff values of multiple tumor markers. Expert Systems with Applications (2011), doi:10.1016/j.eswa.2011.11.089

ESWA 7195 5 December 2011

2 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 Q. Su et al. / Expert Systems with Applications xxx (2011) xxxxxx

No. of Pages 9, Model 5G

Under this consideration, aiming at improving DE for CRC, a novel diagnosis strategy with multiple tumor markers and multi-cutoff values are studied systematically. An optimization algorithm is designed for setting multi-cutoff values for multiple tumor markers (MVMTM). In this work, three tumor markers, i.e., CEA, CA 19-9, and CA 50, are used simultaneously. And no more than three cutoff values are permitted for each tumor marker. With the algorithm, the optimal cutoff values are calculated out based on a real diagnosis dataset with 124 cases. Furthermore, other 88 cases are employed to validate the effectiveness of the algorithm. This paper is organized as follows. In Section 2, the evaluation method of DE is addressed and the algorithm related technologies including the rough set theory (RST) and the genetic algorithm (GA) are introduced. The details of the MVMTM are elaborated in Section 3. Then, in Section 4, the experimental study is conducted to demonstrate the effectiveness of the algorithm. Furthermore, some discussions are given in Section 5. Section 6 concludes the paper.

(Yu & Xu, 1998), etc. Due to these uncertainties, many researches assign the same weights to Sp and Se so that the point located at the upper left-hand corner in Fig. 1 is usually chosen as the cutoff values (Streiner & Cairney, 2007). Then, DE in Eq. (1), dened as the average of Sp and Se, is employed as the criterion for choosing cutoff values.

130 131 132 133 134 135

136 138

Se Sp DE 2
2.2. Rough set theory (RST)

139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171

104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129

2. Related technologies 2.1. Measurement of diagnostic efciency (DE) for cutoff values Suppose a set of sample cases are tested, the diagnosis result for each case can be either normal or abnormal. However, sometimes, wrong decisions can be made in which the normal case is judged as abnormal or vice versa. To measure these errors, sensitivity (Se) and specicity (Sp) are proposed. Sensitivity (Se) represents the probability that an abnormal case is correctly judged as abnormal. Specicity (Sp) stands for the probability that a normal case is correctly diagnosed as normal. Suppose there are seven possible cutoff values, namely, C1, C2, C3, C4, C5, C6, and C7. For each cutoff value, the corresponding sensitivity (Se) and specicity (Sp) can be computed. Then, the coordinate point (1-Sp, Se) can be drawn on the coordinate plane. As shown in Fig. 1, when all the seven coordinate points are obtained, a curve entitled the receiver operating characteristic (ROC) (Wan & Zhang, 2007) can be derived. This curve can be used to assist clinicians in choosing the appropriate cutoff value according to the required levels of Sp and Se. To take a special example, in Fig. 1, C1 would be chosen to be the cutoff value if clinicians prefer high Sp while do not care about Se. In other words, actually, cutoff value is chosen based on the preferred levels of Sp and Se. Similarly, in our research, multi-cutoff values should also be determined according to the requirements of Sp and Se. However, the requirements of Sp and Se are heavily related to the characteristics of the disease, diagnosis cost, and disease prevalence in the given area

The RST is a fairly new methodology developed for extracting useful knowledge from imprecise, uncertain, and vague information. Since being put forward by Pawlak in 1982, the RST has been successfully applied in many elds including machine learning, data mining, and intelligent data analyzing, etc. RST is suitable for medical data analysis since it does not need any preliminary or additional requirements for the data (An et al., 2007). Wakulicz-Deja and Paszek tried to apply RST in diagnosis of Mitochondrial Encephalomyopathies (MEM) (WakuliczDeja & Peszek, 2003). With the capabilities of attribute reduction and decision rule generation, RST can improve the classication quality and diagnosis efciency remarkably. In order to generate decision rules in medical diagnosis, Ilczuk, Mlynarski, WakuliczDeja, Drzewiecka, and Kargul (2005) constructed an algorithm entitled LEM2 based on RST. Fakih and Das (2006) stated that the RST was useful in extracting diagnostic information in the form of rules from the medical databases. In one word, RST offers a useful mechanism for analyzing and distilling essential attributes and rules from medical data (Pattaraintakorn and Cercone, 2008). These research results imply that RST can provide signicant assistance for diagnosis problem encountered in this paper. In RST, an information system can be described as S hU; A; V; f i. In which, U is a nite set of objects fx1 ; x2 ; . . . ; xn g; A is a nite set of attributes fa1 ; a2 ; . . . ; am g; V is the value range of each attribute fv a1 ; v a2 . . . v am g; f : U A ! V is called an information function, and f x; a 2 v a , for 8a 2 A; 8x 2 U. Generally, diagnosis can be regarded as a decision process including two types of attributes, i.e., condition attributes C and decision attributes D. Thus, the corresponding information system can be modeled as S hU; C [ D; V; f i. Let P # C, xi, xj 2 U. We say that xi and xj are indiscernible if the following function is satised.

INDP fxi ; xj 2 U Uj8a 2 P; f xi ; a f xj ; ag

172 174
175 176 177 178 179 180 181 182 183 184 185

C5 C4 C3

C6

C7

C2

Indiscernibility is an imperative property in medical diagnosis. Due to the complexity of disease, it is likely that some persons with similar test results have totally different diagnosis results. These cases are called the indiscernible cases and can be properly resolved by RST. Using RST, a set of decision rules can be derived referring to the information system S hU; C [ D; V; f i. And these rules can be expressed as c1 xi ; . . . ; cm xi ! d1 xi ; . . . ; dn xi or C x ! Dx where fc1 ; . . . cm g C and fd1 ; . . . dn g D. To evaluate the effectiveness of a decision rule, three criteria, i.e., support, accuracy, and coverage, are dened as follows.

186 188 189

cardC x \ Dx Supx C; D cardU

C1

cardC x \ Dx Accx C; D cardC x cardC x \ Dx Cov x C; D cardDx

191 192

Fig. 1. ROC curve.

194

Please cite this article in press as: Su, Q., et al. An algorithm designed for improving diagnostic efciency by setting multi-cutoff values of multiple tumor markers. Expert Systems with Applications (2011), doi:10.1016/j.eswa.2011.11.089

ESWA 7195 5 December 2011

Q. Su et al. / Expert Systems with Applications xxx (2011) xxxxxx 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227

No. of Pages 9, Model 5G

3 254 255 256 257 258 259 261 262 263 264 265 266 267 269 268 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284

2.3. Genetic algorithm (GA) Since multi-cutoff values are permitted and multiple tumor markers are applied simultaneously, it would be extremely arduous and even impossible to explore all the feasible combinations in nding the optimal solution. In this circumstance, a kind of heuristic global searching algorithm called genetic algorithm (GA) is employed. The basic principle of GA can be explained as follows. A feasible solution can be modeled as a chromosome which consists of a series of genes. GA algorithm maintains a group of chromosomes, population, to explore the solution space. The stochastic global searching functionality is realized by genetic operations of selection, crossover, and mutation. With these operations, a new set of population can be generated based on the previous population. To ensure the evolutionary progress between populations, an evaluation function entitled tness function is dened. This function can determine the chance of survival for each chromosome so as to guarantee that the tter individual has higher chance to pass its outstanding genes to offspring. Due to its conceptual simplicity, the ease of programming, and small number of parameters, GA is becoming one of the most popular evolutionary algorithms (Modchang, Triampo, & Lenbury, 2008). Lollini et al. applied the genetic algorithm to nd out the possible minimal vaccination schedule without enormous number of biological experiments (Lollini, Motta, & Pappalardo, 2006). Castiglione et al. took advantage of GA to nd an optimal therapeutic schedule that maximized immune restoration, minimized the viral count and minimized the dose of drug administered to the simulated patient (Castiglione, Pappalardo, Bernaschi, & Motta, 2007). Smigrodzki et al. reported the development of a modied GA suited for detection of biologically meaningful patterns of mitochondrial mutations (Smigrodzki et al., 2005). These researches demonstrate that, with strong self-adaptive global search ability, GA can be used to solve the large scale combination optimization problem.

According to RST theory, the diagnosis problem in this research can be modeled as inferring the nal diagnosis result, normal (d1) or abnormal (d2), with the test results fC 1 ; . . . ; C m g of several tumor markers. A set of clinical records, including the test results as well as the diagnosis result, can make up the information system S hU; C [ D; V; f i. Thus, the DE can be evaluated as follows. Algorithm: RST based DE calculation. Input: A dataset S hU; C [ D; V; f i and a set of cutoff values for the multiple markers. For a specic patient xi 2 U, its condition attributes and diagnosis result are denoted as C xi and Dxi respectively. Output: Diagnosis output (DO) of xi and DE of the whole dataset. Procedures of the algorithm: Step 1: Discretize the test results of tumor markers according to the corresponding cutoff values. The cutoff values are obtained by GA approach which will be described in the following subsection Genetic algorithm design. Step 2: Derive the decision rules R C x ! Dx from S using value reduction analysis method of RST. (See Table 8 for details.) Step 3: Generate diagnosis output (DO) for xi. (1) In R, nd all the rules Rxi whose values in C are the same as that of xi and record their accuracy. (2) Let P xi ;d1 represents the accuracy of xi is judged as d1; and P xi ;d2 is the accuracy of xi is judged as d2. (3) Thus, the diagnosis output (DO) of xi can be derived according to the following function.

8 > d1 pxi ;d1 > pxi ;d2 < d2 pxi ;d1 > pxi ;d2 DOxi > : Fail pxi ;d1 > pxi ;d2

285

6
287
288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314

228 229 230 231 232 233 234 235 236 237 238 239

3. The MVMTM In order to improve DE, an algorithm entitled MVMTM is developed. This algorithm can be used to search out the optimal multicutoff values for multiple tumor markers. In this algorithm, GA method is utilized to search for an optimal solution. DE presented in Eq. (1) is used as the tness function in GA searching so that the highest DE will be obtained when the optimal solution is reached. In this algorithm, DE is evaluated according to the decision rules derived by RST. As RST based DE calculation is the core part of MVMTM, it will be described rst. Then, GA procedures are detailed. Thereafter, the owchart of MVMTM is given.

Step 4: Compare DOxi with Dxi , and compute sensitivity Se and specicity Sp. Se records the number of abnormal cases that is correctly diagnosed as abnormal; Sp records the number of normal cases that is correctly diagnosed as normal. Step 5: Compute DE by Eq. (1). 3.2. Genetic algorithm design It is a huge amount of work to nd out the optimal cutoff values. Therefore, genetic algorithm (GA) is employed in this research. Referring to the GA principles stated in Section 2.3, the GA algorithm is constructed as follows. 3.2.1. Chromosome and the initial population Any potential solution to our optimization problem can be expressed as a chromosome. Generally, binary encoding method is used in chromosome design. Nevertheless, the large value range and high precision requirement in our problem would induce a long binary chromosome. Moreover, the multiple cutoff values of various tumor markers would remarkably increase the complexity of chromosome. Hence, a real-value encoding method is used in chromosome design. This encoding method is more natural so that there are no differences between the genotype (coding) and the phenotype (search space) (Yan, Zheng, Jiang, Peng, & Xiao, 2008; Blanco, Delgado, & Pegalajar, 2001). For instance, the following chromosome represents a possible solution to setting two cutoff values for each of the three tumor markers.

240 241 242 243 244 245 246 247 248 249 250 251 252 253

3.1. RST based DE calculation To now, little effort has been dedicated to the diagnosis analysis with multiple tumor markers and multiple cutoff values (MTM MCV). The traditional diagnosis methods, namely, parallel test and serial test, are only suitable for multiple tumor markers with single cutoff value. Therefore, a new diagnosis strategy should be designed for the MTMMCV problem. To this end, it is important to extract decision rules from the medical dataset. Because of the characteristics of clinical data such as inevitable conictions, it is improper to ignore any decision rules. In order to maintain all possible rules in the practical data, value reduction analysis method is employed because it is able to merge the repeated rule patterns in the data set and output the smallest rules set (Chang, Wang, & Wu, 1999).

Chr 3:42 5:78 30:45 90:77 4:07 187:89

315 317

Please cite this article in press as: Su, Q., et al. An algorithm designed for improving diagnostic efciency by setting multi-cutoff values of multiple tumor markers. Expert Systems with Applications (2011), doi:10.1016/j.eswa.2011.11.089

ESWA 7195 5 December 2011

4 318 319 320 321 322 323 Q. Su et al. / Expert Systems with Applications xxx (2011) xxxxxx

No. of Pages 9, Model 5G

324 326
327 328 329 330 331 332 333 334 335 336

In (7), the rst two values of 3.42 and 5.78 represent the cutoff values for the rst marker; the third and fourth values of 30.45 and 90.77 represent the cutoff values for the second marker; and the last two values of 4.07 and 187.89 represent the cutoff values for the third marker. These numbers are randomly set according to the value range of the corresponding marker, denoted as FieldDR.

Genemp0 MaxFieldDR Genemp

10

360 362
363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395

FieldDR

 8

In which, Max(FieldDR) represents the upper limit value of the corresponding marker referring to Eq. (8). An example of mutation operation is shown in Fig. 3. Here, the 5th gene is randomly chosen as the mutation point. And the corresponding FieldDR is [0, 200], thus the new gene changed to 200 1.72 = 198.28. 3.2.3. Best chromosome maintaining strategy Generally speaking, along with the evolutionary progress, more and more superior chromosomes will appear. However, sometimes, because of the randomization in operations of selection, crossover, and mutation, the best chromosomes may be destroyed. This problem could induce an adverse impact on the efciency of GA algorithm. To avoid this shortcoming, the best chromosome maintaining strategy (BCMS) is adopted in our GA algorithm. And the following two steps are included. Step 1: Before selection operation, nd out the best chromosomes, say, 2% of the population that possess the largest DE. Step 2: After the operations of selection, crossover and mutation, nd out the same amount of the worst chromosomes that possess the smallest DE and replace them with the best ones obtained in step 1. 3.2.4. Summary of MVMTM Fig. 4 is the ow chart of MVMTM. The MVMTM can take advantage of the stochastic global search ability of GA to nd out the optimal cutoff values. Since DE is employed as the tness function in GA operations, with the repetitive calculations, the DE of population will be gradually enhanced and the optimal solution will be derived automatically. To prevent the deadlock in calculation, two stop conditions are designed in the GA algorithm. (1) The difference between the maximal DE and the average DE in the population is less than 104; or (2) The evolution computation times is equal to a predetermined level, say 250 in this research. 4. Experimental study 4.1. Experiment dataset In this paper, to improve the diagnostic accuracy of CRC, three tumor markers (CEA, CA 19-9, and CA 50) were considered simultaneously and at most three cutoff values are permitted for each marker. Totally, 212 real medical cases were collected from Renji

550 550 500 500 200 200

Similarly, a number of chromosomes can be generated randomly and used as the initial population. The population of MVMTM is set to be 200. Each chromosome in population is corresponding to a candidate scheme for setting the cutoff values for the three markers. And the cutoff values can be used for discretizing the continuous diagnosis data. Thereafter, the corresponding DE value can be evaluated by the RST based DE calculation algorithm addressed in Section 3.1. And the derived DE value can act as the tness function for the chromosome evaluating and optimizing through evolution operations of selection, crossover, and mutation. 3.2.2. Operations 3.2.2.1. Selection. The functionality of selection operation is to choose the superior chromosomes with larger DEs and assign higher probability to these chromosomes to pass their genes to the new population. The probability for each chromosome Chrj is calculated as follows:

337 338 339 340 341 342

343 345
346 347

DEChr j PChrj Pn i DEChr i

In which, DE(Chrj) is the DE of chromosome Chrj; i = 1, 2, 3, . . . , n; n is the number of chromosomes in population. 3.2.2.2. Crossover. As shown in Fig. 2, a pair of chromosomes Chr1 and Chr2 are chosen for crossover operation. First, a crossover interval, say 35 for this example, is decided randomly. Then, all the genes located in the interval should be swapped to generate new chromosomes Chr1 and Chr2. The crossover rate of MVMTM is set to be 0.8. 3.2.2.3. Mutation. According to the characteristics of the real number chromosome, the following mutation rule is designed and applied in this study. The mutation rate is set to be 0.2 in MVMTM. For a chromosome, a mutation point is decided randomly. The gene located at the mutation point, denoted as Gene(mp), is processed as follows:

348 349 350 351 352 353

396 397 398 399 400 401

354 355 356 357 358 359

Fig. 2. Crossover operation.

Fig. 3. Mutation operation.

Please cite this article in press as: Su, Q., et al. An algorithm designed for improving diagnostic efciency by setting multi-cutoff values of multiple tumor markers. Expert Systems with Applications (2011), doi:10.1016/j.eswa.2011.11.089

ESWA 7195 5 December 2011

Q. Su et al. / Expert Systems with Applications xxx (2011) xxxxxx Table 1 Training set.

No. of Pages 9, Model 5G

CRC group No. of cases M/F Mean (range) age, years 75 47/28 64.5 (2989)

Control group 49 25/44 45.7 (1683)

Table 2 Validation set. CRC group No. of cases M/F Mean (range) age, years 47 30/17 60.3 (2780) Control group 41 22/19 47.0 (1879)

suggested cutoff value was 4.8 lg/L. Serum CA19-9 values were measured by chemiluminescence assay (Roche, Mannheim, Germany) with an Elecsys 2010 instrument (Roche, Mannheim, GerCA 19-9 many). The value range of CA 19-9 was 0500 U/ml and the suggested cutoff value was 33 U/ml. Serum CA 50 values were measured by radioimmunoassay (Union Medical & Pharmaceutical Tech, Ltd, Tianjin, PR China) with a DFM 96 type radio-immunity detection (Zhongcheng mechanic electron technology Co. Hefei, PR China). The detect range of CA50 was 0200 U/ml and the suggested cutoff value was 25 U/ml.

423 424 425 426 427 428 429 430 431 432

CA 50

4.3. DE achieved by parallel test and serial test It is difcult to nd a tumor marker that has both high sensitivity and specicity. So, joint test of two or more tumor markers is usually applied for enhancing DE. In practices, the DE of joint test is usually evaluated either by serial rule or parallel rule. In serial rule, if the test value of any tumor marker is above the corresponding cutoff value, the case is diagnosed as abnormal (CRC patient). In parallel rule, if the test values are all above the respective cutoff values, the case is diagnosed as abnormal (CRC patient). With the suggested cutoff values of CEA, CA19-9 and CA50 given above, the sensitivity, specicity, and DE of the training samples can be evaluated (see Table 3). 4.4. DE achieved by MVMTM

433 434 435 436 437 438 439 440 441 442 443 444

Fig. 4. Flow chart of MVMTM.

402 403 404 405 406 407 408 409 410 411 412 413 414

Hospital, a teaching hospital located in Shanghai, China. These 212 cases are separated into two datasets: training set and validation set. Training set includes 124 cases dated from November 2004 to 2006. Validation set includes 88 cases dated from December 2006 to March 2008. All three tumor markers were all tested and well recorded for each case. These cases are classied into two groups: CRC group (abnormal) and control group (normal). All cases in CRC group were veried by pathological test. The data of control group were collected from healthy persons who performed the physical examinations in the hospital. They had no history of any gastrointestinal diseases and had been follow-upped for more than 6 months. The information of training set and validation set is summarized in Tables 1 and 2. 4.2. Test method of CEA, CA19-9 and CA50 Serum samples were collected from subjects three days before CA 19-9 starting any medication. Clotted blood was CA 50 centrifuged at 4 C, and the serum samples were used for assays on the same day or frozen at 70 C before analyzing. Serum CEA values were measured by chemiluminescence assay (DPC, Tianjin, PR China), with an Immulite 1000 instrument (DPC, Tianjin, PR China). The value range of CEA was 0550 lg/L and the

CA 19-9

CA 50
445 446 447 448 449 450 451 452 453 454 455 456 457

415 416 417 418 419 420 421 422

Theoretically, the number of cutoff values for each tumor marker is unlimited in our MVMTM, but too many cutoff values are impractical and unnecessary. In this study, the number of cutoff values is limited to 1, 2, or 3 for each marker. Therefore, for the three markers, totally 27 different diagnosis schemes can be constructed. In order to nd out the optimal diagnosis scheme efciently, orthogonal experiment technology is employed. According to the corresponding orthogonal table of L9 (33), 27 possible combinations can be simplied to nine orthogonal schemes as listed in the rst column in Table 4. Here, XYZ represents that there are X cutoff values for CEA, Y cutoff values for CA 19-9, and Z cutoff values for CA 50.
Table 3 Parallel test and serial test of the training set. Training dataset Joint test Serial test CRC group & ontrol group Specicity Sensitivity DE 0.90 0.67 0.78 Parallel test 1 0.09 0.55
Q2

Please cite this article in press as: Su, Q., et al. An algorithm designed for improving diagnostic efciency by setting multi-cutoff values of multiple tumor markers. Expert Systems with Applications (2011), doi:10.1016/j.eswa.2011.11.089

ESWA 7195 5 December 2011

6 Q. Su et al. / Expert Systems with Applications xxx (2011) xxxxxx

No. of Pages 9, Model 5G

Table 4 Multi-cutoff values of the orthogonal experiments by MVMTM.

Table 5 DE in training set derived by MVMTM algorithm.

Table 6 DE in validation set by MVMTM in 3-3-2. Validation dataset CRC group & control group Specicity Sensitivity DE # of cutoff values 3-3-2 0.87 0.86 0.87

Table 7 Parallel test and serial test of the validation set. Validation dataset Joint test Serial test CRC group & control group Specicity Sensitivity DE 0.90 0.65 0.77 Parallel test 1 0.15 0.58

determined individually for each tumor marker even though, sometimes, several tumor markers are used jointly. In this study, a novel diagnosis strategy is proposed in which several markers are applied together and analyzed simultaneously. More important, more than one cutoff values are permitted for each of the marker. To search out the optimal cutoff values for multiple markers, an algorithm entitled MVMTM is developed based on the RST and GA method. The advantage of MVMTM is mainly resulted from the heuristic global search ability of GA plus the probabilistic reasoning ability of RST.

473 474 475 476 477 478 479 480 481 482

5.1. The essential part in MVMTM As stated in Section 3, the RST based DE calculation is the essential part in MVMTM. As we known, when more than one cutoff values are set for tumor markers, it is impossible to obtain the diagnosis result (normal or abnormal) using serial or parallel test methods. Thus, the value reduction analysis method of RST is applied to derive the decision rules R C x ! Dx . With these rules, the diagnosis output (DO) can be determined by Eq. (6). Thereafter, sensitivity Se, specicity Sp, and DE can be quantitatively calculated. With the optimal cutoff values derived by MVMTM, the original dataset can be discretized accordingly. In the case study, since three cutoff values are set for CEA, the value of CEA can be separated into four different levels as 1, 2, 3, and 4. Similarly, three cutoff values separate the value of CA 19-9 into four levels and two cutoff values separate the value of CA 50 into three levels. With these cutoff values, using the value induction analysis method of RST, the decision rules can be obtained (as listed in Table 8). Totally, 23 decision rules are derived from the training set. In the column of diagnosis, 0 denotes normal and 2 denotes abnormal. Twenty-three rules may be a big number compared with 124 cases in the testing set. However, there is no worry about over-tting. Actually, it is a merit of the proposed algorithm MVMTM. In

483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505

458 459 460 461 462 463 464 465 466 467 468 469

By MVMTM, the optimal cutoff values for the three markers can be derived for each scheme, as listed in Table 4. Table 5 lists the sensitivity, specicity, and DE for these nine schemes. Given the above analysis results, one can nd that the 3-3-2 is the best scheme in which three cutoff values are set for CEA, 3 for CA 19-9, and two for CA 50. This scheme can enhance the value of DE from 0.78 (Table 3) to 0.86 (Table 5). Furthermore, the validation dataset is employed to testify the effectiveness of scheme 3-3-2 derived above. Table 6 shows that a high DE of 0.87 is achieved in validation. Compared with the parallel or serial tests (see Table 7), it is obvious that MVMTM can improve the DE of CRC from about 0.770.86. 5. Discussions In most of the existing clinical researches, only one cutoff value is set for each tumor marker. Moreover, the cutoff value is

470 471 472

Please cite this article in press as: Su, Q., et al. An algorithm designed for improving diagnostic efciency by setting multi-cutoff values of multiple tumor markers. Expert Systems with Applications (2011), doi:10.1016/j.eswa.2011.11.089

ESWA 7195 5 December 2011

Q. Su et al. / Expert Systems with Applications xxx (2011) xxxxxx Table 8 Decision rules derived by RST. No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
a

No. of Pages 9, Model 5G

CEA 1 1
a

CA 19-9 1 2 1 3 3 1 3 1 2 2 3 3 2 4 1 1
a a

CA 50 1 2 3 1 3 2 2 1 1 1
a

Diagnosis 0 0 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2 2 2 2 2

Support 0.25 0.0242 0.0161 0.0484 0.0081 0.0161 0.0081 0.0081 0.0081 0.0081 0.129 0.0161 0.0645 0.0161 0.0887 0.1048 0.0242 0.0242 0.0806 0.0806 0.0242 0.0081 0.0081

Accuracy 0.7381 1 1 1 0.3333 0.6667 0.25 0.0714 0.1111 0.0909 1 0.6667 0.8889 1 0.2619 0.9286 1 1 0.9091 1 0.75 1 0.3333

Coverage 0.6327 0.0612 0.0408 0.1224 0.0204 0.0408 0.0204 0.0204 0.0204 0.0204 0.2133 0.0267 0.1067 0.0267 0.1467 0.1733 0.04 0.04 0.1333 0.1333 0.04 0.0133 0.0133

Table 10 t-Test results for each pairwise classication methods for MVMTM. Null hypothesis MVMTM 6 SVM MVMTM 6 BPNN MVMTM 6 KNN MVMTM 6 Decision tree MVMTM 6 serial test P value 0.0577 0.0210 0.0078 0.0009 0.0271

1 1 1 1 2 1 2 2 1 1
a

3 1
a

Table 11 t-Test results for each pairwise classication methods for serial test. Null hypothesis Serial Serial Serial Serial test = SVM test = BPNN test = KNN test 6 Decision tree P value 0.3727 0.7088 0.6184 0.044

1 2 3 4 2 2 1
a

1 1
a a

2
a

3 2 1

1 2 2 3 2

Means that it can be any value.

506 507 508 509 510 511 512 513 514 515 516 517 518 519

Section 3.1, it mentioned that the classical value reduction analysis method in RST is able to merge the repeated rule patterns in the data set and output the smallest rule set. Certainly, we can decrease the number of rules by setting limits on support, coverage and accuracy. However, MVMTM does not need any limit because it allows of conicting rules such as the 1st rule and the 15th rule. These situations may occur in real medical diagnosis. It is quite possible for two patients who have the same test results in tumor markers but have different diagnosis results. This phenomenon is in accordance with the indiscernibility characteristic of RST that has been introduced above. That is why we choose value reduction analysis method in RST to generate rules. The value reduction analysis method of RST is able to maintain these conicting rules and reect the clinical facts. There is no over-tting problem because

the nal diagnosis is made according to the value of the accuracy. Some rules with smaller accuracies will not inuence the nal decision. For example, according to the cutoff values in 3-3-2 scheme, a case with 3.2, 10.2, 13 can be discretized as 1, 1, 1 in three tumor markers. Then, two rules (the 1st rule and the 15th rule) in Table 8 can be applied to this case. Here, the 1st rule (1, 1, 1 ? 0) has an accuracy of 0.7381 and the 15th rule (1, 1, 1 ? 2) has an accuracy of 0.2619. Obviously, the 1st rule outweighs the 15th rule. So, the nal diagnosis output is normal. 5.2. Further comparison Besides the traditional serial test, the new proposed algorithm is compared with some baseline statistical machine learning algorithms, namely, SVM (Support Vector Machine), BPNN (Back-Propagation Neutral Network), KNN (K-Nearest Neighbor) and Decision Tree. Here, SVM is employed using software LibSVM-2.89 (Chang & Lin, 2001). BPNN is tested by the NN Toolbox 5.0 in MATLAB 7.2. KNN is run by algorithm IBK in Weka 3.6 (Hall

520 521 522 523 524 525 526 527 528

529 530 531 532 533 534 535 536

Table 9 DE comparison among various classication methods. Fold #1 SVMa Sp Se DE Sp Se DE Sp Se DE Sp Se DE Sp Se DE Sp Se DE

2

Fold #2 0.9375 0.7308 0.8341 0.846 0.75 0.798 0.6875 0.7307 0.7091 0.5625 0.8077 0.6851 1 0.6538 0.8269 1 0.8462 0.9231

Fold #3 0.8125 0.7692 0.7909 0.8125 0.6923 0.7524 0.625 0.8846 0.7548 0.5 0.8846 0.6923 0.875 0.6923 0.7837 0.875 0.7308 0.8029

Fold #4 0.85 0.7727 0.8114 0.8 0.7727 0.7864 0.8 0.7273 0.7636 0.85 0.6818 0.7659 0.8 0.5909 0.6955 0.85 0.8636 0.8568

Fold #5 0.8824 0.7778 0.8301 0.8824 0.7778 0.8301 0.7059 0.9259 0.8159 0.7647 0.7047 0.7347 0.8824 0.7037 0.7931 0.8235 0.8889 0.8562

Average 0.8601 0.7501 0.8051 0.8318 0.7486 0.7902 0.7273 0.8037 0.7655 0.7173 0.7358 0.7265 0.9024 0.6581 0.7803 0.8642 0.8359 0.8501

St. dev. 0.0515 0.0336 0.0309 0.0329 0.0339 0.0280 0.0807 0.0942 0.0393 0.1788 0.1114 0.0364 0.0773 0.0443 0.0501 0.0848 0.0611 0.0478

0.8182 0.7 0.7591 0.8182 0.75 0.784 0.8182 0.75 0.7841 0.9091 0.6 0.7546 0.9545 0.65 0.8023 0.7727 0.85 0.8114

BPNNb

KNNc

Decision tree

Serial test

MVMTM

RBF kernel Kx; y eckxyk ; penalty parameter C and kernel parameters c have been optimized with internal command of LibSVM-2.89. A 3-layer BPNN with 3,7 and 2 as the number of input, hidden and output layers respectively; The hidden nodes use TANSIG transfer function and the output nodes use LOGSIG transfer function. c The parameter K in IBK is tested from 1 to 10 for each fold. Only highest DEs are listed here.
b

Please cite this article in press as: Su, Q., et al. An algorithm designed for improving diagnostic efciency by setting multi-cutoff values of multiple tumor markers. Expert Systems with Applications (2011), doi:10.1016/j.eswa.2011.11.089

ESWA 7195 5 December 2011

8 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 Q. Su et al. / Expert Systems with Applications xxx (2011) xxxxxx

No. of Pages 9, Model 5G

et al., 2009). Decision Tree is run by algorithm J48 in Weka 3.6. 33-2 scheme of MVMTM is utilized as the benchmark. 5-fold cross-validation is conducted and the comparison results are listed in Table 9. We also performed a t-test to verify whether MVMTM have better performance than other ve methods (see Table 10). Table 10 shows that the average DE of MVMTM is signicantly better than that of KNN and Decision Tree at the 99% condence interval. Likewise, it is signicantly better than Serial Test and BPNN at the 95% condence interval. It is signicantly better than SVM at the 90% condence interval. In addition, Table 11 compares the efcacy of Serial Test and other four methods. It shows that serial test works better than Decision Tree at the 95% condence interval and it has not significant difference with SVM, BPNN and KNN. These comparisons demonstrate that the medical method (serial test) is efcient compared with other methods. However, the new proposed algorithm MVMTM can improve the DE further.

554 555 556 557 558 559 560 561 562 563 564 565 566 567

6. Conclusions In this paper, focusing on how to set the multi-cutoff values for the three tumor markers of CEA, CA 19-9, and CA 50 in CRC diagnosis, an articial intelligent algorithm entitled MVMTM has been proposed based on the RST and GA. The experimental study demonstrates that, with the MVMTM, the DE of CRC can be enhanced substantially. Although some progress has been made in this study, many issues are still open for future research. For instance, the algorithm should be improved further in terms of the calculation speed and quality. Meanwhile, some other types of combination of different tumor markers should be studied systematically. In addition, the application of this algorithm to diagnosis of other diseases should be explored. Acknowledgements This paper is sponsored by National Nature Science Foundation of China (70672077 and 70832005), Medical & Engineering Interdisciplinary Research Foundation of Shanghai Jiao Tong University (YG2007ZD10), Shanghai Leading Academic Discipline Project (S30203) and sponsored by Research Center for Healthcare Management, SEM, Tsinghua University. References
An, A., Stefanowski, J., Ramanna, S., Butz, C., Pedrycz, W., Wang, G., et al. (2007). Visualization of rough set decision rules for medical diagnosis systems. In Rough sets, fuzzy sets, data mining and granular computing (Vol. 4482, pp. 371378). Berlin/Heidelberg: Springer. Armitage, N., Davidson, A., Tsikos, D., & Wood, C. (1984). A study of the reliability of carcinoembryonic antigen blood levels in following the course of colorectal cancer. Clinical Oncology, 10, 141147. Blanco, A., Delgado, M., & Pegalajar, M. C. (2001). A real-coded genetic algorithm for training recurrent neural networks. Neural Networks, 14, 93105. Carpelan-Holmstrom, M., Haglund, C., Kuusela, P., Jarvinen, H., & Roberts, P. (1995). Preoperative serum levels of CEA and CA 242 in colorectal cancer. British Journal of Cancer, 71, 868872.

568 569 570 571 572 573 574

575 576 577 578 579 580 581 582 583 584 585 586 587

Carriquiry, L. A., & Pineyro, A. (1999). Should carcinoembryonic antigen be used in the management of patients with colorectal cancer. Diseases of the Colon & Rectum, 42, 921929. Castiglione, F., Pappalardo, F., Bernaschi, M., & Motta, S. (2007). Optimization of HAART with genetic algorithms and agent-based models of HIV infection. Bioinformatics, 23, 33503355. Chang, C.-C., & Lin, C.-J. (2001). Libsvm: A library for support vector machines. Software available at <http://www.csie.ntu.edu.tw/~cjlin/libsvm>. Chang, L.-Y., Wang, G.-Y., & Wu, Y. (1999). An approach for attribute reduction and rule generation based on rough set theory. Journal of Software, 11, 371378. Duffy, M. J. (2001). Carcinoembryonic antigen as a marker for colorectal cancer: Is it clinically useful? Clinical Chemistry, 47, 624630. Fakih, S. J., & Das, T. K. (2006). Lead: A methodology for learning efcient approaches to medical diagnosis. IEEE Transactions on Information Technology in Biomedicine, 10, 220228. Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., & Witten, I. H. (2009). The weka data mining software: An update. SIGKDD Explorations, 11. Ilczuk, G., Mlynarski, R., Wakulicz-Deja, A., Drzewiecka, A., & Kargul, W. (2005). Rough set techniques for medical diagnosis systems. Computers in Cardiology, 32, 837840. Krner, H., Sreide, K., Stokkeland, P., & Sreide, J. (2007). Diagnostic accuracy of serum-carcinoembryonic antigen in recurrent colorectal cancer: A receiver operating characteristic curve analysis. Annals of Surgical Oncology, 14, 417423. Lewis, J. D. (2000). Prevention and treatment of colorectal cancer: Pay now or pay later. Annals of Internal Medicine, 133, 647649. Lollini, P.-L., Motta, S., & Pappalardo, F. (2006). Discovery of cancer vaccination protocols with a genetic algorithm driving an agent based simulator. BMC Bioinformatics, 7, 352. Lucarotti, M., Habib, N., Kelly, S., Rothnie, N., Nelson, O., Lindholm, L., et al. (1991). Clinical evaluation of combined use of CEA, CA19-9 and CA50 in the serum of patients with pancreatic carcinoma. European Journal of Surgery Oncology, 17, 5153. Modchang, C., Triampo, W., & Lenbury, Y. (2008). Mathematical modeling and application of genetic algorithm to parameter estimation in signal transduction: Trafcking and promiscuous coupling of G-protein coupled receptors. Computers in Biology and Medicine, 38, 574582. Moertel, C. G., Fleming, T. R., Macdonald, J. S., Haller, D. G., Laurie, J. A., & Tangen, C. (1993). An evaluation of the carcinoembryonic antigen (CEA) test for monitoring patients with resected colon cancer. The Journal of the American Medical Association, 270, 943947. Moghimi, M., & Ghodosi, I. (2007). Clinical efcacy of combined use of cea-ca19-9 in patients with pancreatic malignancy. Journal of Medical Science, 7, 625629. Pattaraintakorn, P., & Cercone, N. (2008). Integrating rough set theory and medical applications. Applied Mathematics Letters, 21, 400403. Persijn, J. P., & Hart, A. A. M. (1981). Prognostic signicance of CEA in colorectal cancer: A statistical study. Journal of Clinical Chemistry and Clinical Biochemistry, 19. Smigrodzki, R., Goertzel, B., Pennachin, C., Coelho, L., Prosdocimi, F., & Davis Parker, W. Jr., (2005). Genetic algorithm for analysis of mutations in Parkinsons disease. Articial Intelligence in Medicine, 35, 227241. Streiner, D. L., & Cairney, J. (2007). Whats under the ROC? An introduction to receiver operating characteristics curves. Canadian Journal of Psychiatry, 52, 121128. Wakulicz-Deja, A., & Peszek, P. (2003). Applying rough set theory to multi stage medical diagnosing. Fundamenta Informaticae, 54, 387408. Wan, S., & Zhang, B. (2007). Smooth semiparametric receiver operating characteristic curves for continuous diagnostic tests. Statistics in Medicine, 26, 25652586. Weiss, H. L., Niwas, S., Grizzle, W. E., & Piyathilake, C. (2004). Receiver operating characteristic (ROC) to determine cut-off points of biomarkers in lung cancer patients. Disease Markers, 19, 273278. Wichmann, M. W., Lau-Werner, U., Muller, C., Stierber, P., & Schildberg, F.-W. (2000). Carcinoembryonic antigen for the detection of recurrent disease following curative resection of colorectal cancer. Anticancer Research, 20, 49534955. Wood, C. B., Ratcliffe, J. G., Burt, R. W., Malcolm, A. J. H., & Blumgart, L. H. (1980). The clinical signicance of the pattern of elevated serum carcinoembryonic antigen (CEA) levels in recurrent colorectal cancer. British Journal of Surgery, 67, 4648. Yan, H., Zheng, J., Jiang, Y., Peng, C., & Xiao, S. (2008). Selecting critical clinical features for heart diseases diagnosis with a real-coded genetic algorithm. Applied Soft Computing, 8, 11051111. Yu, C., & Xu, Y. (1998). Basic principals of ROC analysis. Chinese Journal of Epidemiology, 9, 413415.

588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661

Please cite this article in press as: Su, Q., et al. An algorithm designed for improving diagnostic efciency by setting multi-cutoff values of multiple tumor markers. Expert Systems with Applications (2011), doi:10.1016/j.eswa.2011.11.089