Journal Article Page
European UrologyVolume 57, issue 3, pages 363-550, March 2010
The Application of Artificial Intelligence to Microarray Data: Identification of a Novel Gene Signature to Identify Bladder Cancer Progression
Accepted 27 October 2009, Published online 6 November 2009, pages 398 - 406
New methods for identifying bladder cancer (BCa) progression are required. Gene expression microarrays can reveal insights into disease biology and identify novel biomarkers. However, these experiments produce large datasets that are difficult to interpret.
To develop a novel method of microarray analysis combining two forms of artificial intelligence (AI): neurofuzzy modelling (NFM) and artificial neural networks (ANN) and validate it in a BCa cohort.
Design, setting, and participants
We used AI and statistical analyses to identify progression-related genes in a microarray dataset (n = 66 tumours, n = 2800 genes). The AI-selected genes were then investigated in a second cohort (n = 262 tumours) using immunohistochemistry.
We compared the accuracy of AI and statistical approaches to identify tumour progression.
Results and limitations
AI identified 11 progression-associated genes (odds ratio [OR]: 0.70; 95% confidence interval [CI], 0.56–0.87; p = 0.0004), and these were more discriminate than genes chosen using statistical analyses (OR: 1.24; 95% CI, 0.96–1.60; p = 0.09). The expression of six AI-selected genes (LIG3, FAS, KRT18, ICAM1, DSG2, and BRCA2) was determined using commercial antibodies and successfully identified tumour progression (concordance index: 0.66; log-rank test: p = 0.01). AI-selected genes were more discriminate than pathologic criteria at determining progression (Cox multivariate analysis: p = 0.01). Limitations include the use of statistical correlation to identify 200 genes for AI analysis and that we did not compare regression identified genes with immunohistochemistry.
AI and statistical analyses use different techniques of inference to determine gene–phenotype associations and identify distinct prognostic gene signatures that are equally valid. We have identified a prognostic gene signature whose members reflect a variety of carcinogenic pathways that could identify progression in non–muscle-invasive BCa.
The care of patients with urothelial cell carcinoma (UCC) of the bladder could be significantly improved if their tumour behaviour were accurately identified at diagnosis. Patients with nonprogressive superficial disease could be spared endoscopic surveillance and bacillus Calmette-Guérin immunotherapy, while those at greatest risk of progression could opt for early cystectomy. For invasive tumours, the use of systemic chemotherapy could be rationalised to cases with the greatest progression risk. Tumour behaviour can be difficult to determine from histopathology alone. For example, the progression risk for non-muscle UCC varies between <1% and >50%  and . Furthermore, as stage and grade are often linked, when one is fixed (eg, stage), the other performs poorly (eg, grade) at identifying tumour progression. It is hoped that molecular knowledge will reveal an understanding of tumour biology that allows accurate phenotype identification.
As current biomarkers are insufficiently robust for clinical practice, microarrays have been used to identify new candidates  and . Microarray experiments reveal great insights into tumour biology, but the cost and magnitude of these experiments prohibit large sample-size analyses. Thus, microarray datasets have high dimensionality (large imbalance between gene number and sample size) that leads to analytical difficulties , , and . Successful analysis requires the identification of genes related to tumour class and the removal of noncontributing variables. Poor analysis leads to data overfitting and irreproducible results . Traditional analytical techniques, such as hierarchical clustering, assume biological linearity and use statistical proximity to infer class–gene relationships (“feature selection”). They perform poorly in datasets contaminated with variable noise. Artificial intelligence (AI) is a machine learning approach without these prerequisites. Various AI techniques exist , and successful microarray analysis has been reported using artificial neural networks (ANN)  and  and support vector machines (SVMs)  and  in nonurothelial malignancies. However, the hidden working layer of an ANN prevents model understanding and hinders its acceptance by the scientific community , while SVMs still use proximity to infer class–gene associations and function poorly with respect to interpretability .
An alternative form of AI is the neurofuzzy model (NFM). This model has a similar design to an ANN but uses a transparent fuzzy logic internal structure . This transparency allows model understanding and parameter interrogation and can facilitate the inclusion of a priori qualitative knowledge. When used to identify tumour progression, we have previously found that NFM is accurate, reproducible, and appears superior to regression-based classifications  and . We hypothesised that NFM could improve microarray analysis and identify prognostic gene panels that could accurately predict the behaviour of UCC. To test this hypothesis, we examined a previously reported non–muscle-invasive UCC microarray dataset to find genes associated with progression to invasion. Genes associated with progression were then tested in a new, larger UCC cohort using immunohistochemistry.
2. Materials and methods
2.1. Patients and tumours
We studied two patient populations (Table 1). For microarray analysis, we used 66 tumours from 34 patients treated at the Ludwig Maximilian University, Germany . Progression to muscle invasion occurred in 10 of 34 patients (29%), and the median follow-up was 43 mo. For immunohistochemical analysis, we studied 262 tumours from separate consecutive patients treated at the University of Regensburg, Germany. We created a tissue microarray (TMA) using paraffin-embedded formalin-fixed tissues with two cores per cases (1.2 mm) . Progression information was available for 182 of 262 (69.5%) patients, and muscle invasion or new metastases occurred in 49 patients (26.9%). The median follow-up was 89 mo (range: 2–154). No patients were in both UCC populations. Normal urothelium from patients with benign prostatic hyperplasia (n = 20) and coexisting UCC (n = 15) was also analysed. Institutional review board approval was obtained from both institutions prior to study commencement.
|Gene array tumours||TMA tumours|
|Stage TNM 1998||Normal||8||100.0||–||–|
|Stage TNM 2004||PUNLMP||1||1.5||22||8.4|
|Time to progression, median (range), mo||21 (0–60)||23 (1–154)|
|OS time, median (range), mo**||43 (0–109)||90 (24–154)|
* In the 34 individual patients.
** In nonprogressing patients.
*** In 134 primary superficial tumours with available follow-up information.
TMA = tissue microarray; PUNLMP = papillary urothelial neoplasm of low malignant potential; CIS = carcinoma in situ; UCC = urothelial cell carcinoma; OS = overall survival.
2.2. RNA extraction and gene expression microarray analysis
The microarray (metg001A) contained 2800 genes (6117 probe sets) annotated by the GoldenPath assembly. The microarray experiments and data processing are reported in detail elsewhere .
2.3. Artificial intelligence feature selection
To analyse the microarray data, we used a “committee of models” approach that assimilated findings from each individual AI model (Fig. 1), as we wanted to determine gene-progression relationships that were not dependent upon one AI structure. We initially performed a dimension reduction using the Pearson coefficient to identify the 200 genes most associated with progression. These selected genes were then analysed using iterative ANN and NFM models in two structures, which we termed selectivity and averaging (Fig. 1). These structures enable simultaneous analysis of all genes rather than a “leave-one-out” approach. ANNs were produced within Statistica v.7 (StatSoft, Bedford, UK). NFMs were produced within Matlab v.6.5 (The MathWorks, Natick, MA, USA), and progression predictions were performed using an in-house software suite  and . The data were divided into 90% for training (60% for learning and 30% for validation) and 10% for testing. Ensembling and cross-validation were used to maximise data .
We ranked the 200 genes according to the size of model error induced by their alteration. Those with the largest error were ranked highest, as an alteration of their values produced the largest disturbance in the models accuracy. For each gene, a committee ranking was produced from the average score of the individual AI models. A panel of progression-related genes was produced from those with the highest ranking. This committee panel was compared with the original gene panel selected using Pearson's linear regression coefficient and GeneCluster v.2.0 software (Broad Institute, Cambridge, MA, USA) . This original panel included 11 members:
- • Fatty acid binding protein 4, adipocyte (FABP4)
- • Glutathione S-transferase μ 4 (GSTM4)
- • Serpin peptidase inhibitor, clade A (α-1 antiproteinase, antitrypsin), member 1 (SERPINA1)
- • Histone deacetylase 1 (HDAC1)
- • TPX2, microtubule-associated, homolog (TPX2; formerly known as C20ORF1)
- • Dynein, light chain, roadblock-type 1 (DYNLRB1; formerly known as DNLC2A)
- • PTK6 protein tyrosine kinase 6 (PTK6)
- • Ubiquitin C (UBC)
- • O-6-methylguanine-DNA methyltransferase (MGMT)
- • Integrin β 3 binding protein (ITGB3BP)
- • Poly[A] binding protein interacting protein 2 (PAIP2).
To evaluate the committee approach, we analysed the expression of its highest-ranking members using immunohistochemistry in a new UCC cohort  and . Commercially manufactured antibodies were available for six members:
- • Ligase III, DNA, ATP-dependent (LIG3; clone 6G9; Abcam, Cambridge, UK; dilution 1:50)
- • Breast cancer 2, early onset (BRCA2; Abcam, Cambridge, UK; dilution 1:10)
- • Fas (TNF receptor superfamily, member 6) (FAS; formerly known as TNFRSF6; Abcam, Cambridge, UK; dilution 1:25)
- • Keratin 18 (KRT18; clone CK2; Chemicon, Billerica, MA, USA; dilution 1:50)
- • Desmoglein 2 (DSG2; clone 3G132; Abcam, Cambridge, UK; dilution 1:10)
- • Intercellular adhesion molecule 1 (ICAM1; clone 23G12; Lab Vision, Fremont, CA, USA; dilution 1:10).
For negative controls, the primary antibody was omitted. Immunostained sections were scored independently for the percentage of positive tumour cells by uropathologists (PW, AH). The abnormal status for each protein was defined according to its cellular function, its contrast with normal urothelial expression, and from previous reports. For ICAM1, a case was considered positive if >30% of intratumoural blood vessels were stained. For LIG3, BRCA2, FAS, and DSG2, abnormal expression was defined as a loss or reduction of staining (0% or ≤30% positively stained cells). For both, normal urothelium had expression in >50% of cells. Abnormal KRT18 expression was defined as increased immunostaining (≥80% cells with positive staining) with respect to normal samples, which were negative in 90% of cases.
2.5. Statistical analysis
All analyses were two tailed and carried out using SPSS v.14 software (SPSS Inc, Chicago, IL, USA). Categorical variables were compared using the χ2 test and continuous variables with a t test. Disease progression was defined when a non–muscle-invasive tumour became invasive or a muscle-invasive tumour developed metastases. Progression-specific survival probability following tumour resection was analysed using the Kaplan-Meier method and log-rank test. Patients without progression were censored when they were last reviewed or when they died of other causes. The concordance index was calculated as reported . A p value of <0.05 was interpreted as statistically significant. Cox regression multivariate analysis was used to compare the prognostic value of the various gene panels with clinicopathologic parameters.
3.1. Dimension reduction
We aimed to produce a prognostic gene panel of around 11 members to allow comparison with the original panel chosen by statistical methods. Analysis of predictive ANN and NFM models with increments of 1 to 200 members revealed that this was feasible (Fig. 2). For NFM, the modelling error with 11 genes (root mean squared [RMS]: 0.135) was similar to that for >157 genes (both concordance index: 1.0). For ANN, the error did not change until >140 gene inputs were used (RMS: 0.37 for 11 genes) and was larger than the equivalent for NFM.
3.2. Gene ranking and comparison of feature selection panels
We ranked the 200 genes according to their average score from the various AI models (Table 2) and selected the 11 highest-ranked genes to compare with the original panel. Using gene expression dichotomised around the mean, both panels were able to stratify tumour progression, although the committee panel appeared more discriminate. For example, the findings of the committee panel are typical (Fig. 3a), while individual members are associated with tumour progression (eg, LIG3: p = 0.01; KRT18: p = 0.04; log-rank values). The best prediction of progression occurs when the members are used in combination (≥3 of 11 abnormal genes: p = 0.007; ≥4 of 11: p = 0.0004; ≥5 of 11: p = 0.002; log-rank values). In multivariate analysis, the committee panel (odds ratio [OR]: 0.70; 95% confidence interval [CI], 0.56–0.87; log-rank p = 0.0004) was better at identifying progression than grade (OR: 0.38; 95% CI, 0.15–0.91; p = 0.001) and stage (OR: 0.65; 95% CI, 0.1–4.31; p = 0.03) than the original panel (OR: 1.24; 95% CI, 0.96–1.60; p = 0.09). No members were shared between the committee and original panels.
|1||PPA1||Pyrophosphatase (inorganic) 1 (formerly known as PP)||Phosphate metabolism/metabolism|
|2||FAS||Fas (TNF receptor superfamily, member 6) (also known as CD95; formerly known as TNFRSF6)||Apoptosis/immune response/signal transduction|
|3||LIG3||Ligase III, DNA, ATP-dependent||Cytokinesis/DNA replication and repair/meiosis|
|4||BRCA2||Breast cancer 2, early onset||Cell cycle control/double-strand break repair/DNA replication/chromatin architecture/apoptosis|
|5||ICAM1||Intercellular adhesion molecule 1 (also known as CD54)||Cell–cell adhesion|
|6||RND3||Rho family GTPase 3 (formerly known as ARHE)||Cell adhesion/signal transduction/cytoskeleton organisation|
|7||NACA||Nascent-polypeptide-associated complex α subunit||Protein biosynthesis/nascent polypeptide association|
|8||DSG2||Desmoglein 2||Cell adhesion/homophilic cell adhesion|
|9||KRT18||Keratin 18||Embryogenesis and morphogenesis|
|10||C1orf115||Chromosome 1 open reading frame 115 (also known as FLJ14146)||Unknown function|
|11||TP53BP2||Tumour protein p53 binding protein, 2 (also known as ASPP2)||Cell cycle/apoptosis regulation/signal transduction|
* Genes shown in bold were analysed by immunohistochemistry.
3.3. Analysis of the committee panel in a second tumour cohort
Six of the 11 members in the committee panel (LIG3, BRCA2, FAS, KRT18, DSG2, and ICAM1; Fig. 4a) have commercially manufactured antibodies with proven reproducible staining patterns in formalin-fixed paraffin–embedded tissue. Using these antibodies, we performed immunohistochemistry on the 262-tumour TMA. When protein expression was analysed with respect to tumour histology, various associations were seen. For example, LIG3 and ICAM1 were associated with tumour stage and grade (χ2: p < 0.05; Table 3) when compared to tumours with normal expression. However, when expression of individual proteins with respect to tumour behaviour was analysed, few significant relationships were present. Only abnormal FAS expression was significantly associated with tumour progression (log-rank test: p = 0.003).
|FAS Abnormal/ total (%)||χ2||LIG3 Abnormal/total (%)||χ2||ICAM1 Abnormal/total (%)||χ2||DSG2 Abnormal/total (%)||χ2||BRCA2 Abnormal/total (%)||χ2||KRT18 Abnormal/total (%)||χ2|
|Grade||1||16/69 (28)||–||42/66 (64)||–||46/61 (75)||–||35/67 (52)||–||23/69 (33)||–||29/61 (48)||–|
|2||15/65 (23)||–||44/66 (67)||–||41/62 (66)||–||37/65 (57)||–||22/66 (33)||–||30/60 (50)||–|
|3||13/93 (14)||0.093||33/95 (35)||0.0001||40/93 (43)||0.0001||63/93 (68)||0.119||40/95 (42)||0.398||46/88 (52)||0.850|
|Stage||PUNLMP||7/19 (37)||–||10/18 (56)||–||11/16 (69)||–||12/18 (67)||–||5/18 (28)||–||7/16 (44)||–|
|pTa||28/112 (25)||–||70/110 (64)||–||72/103 (70)||–||55/109 (51)||–||37/113 (33)||–||51/100 (51)||–|
|pT1||6/42 (14)||–||17/43 (40)||–||26/44 (59)||–||31/43 (72)||–||18/42 (43)||–||21/41 (51)||–|
|pT2||6/49 (12)||–||21/51 (41)||–||18/48 (38)||–||34/50 (68)||–||23/52||–||24/48 (50)||–|
|pT3||0/2 (0)||–||0/2 (0)||–||0/2 (0)||–||1/2 (50)||–||0/2 (0)||–||2/2 (100)||–|
|pT4||0/3 (0)||0.119||1/3 (33)||0.018||0/3 (0)||0.001||2/3 (66)||0.128||2/3 (67)||0.344||0/2 (0)||0.506|
|CIS||Absent||42/200 (21)||–||110/198 (67)||–||115/189 (61)||–||117/169 (60)||–||75/201 (37)||–||90/182 (50)||–|
|Present||5/27 (19)||0.765||9/29 (31)||0.014||12/27 (44)||0.105||18/29 (62)||0.807||10/29 (35)||0.768||15/27 (56)||0.554|
|Growth||Papillary||42/187 (23)||–||99/185 (53)||–||115/177 (63)||–||103/184 (56)||–||69/187 (37)||–||88/170 (52)||–|
|Solid||4/39 (10)||0.085||20/41 (49)||0.583||14/38 (37)||0.003||31/40 (78)||0.012||15/42 (36)||0.886||17/38 (45)||0.433|
|PFS*||Progression||12/34 (35)||0.003||21/33 (64)||0.566||17/31 (55)||0.155||23/33 (70)||0.081||13/33 (40)||0.802||17/30 (57)||0.458|
|No||10/84 (12)||–||48/83 (58)||–||54/78 (69)||–||42/81 (52)||–||31/84 (37)||–||35/72 (49)||–|
* Only superficial tumours were analysed for this outcome.
PUNLMP = papillary urothelial neoplasm of low malignant potential; PFS = progression-free survival.
Note: For each variable, the numerator is the number of abnormally immunostained tumours, and the denominator is the number successfully analysed for that protein.
We then analyzed the six proteins together as a committee panel using only superficial tumours (n = 134). Each tumour was scored according to the number of proteins with abnormal staining, and this value was expressed as a percentage of the total number successfully immunostained for that sample. Only samples with four or more stained proteins were evaluated. When progression was analysed with respect to this score, significantly worse outcomes were present in tumours with higher than lower scores (Fig. 4b). As with its use in the first tumour cohort, the panel's discriminating ability was maximal at its mean content (concordance index: 0.66; log-rank test: p = 0.02 for 40% and p = 0.01 for 50%). In multivariate analysis, the committee panel was better at stratifying progression (Cox OR: 1.2; 95% CI, 1.1–1.3; p = 0.014) than tumour stage (OR: 1.44; 95% CI, 0.82–2.53; p = 0.2), grade (OR: 0.93; 95% CI, 0.53–1.66; p = 0.8), the presence of carcinoma in situ (CIS; OR: 1.3; 95% CI, 0.54–3.12; p = 0.6), growth pattern (OR: 0.74; 95% CI, 0.26–2.12; p = 0.6), and multifocality (OR: 1.61; 95% CI, 0.61–4.24; p = 0.3).
We have used AI to examine the relationship between gene expression and progression. To evaluate this approach, rather than specific model designs, we used a committee of models to merge gene rankings from individual models and structures. AI can identify complex relationships within nonlinear data contaminated by variable noise and, as such, can outperform statistical regression  and . AI modelling is a generic process, and these methods could be applied to reinterrogate microarray datasets for prognostic and functional data.
Our approach reduced 200 genes to 11, with minimal deterioration in progression identification. The highest-ranked genes appeared better at predicting tumour outcome than those selected using traditional analysis and pathologic criteria. The fuzzy logic layer of our committee NFM is shown in Fig. 3b. This rule base consists of parallel rules in which the fuzzy logic component can be visualized. In rule 1 (top line), high KRT18 in combination with low DSG2 and FAS expression leads to rapid tumour progression (final box). This supports known carcinogenic functions of these genes, as KRT18 is an oncogene and the others are tumour suppressors . One can also see that the discriminatory effects in tumour protein p53 binding protein, 2 (TP53BP2) are less apparent than for other genes (TP53BP2 was ranked 11th; Table 2).
The ability of AI to determine nonlinear relationships is demonstrated in our results. Of the 11 genes that compose the committee panel, only FAS was individually associated with tumour progression. However, the cumulative use of this panel allowed accurate progression discrimination (Fig. 4b). The members of the committee panel represent various carcinogenic pathways. Their association with progression may be directly through carcinogenic roles or as bystanders associated with progression. Their diversity in roles suggests that they may function as synergistic facilitators of progression. Apoptosis evasion is represented by reduced expression of FAS, TP53BP2, and Rho family GTPase 3 (RND3; formerly known as ARHE).FAS is important for apoptosis induction, and decreased expression is associated with advanced bladder cancer (BCa) stage, grade, and progression . TP53BP2 (also known as apoptosis stimulating protein of p53 2 [ASPP2]) plays a key role in apoptosis induction through the activation of p53. Reduced TP53BP2 expression abrogates the onset of apoptosis in cancer but has not been reported in UCC. Tumour invasion is represented by reduced cellular adhesion (ICAM1 and DSG2) and cytoskeletal reorganisation through increased KRT18 and reduced RND3 expression. DSG2 is a cellular adhesion molecule whose loss reduces adhesion, increases invasion, and speeds tumour progression . ICAM1 is also an intercellular adhesion molecule and is frequently epigenetically silenced in UCC (>70%) . KRT18 is a cytokeratin known to be expressed in the umbrella layer of urothelium whose expression increases with urothelial carcinogenesis . RND3 is a Rho signal transduction member with roles in many cellular processes (cytoskeleton organisation, membrane trafficking, cell growth, and apoptosis) , whose loss is reported in prostate cancer. Deranged DNA repair is represented by BRCA2 and LIG3. Although neither is directly linked with bladder carcinogenesis, it is possible that loss of both is required for carcinogenic alteration. BRCA2-deficient cells have reduced DNA ligation capacity, which can be reversed by LIG3 administration .
AI can analyse microarray datasets in a complementary manner to statistical analyses. Both methods use different techniques of inference to determine gene–phenotype associations and thus identify distinct prognostic gene signatures that are equally valid. We have identified a new prognostic gene signature in UCC, whose members reflect a variety of carcinogenic pathways. This signature requires validation in new tumour cohorts to assess its ability to identify progression in non–muscle-invasive BCa.
Author contributions: James W.F. Catto had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.
Study concept and design: Catto, Abbod, Wild, Hamdy, Linkens, Hartmann.
Acquisition of data: Catto, Abbod, Linkens, Wild, Pilarsky, Denzinger, Burger, Stoehr, Knuechel, Hartmann.
Analysis and interpretation of data: Catto, Abbod, Linkens, Wild, Hartmann.
Drafting of the manuscript: Catto, Abbod, Linkens, Wild, Rosario, Rehman, Hartmann.
Critical revision of the manuscript for important intellectual content: Catto, Abbod, Linkens, Wild, Rosario, Rehman, Hartmann.
Statistical analysis: Catto, Abbod, Rosario, Hartmann.
Obtaining funding: Catto, Hartmann, Hamdy.
Administrative, technical, or material support: Catto, Hartmann, Hamdy.
Supervision: Linkens, Hamdy, Hartmann.
Other (specify): None.
Financial disclosures: I certify that all conflicts of interest, including specific financial interests and relationships and affiliations relevant to the subject matter or materials discussed in the manuscript (eg, employment/affiliation, grants or funding, consultancies, honoraria, stock ownership or options, expert testimony, royalties, or patents filed, received, or pending), are the following: None.
Funding/Support and role of the sponsor: None.
-  B.W.G. van Rhijn, M. Burger, Y. Lotan, et al. Recurrence and progression of disease in non-muscle-invasive bladder cancer: from epidemiology to treatment strategy. Eur Urol. 2009;56:430-442 Abstract, Full-text, PDF, Crossref.
-  R.J. Sylvester, A.P.M. van der Meijden, W. Oosterlinck, et al. Predicting recurrence and progression in individual patients with stage Ta T1 bladder cancer using EORTC risk tables: a combined analysis of 2596 patients from seven EORTC trials. Eur Urol. 2006;49:466-477 Abstract, Full-text, PDF, Crossref.
-  L. Dyrskjot, T. Thykjaer, M. Kruhoffer, et al. Identifying distinct classes of bladder carcinoma using microarrays. Nat Genet. 2003;33:90-96 Crossref.
-  L. Dyrskjot, K. Zieger, F.X. Real, et al. Gene expression signatures predict outcome in non-muscle-invasive bladder carcinoma: a multicenter validation study. Clin Cancer Res. 2007;13:3545-3551 Crossref.
-  D.F. Ransohoff. Rules of evidence for cancer molecular-marker discovery and validation. Nat Rev Cancer. 2004;4:309-314 Crossref.
-  F.C. Hamdy, J.W. Catto. Less is more: artificial intelligence and gene-expression arrays. Lancet. 2004;364:2003-2004 Crossref.
-  A. Dupuy, R.M. Simon. Critical review of published microarray studies for cancer outcome and guidelines on statistical analysis and reporting. J Natl Cancer Inst. 2007;99:147-157 Crossref.
-  M.F. Abbod, J.W. Catto, D.A. Linkens, F.C. Hamdy. Application of artificial intelligence to the management of urological cancer. J Urol. 2007;178:1150-1156 Crossref.
-  J. Khan, J.S. Wei, M. Ringner, et al. Classification and diagnostic prediction of cancers using gene expression profiling and artificial neural networks. Nat Med. 2001;7:673-679 Crossref.
-  Lancashire LJ, Powe DG, Reis-Filho JS, et al. A validated gene expression profile for detecting clinical outcome in breast cancer using artificial neural networks. Breast Cancer Res Treat. In press. doi:10.1007/s10549-009-0378-1.
-  M.P. Brown, W.N. Grundy, D. Lin, et al. Knowledge-based analysis of microarray gene expression data by using support vector machines. Proc Natl Acad Sci U S A. 2000;97:262-267 Crossref.
-  T.M. Huang, V. Kecman. Gene extraction for cancer diagnosis by support vector machines—an improvement. Artif Intell Med. 2005;35:185-194 Crossref.
-  G. Schwarzer, W. Vach, M. Schumacher. On the misuses of artificial neural networks for prognostic and diagnostic classification in oncology. Stat Med. 2000;19:541-561 Crossref.
-  T. Hastie, R. Tibshirani, J. Friedman. The elements of statistical learning: data mining, inference and prediction. (Springer-Verlag, New York, NY, 2001)
-  J.W. Catto, M.F. Abbod, D.A. Linkens, F.C. Hamdy. Neuro-fuzzy modeling: an accurate and interpretable method for predicting bladder cancer progression. J Urol. 2006;175:474-479 Crossref.
-  J.W. Catto, M.F. Abbod, D.A. Linkens, S. Larre, D.J. Rosario, F.C. Hamdy. Neurofuzzy modeling to determine recurrence risk following radical cystectomy for non-metastatic urothelial carcinoma of the bladder. Clin Cancer Res. 2009;15:3150-3155 Crossref.
-  P.J. Wild, A. Herr, C. Wissmann, et al. Gene expression profiling of progressive papillary noninvasive carcinomas of the urinary bladder. Clin Cancer Res. 2005;11:4415-4429 Crossref.
-  J.M.M. van Oers, P.J. Wild, M. Burger, et al. FGFR3 mutations and a normal CK20 staining pattern define low-grade noninvasive urothelial bladder tumours. Eur Urol. 2007;52:760-768 Abstract, Full-text, PDF, Crossref.
-  M. Chen, D.A. Linkens. A systematic neurofuzzy modelling framework with application to material property prediction. IEEE Trans SMC Part B: Cybernetics. 2001;31:781-790
-  M.F. Abbod, J.W.F. Catto, M. Chen, D.A. Linkens, F.C. Hamdy. Artificial intelligence for the prediction of bladder cancer. Biomed Eng Appl Basis Comm. 2004;16:49-58 Crossref.
-  A.M. Molinaro, R. Simon, R.M. Pfeiffer. Prediction error estimation: a comparison of resampling methods. Bioinformatics. 2005;21:3301-3307 Crossref.
-  J.W.F. Catto, G. Xinarianos, J.L. Burton, M. Meuth, F.C. Hamdy. Differential expression of hMLH1 and hMSH2 is related to bladder cancer grade, stage and prognosis, but not microsatellite instability. Int J Cancer. 2003;105:484-490 Crossref.
-  C.B. Begg, L.D. Cramer, E.S. Venkatraman, J. Rosai. Comparing tumour staging and grading systems: a case study and a review of the issues, using thymoma as a model. Stat Med. 2000;19:1997-2014
-  J.W.F. Catto, D.A. Linkens, M.F. Abbod, et al. Artificial intelligence in predicting bladder cancer outcome: a comparison of neuro-fuzzy modeling and artificial neural networks. Clin Cancer Res. 2003;9:4172-4177
-  E.A. Reedy, B.M. Heatfield, B.F. Trump, J.H. Resau. Correlation of cytokeratin patterns with histopathology during neoplastic progression in the rat urinary bladder. Pathobiology. 1990;58:15-27 Crossref.
-  K. Yamana, V. Bilim, N. Hara, et al. Prognostic impact of FAS/CD95/APO-1 in urothelial cancers: decreased expression of Fas is associated with disease progression. Br J Cancer. 2005;93:544-551 Crossref.
-  K.M. Rieger-Christ, L. Ng, R.S. Hanley, et al. Restoration of plakoglobin expression in bladder carcinoma cell lines suppresses cell migration and tumorigenic potential. Br J Cancer. 2005;92:2153-2159 Crossref.
-  M.G. Friedrich, S. Chandrasoma, K.D. Siegmund, et al. Prognostic relevance of methylation markers in patients with non-muscle invasive bladder carcinoma. Eur J Cancer. 2005;41:2769-2778 Crossref.
-  J. Bektic, K. Pfeil, A.P. Berger, et al. Small G-protein RhoE is underexpressed in prostate cancer and induces cell cycle arrest and apoptosis. Prostate. 2005;64:332-340 Crossref.
-  M. Bogliolo, R.M. Taylor, K.W. Caldecott, G. Frosina. Reduced ligation during DNA base excision repair supported by BRCA2 mutant cells. Oncogene. 2000;19:5781-5787 Crossref.
a Academic Urology Unit, University of Sheffield, Sheffield, United Kingdom
b School of Engineering and Design, Brunel University, West London, United Kingdom
c Institute for Surgical Pathology, University Hospital Zurich, Zurich, Switzerland
d Department of Automatic Control and Systems Engineering, University of Sheffield, Sheffield, United Kingdom
e Department of Surgery, University of Dresden, Dresden, Germany
f Department of Urology, University of Regensburg, Regensburg, Germany
g Department for Pathology, University of Erlangen, Erlangen, Germany
h Institute of Pathology, University of Aachen, Aachen, Germany
i Nuffield Department of Surgery, University of Oxford, Oxford, United Kingdom
© 2009 European Association of Urology, Published by Elsevier B.V.
Recommend this article
Currently this article has a rating of 1. Please log in to recommend it.