© 2004 Foundation for Promotion of Cancer Research
The Use of Artificial Neural Network Analysis to Improve the Predictive Accuracy of Prostate Biopsy in the Japanese Population
1 Department of Urology, Kurashiki Central Hospital, Kurashiki, Okayama, Japan and 2 Department of Urology, Tohoku University, School of Medicine, Sendai
For reprints and all correspondence: Yoshiyuki Matsui, Department of Urology, Faculty of Medicine, Kyoto University, Shogoin Kawahara-cho 54, Sakyo-ku, Kyoto 606-8507, Japan. E-mail: ym1108{at}kuhp.kyoto-u.ac.jp
Received May 20, 2004; accepted August 9, 2004
 |
Abstract |
|---|
 TOP Abstract INTRODUCTIONSUBJECTS AND METHODSRESULTSDISCUSSIONReferences |
|---|
Objective: We examined the efficacy of an artificial neural network analysis (ANNA) based on parameters available from previously existing examinations for improving the predictive accuracy of prostate cancer screening in the Japanese population.
Methods: Two hundred and twenty-eight patients with prostate-specific antigen (PSA) of 2–10 ng/ml were enrolled in this study. Two artificial neural network analysis (ANNA) models were constructed: ANNA1 with patient age, total PSA, free to total PSA ratio, prostate volume, transition zone volume (TZ), PSA density (PSAD) and PSA-TZ density (PSATZ) as input variables, and ANNA2 with presumed circle area ratio (PCAR), digital rectal examination (DRE) findings and chief complaint added as variables. The predictive accuracies of the ANNA models were compared with conventional PSA and volume-related parameters and a logistic regression (LR) model by receiver operating characteristic (ROC) curve analysis.
Results: Of 228 patients, 58 (25.5%) were diagnosed with prostate cancer. While ANNA2 had a slightly larger area under the curve (AUC) than ANNA1 (0.782 versus 0.793, P = 0.8477), the AUC of ANNA2 was significantly greater than those of ln(PSA), PSAD, PSATZ and free to total PSA ratio (P = 0.0004, 0.0230, 0.0304, and 0.0037, respectively). The accuracy of ANNA2 was significantly better than that of LR analysis at 90 and 95% sensitivity levels (P = 0.0051 and P < 0.0001, respectively). At 95% sensitivity level, ANNA2 reduced unnecessary biopsies by 40.0% with a negative predictive value of 95.7%.
Conclusions: To determine the indication of prostate biopsy for PSA value in the range of 2–10 ng/ml, the ANNA model has the possibility to reduce unnecessary biopsies without missing many cases of cancers.
Key Words: prostate cancer • artificial neural network analysis • prostate biopsy • logistic regression analysis
|
INTRODUCTION |
|---|
|
TOPAbstractINTRODUCTIONSUBJECTS AND METHODSRESULTSDISCUSSIONReferences |
|---|
Measurement of prostate-specific antigen (PSA) is the most valuable means of early detection of prostate cancer (1), but its specificity as a screening test is relatively low, especially for PSA values <10 ng/ml (2). To improve the specificity of the screening algorithm, use of several volume-related parameters, PSA velocity or doubling time, age-specific PSA and several molecular forms of PSA has been proposed (3–6). Complexed PSA, human glandular kallikrein 2 and subfractions of free PSA such as pro-PSA, benign PSA and intact PSA are now considered as candidates for new tumor markers (7–10). Several authors have also reported the effectiveness of using the ratio of free to total PSA and PSA–
1-antichymotrypsin (ACT) in the Japanese population (11,12), while Okihara et al. suggested that a complex PSA did not provide any additional value in differentiating cancer from non-cancer cases in men with a total PSA between 4.0 and 10.0 ng/ml (13). However, these trials involved a small number of patients. These new biomarkers have not yet become widely used in the Japanese population, because their effectiveness has not been validated using a large sample size of the Japanese population and, in addition, most of the tests are not covered by national health insurance. We consider that, as prostate cancer has become more prevalent in Japan, the necessity of establishing an original predictive method for our population has increased. We previously reported on the effectiveness of an artificial neural network analysis (ANNA) for predicting the pathological stage of clinically localized prostate cancer (14), but there has been no report examining the clinical utility of ANNA in the diagnosis of prostate cancer in the Japanese population. One advantage of regression models is that the contribution of individual variables can be quantified easily, but a neural network (using the computing power of today's PCs) can easily handle non-linear phenomena of any type without requiring linear relationships that reflect simple correlations. Furthermore, as neural network models are used more frequently in clinical practice, more data will be accumulated that can be used constantly to train the ANNA model algorithms. In this study, we examined whether an ANNA model, which is a more complex decision-making algorithm based on a combination of clinical parameters available from previously existing examinations, can improve the predictive accuracy of prostate biopsy in the Japanese population.
|
SUBJECTS AND METHODS |
|---|
|
TOPAbstractINTRODUCTIONSUBJECTS AND METHODSRESULTSDISCUSSIONReferences |
|---|
From August 2000 to December 2002, 340 men referred for lower urinary tract symptoms or for early detection of prostate cancer underwent transrectal ultrasound (TRUS)-guided biopsies at Kurashiki Central Hospital. In all, 228 patients who had a PSA of 2–10 ng/ml were enrolled in this study.
Serum samples collected before any prostate manipulation were centrifuged within 2–3 h after venipuncture, and serum total PSA (tPSA) and free PSA (fPSA) levels were measured with the Architect total PSA and free PSA assay (Abbott Laboratories, Abbott Park, IL) on the same day, in order to prevent deterioration of fPSA stability.
Digital rectal examiniaiton (DRE) findings were classified by examining urologists as normal, abnormal but not strongly suggestive of prostate cancer (rule out prostate cancer), or strongly suggestive of prostate cancer (suspected prostate cancer).
All TRUS examinations were performed by one well-trained ultrasound technician. The total prostate volume (PV) and transition zone volume (TZV) were measured using the prolate ellipsoid formula. PSA density (PSAD) and PSA transition zone density (PSATZ) were calculated by dividing serum tPSA by PV or TZV, respectively. Presumed circle area ratio (PCAR), a parameter of benign prostatic hyperplasia, was calculated according to the formula as previously suggested (15).
In all patients, standard sextant biopsy plus an additional four biopsies from a more lateral portion were performed as a TRUS-guided prostate biopsy (16). For repeat biopsy, two additional transition zone biopsies were performed each time. Histopathological findings of biopsy specimens were classified as benign or positive for prostate cancer.
STATISTICAL METHODS
The non-parametric Mann–Whitney U-test was used for comparison between each group with or without prostate cancer on biopsy to compare the distributions of explanatory variables including tPSA, proportion of fPSA (f/T ratio), PV, TZV, PSAD, PSATZ, DRE, PCAR, age and chief complaint responsible for referral.
The results of several statistical analyses were compared with those of conventional PSA and volume-related parameters in this study to determine the most effective method for predicting the presence of prostate cancer.
ARTIFICIAL NEURAL NETWORK ANALYSIS (ANNA)
The neural network used in this application was the Bayesian neural tool of SPSS (Statistical Package for the Social Sciences, SPSS Inc., Chicago, IL) Neural Connection 2.1 software. The Bayesian neural tool used was a modified multilayer perceptron (MLP), which has a standard feed-forward topology and successive layers of adaptive weight.
ANNA models were inspired by the structure and function of biological neurons and are computational methodologies that perform multifactorial analyses. In biology, nervous system networks are composed of a large number of neuron cells that are extensively interconnected to each other. They interact with other neurons through a complicated web of branches and finally produce an output signal. As a neural network, artificial neural network models also contain layers of simple points (nodes) of data that interact through carefully weighted connections to produce an outcome. In an artificial neuron, the input nodes (diagnostic variables) are multiplied by optimized weight values and summed together. As with regression models, a bias weight (intercept or constant in statistical terminology) is used to optimize the prediction. The sum is transformed by a non-linear activation function into a desired range, usually a value between 0 and 1, which describes the probability of a certain outcome. A limitation of regression models and single-neuron networks is that they cannot detect complex relationships in a data set. Using hidden nodes (layers), MLP can combine several neurons to detect more complex functions and interactions between input variables. The number of hidden layers and the number of neurons in each layer are unrestricted, but typically one or two hidden layers are used. The output nodes (layers) produce the network response. In this study, the output node produced a binary value (0 = no malignancy, 1 = prostate cancer). The parameter and bias weights are obtained by training the neural network model on data with known values for the output. During the training process, the model parameters are chosen so that the output is as close to the correct values as possible by minimizing the prediction errors. Each weight and bias was modified to converge toward values representing a solution to the prediction problem. The weights were constantly updated to reflect this gradual convergence and to contribute further to an overall reduction in the root mean square error.
The disadvantage of MLP had been its tendency to overfit without a large validation data set. However, the Bayesian neural tool could automatically decrease overfitting and produce a generalized model even with a limited data set.
To estimate suitable initial weights for the ANNA model, 227 (228 – 1) patients were selected randomly and divided into two groups. The model was initially fit using 70% of the patients (training set) and validated on the remaining 30% (validation set). After five sessions of randomization, fitting and testing, the weights producing the smallest sum of squared errors on the corresponding validation set were selected as initial weights for the ANNA model used on the finally adapted training set.
We constructed one ANNA model (ANNA1) using age and conventional PSA and volume-related parameters (tPSA, f/T ratio, PV, TZV, PSAD and PSATZ) as input variables, and in another ANNA model (ANNA2), PCAR, DRE finding and chief complaint (lower urinary symptoms or result of PSA screening for early detection of prostate cancer) were added as variables.
LOGISTIC REGRESSION (LR) ANALYSIS
For comparison with the ANNA model, prostate cancer risks were estimated wih every explanatory variable using logistic regression (LR) analysis with the likelihood ratio test. The binary response variable in the LR model was the presence or absence of prostate cancer on biopsy. Logarithmic transformation of tPSA was used to decrease deviance and obtain a better fit with the LR model.
The generalizability of the LR and ANNA models was tested by ‘leave-one-out cross-validation’, dividing the patients into a training (training + validation) set with all patients but one, who formed the test set (17). Permutations were then made so that every patient in turn formed the test set. The benefit of this validation technique is the large training sets and therefore better model obtained, as suggested by Finne et al. (18).
To obtain the desired sensitivity level among the tested subjects, a cut-off of the output value was determined on the basis of the respective sensitivity level in the training set. For comparison for predictive accuracy, the area under the receiver operating characteristic (ROC) curve was used. Areas under the curve (AUCs) were calculated and compared with the McNemar test, as modified by Bonferoni–Holm.
The limit of significance for all tests was P < 0.05.
|
RESULTS |
|---|
|
TOPAbstractINTRODUCTIONSUBJECTS AND METHODSRESULTSDISCUSSIONReferences |
|---|
Patient characteristics are shown in Table 1. Of 228 patients, 58 (25.5%) were diagnosed with prostate cancer on biopsy. Although PCAR, chief complaint and DRE findings did not differ significantly between patients with and without prostate cancer, age distribution, tPSA, PV, TZV, PSAD, PSATZ and f/T ratio did (P = 0.0152, P = 0.0153, P = 0.0094, P = 0.0024, P < 0.0001, P < 0.0001 and P = 0.002, respectively). However, for multivariate logistic regression analysis, only patient age and PSATZ provided independent diagnostic information, and these two variables were used to construct the LR model (Table 2).
|
|
The cut-offs and specificities at 50, 90 and 95% sensitivity, as well as the AUC for each variable and results of statistical analysis are shown in Table 3. Compared with ANNA1, although the difference between models did not reach significance, ANNA2 had a larger AUC (0.782 versus 0.793, P = 0.8477). We therefore used ANNA2 as the final ANNA model for comparison with other variables. The AUC of ANNA reached 0.793 and was significantly greater than those of ln(PSA), PSAD, PSATZ and f/T ratio (AUCs of 0.607 with P = 0.0004, 0.673 with P = 0.0230, 0.671 with P = 0.0304 and 0.636 with P = 0.0037, respectively). Although the difference between them was not significant, the AUC of ANNA tended to be greater than that of LR analysis (0.793 versus 0.747, P = 0.3493) (Fig. 1). Furthermore, at 90 and 95% sensitivity levels, the accuracy of ANNA was significantly better than not only those of conventional PSA and volume-related parameters [ln(PSA), PSAD, PSATZ and f/T ratio] but also that of the LR analysis. At 90% sensitivity, the ANNA used in this study reduced unnecessary biopsies by 45.3% with a negative predictive value of 92.8% (six patients with prostate cancer would have been missed), while at 95% sensitivity, it reduced unnecessary biopsies by 40.0%, with a negative predictive value of 95.7% (three patients with prostate cancer would have been missed).
|
|
|
DISCUSSION |
|---|
|
TOPAbstractINTRODUCTIONSUBJECTS AND METHODSRESULTSDISCUSSIONReferences |
|---|
ANNA, a complex computational method designed to replicate human decision making by emulating human neuron connectivity, can be trained to recognize patterns derived from input variables and their associated outcomes and then to apply these patterns to new cases, unlike conventional statistical methods that process information step-by-step according to a given set of rules. Properly trained neural networks have been demonstrated repeatedly to have predictive accuracy superior to those of other predictive techniques, although some investigators have reported that their predictive power was not substantially superior to that of logistic regression models (19,23).
For screening of prostate cancer, ANNA exhibited better predictive accuracy than PSA-related parameters in European and American populations (24–27). In the study by Babarian et al., a neural network was developed using a retrospective data set for 151 men with PSA values from 2.5 to 4.0 ng/ml. The ANNA developed using input variables such as age, tPSA, prostatic acid phosphatase, creatinine kinase and f/T ratio exhibited better specificity at 92% sensitivity, although it is not significantly different from those for the single parameters (26). Finne et al. reported significantly better accuracy of ANNA using tPSA, f/T ratio, prostate volume and DRE findings as input variables than for a logistic regression model and standard PSA-related parameters. Their ANNA model eliminated 33% of false-positive biopsies at a 95% sensitivity level (27). Recently, Stephan et al. reported a multicenter study to evaluate the diagnostic value of f/T ratio-based ANNA in men with tPSA values from 2 to 20 ng/ml (24). They used tPSA, f/T ratio, prostate volume, DRE findings and patient age as input variables and concluded that use of this ANNA with PSA values from 2 to 10 ng/ml enhanced the specificity of thef/T ratio by 20–22%. Recently, Djavan developed ANNA models for patients with total PSA in the ranges of 2.5–4 ng/ml and 4–10 ng/ml (25). The f/T ratio, PSATZ, PSA velocity, free PSA, TZ volume, tPSA and PSAD were selected as input variables for the ANNA for PSA values from 4 to 10 ng/ml, and PSATZ, f/T ratio, PSAD and prostate volume for PSA values from 2.5 to 4 ng/ml. They reported that the predictive accuracy of the ANNA model was significantly superior to those of the single parameters, but the difference between the AUC of the ANNA and that of the logistic regression model did not reach significance for PSA values of 4–10 ng/ml. In our study, as several authors have already reported, the ANNA model yielded a significantly better outcome compared with any of the single parameters. It did not significantly improve the predictive accuracy compared with the LR model or the outcome of the ROC curve analysis, but it did improve accuracy at high sensitivity levels, though bias may be present due to the use of a different criterion for determining the predictive accuracy in fitting ANNA models compared with the criterion used to test statistical significance for conventional statistical models. The ANNA model eliminated 40% of false-positive PSA results with only 5% loss in sensitivity, suggesting that use of ANNA models may be beneficial to reduce patient risk and health care costs associated with performing biopsies.
Furthermore, although the advantage obtained was not large, adding more clinical variables such as chief complaint, PCAR on TRUS and DRE findings to age and conventional PSA and volume-related parameters tended to improve the predictive accuracy of ANNA in our models. Our final ANNA model could effectively utilize parameters that were not independent predictors to train its algorithm. We speculate that because these parameters closely correlate to the extent of benign prostatic hyperplasia (BPH), PSA increases due to BPH may be taken into account in the final ANNA model. ANNA models are generally more flexible and often have many parameters compared with conventional statistical models, for example logistic models. Our final ANNA model suggests the possibility that ANNA can be extended further not only by adding newer markers such as new molecular forms of PSA but also by using other clinical parameters already available.
ANNA has the advantage that, as ANNA models are used in clinical practice, more data will be generated that can be used constantly to update and improve the training of algorithms. We intend to perform a prospective trial of ANNA to assess the risk of prostate cancer in order to validate and confirm further the generalizability of our ANNA model in the Japanese population.
CONCLUSION
In determining whether prostate biopsy is indicated for PSA values in the range of 2–10 ng/ml, the use of an ANNA model is very beneficial in reducing unnecessary biopsies without missing many cases of cancer. Additional input parameters and further training of the algorithm in clinical practice may improve its performance further.
|
References |
|---|
|
TOPAbstractINTRODUCTIONSUBJECTS AND METHODSRESULTSDISCUSSIONReferences |
|---|
1 Polascik TJ, Oesterling JE, Partin AW. Prostate specific antigen: a decade of discovery—what we have learned and where we are going. J Urol 1999;162:293–306.[CrossRef][Web of Science][Medline]
2 Brawer MK, Cheli CD, Neaman IE, Goldblatt J, Smith C, Schwartz MK et al. Complexed prostate specific antigen provides significant enhancement of specificity compared with total prostate specific antigen for detecting prostate cancer. J Urol 2000;163:1476–80.[CrossRef][Web of Science][Medline]
3 Horninger W, Reissigl A, Klocker H, Rogatsch H, Fink K, Strasser H, Bartsch G. Improvement of specificity in PSA-based screening by using PSA-transition zone density and percent free PSA in addition to total PSA levels. Prostate 1998;37:133–7.[CrossRef][Web of Science][Medline]
4 Maeda H, Arai Y, Ishitoya S, Okubo K, Aoki Y, Okada T. Prostate specific antigen adjusted for the transition zone volume as an indicator of prostate cancer. J Urol 1997;158:2193–6.[CrossRef][Web of Science][Medline]
5 Catalona WJ, Southwick PC, Slawin KM, Partin AW, Brawer MK, Flanigan RC et al. Comparison of percent free PSA, PSA density, and age-specific PSA cutoffs for prostate cancer detection and staging. Urology 2000;56:255–260.[CrossRef][Web of Science][Medline]
6 Ciatto S, Bonardi R, Lombardi C, Zappa M, Gervasi G and Cappelli G. Analysis of PSA velocity in 1666 healthy subjects undergoing total PSA determination at two consecutive screening rounds. Int J Biol Markers 2002; 17:79–83.[Web of Science][Medline]
7 Djavan B, Remzi M, Zlotta AR, Ravery V, Hammerer P, Reissigl A et al. Complexed prostate-specific antigen, complexed prostate-specific antigen density of total and transition zone, complexed/total prostate-specific antigen ratio, free-to-total prostate-specific antigen ratio, density of total and transition zone prostate-specific antigen: results of the prospective multicenter European trial. Urology 2002;60 (Suppl 4A):4–9.[Web of Science][Medline]
8 Marks LS, Llanes AS, Linton HJ, Gasior CL, Millar LS, Mikolajczyk SD et al. BPSA is a potential serum marker for benign prostatic hyperplasia (BPH). J Urol 2001;165 (Suppl):266.
9 Nam RK, Diamandis EP, Toi A, Trachtenberg J, Magklara A, Scorilas A et al. Serum human glandular kallikrein-2 prostate levels predict the presence of prostate cancer among men with elevated prostate-specific antigen. J Clin Oncol 2000;18:1036–1042.
10 Khan MA, Partin AW, Rittenhouse HG, Mikolajczyk SD, Sokoll LJ, Chan DW et al. Evaluation of proprostate specific antigen for early detection of prostate cancer in men with a total prostate specific antigen range of 4.0 to 10.0 ng/ml. J Urol 2003;170:723–6.[CrossRef][Web of Science][Medline]
11 Egawa S, Soh S, Ohori M, Uchida T, Gohji K, Fujii A et al. The ratio of free to total serum prostate specific antigen and its use in differential diagnosis of prostate carcinoma in Japan. Cancer 1997;79:90–8.[CrossRef][Web of Science][Medline]
12 Kuriyama M, Abrahamsson PA, Imai K, Akimoto S, Deguchi N, Shichiri Y et al. Determination of serum prostate-specific antigen–alpha1-antichymotrypsin complex for diagnosis of prostate cancer in Japanese cases. Scand J Urol Nephrol 2001;35:5–10.[Web of Science][Medline]
13 Okihara K, Ukimura O, Nakamura T, Mizutani Y, Kawauchi A, Naya Y et al. Can complexed prostate specific antigen enhance prostate cancer detection in Japanese men? Eur Urol 2004;46:57–64.[CrossRef][Web of Science][Medline]
14 Matsui Y, Egawa S, Tsukayama C, Terai A, Kuwao S, Baba S et al. Artificial neural network analysis for predicting pathological stage of clinically localized prostate cancer in the Japanese population. Jpn J Clin Oncol 2002;32:530–5.
15 Watanabe H. New concept of BPH: PCAR theory. Prostate 1998;37:116–25.[CrossRef][Web of Science][Medline]
16 Bauer JJ, Zeng J, Zhang W, McLeod DG, Sesterhenn IA, Connelly RR et al. Lateral biopsies added to the traditional sextant prostate biopsy pattern increases the detection rate of prostate cancer. Prostate Cancer Prostatic Dis 2000;3:43–6.[CrossRef][Web of Science][Medline]
17 Moody J, Utans J. Architecture selection strategies for neural networks: application to corporate bond rating predictions. In: Refenes A-P, editor. Neural Networks in the Capital Markets, 2nd edn. Chichester, UK: John Wiley & Sons, 1995;287–9.
18 Finne P, Finne R, Stenman UH. Neural network analysis of clinicopathological factors in urological disease: a critical evaluation of available techniques. BJU Int 2001;88:825–31.[CrossRef][Web of Science][Medline]
19 Zernikow B, Holtmannspoetter K, Michel E, Theilhaber M, Pielemeier W, Hennecke KH. Artificial neural network for predicting intracranial haemorrhage in preterm neonates. Acta Paediatr 1998;87:969–75.[CrossRef][Web of Science][Medline]
20 Dybowski R, Weller P, Chang R and Gant V. Prediction of outcome in critically ill patients using artificial neural network synthesized by genetic algorithm. Lancet 1996;347:1146–50.[CrossRef][Web of Science][Medline]
21 Schwarzer G, Vach W and Schumacher M. On the misuse of artificial neural networks for prognostic and diagnostic classification in oncology. Stat Med 2000;19:541–61.[CrossRef][Web of Science][Medline]
22 Freeman RV, Eagle KA, Bates ER, Werns SW, Kline-Rogers E, Karavite D et al. Comparison of artificial neural networks with logistic regression in prediction of in-hospital death after percutaneous transluminal coronary angioplasty. Am Heart J 2000; 140:511–20.[CrossRef][Web of Science][Medline]
23 Duh MS, Walker AM, Pagano M, Kronlund K. Prediction and cross-validation of neural networks versus logistic regression: using hepatic disorders as an example. Am J Epidemiol 1998; 147: 407–13.
24 Stephan C, Cammann H, Semjonow A, Diamandis EP, Wymenga LF, Lein M et al. Multicenter evaluation of an artificial neural network to increase the prostate cancer detection rate and reduce unnecessary biopsies. Clin Chem 2002;48:1279–87.
25 Djavan B, Remzi M, Zlotta A, Seitz C, Snow P, Marberger M. Novel artificial neural network for early detection of prostate cancer. J Clin Oncol 2002;20:921–9.
26 Babaian RJ, Fritsche H, Ayala A, Bhadkamkar V, Johnston DA, Naccarato W et al. Performance of a neural network in detecting prostate cancer in the prostate-specific antigen reflex range of 2.5 to 4.0 ng/ml. Urology 2000;56:1000–6.[CrossRef][Web of Science][Medline]
27 Finne P, Finne R, Auvinen A, Juusela H, Aro J, Maattanen L et al. Predicting the outcome of prostate biopsy in screen-positive men by a multilayer perceptron network. Urology 2000;56:418–22.[CrossRef][Web of Science][Medline]
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  R. N. Dickerson, D. L. Mason, M. A. Croce, G. Minard, and R. O. Brown Evaluation of an Artificial Neural Network to Predict Urea Nitrogen Appearance for Critically Ill Multiple-Trauma Patients JPEN J Parenter Enteral Nutr, November 1, 2005; 29(6): 429 - 435. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||