Int. J. Cancer: 111, (2004) 2004 Wiley-Liss, Inc.

Transcription

1 Int. J. Cancer: 111, (2004) 2004 Wiley-Liss, Inc. Publication of the International Union Against Cancer ALGORITHMS BASED ON PROSTATE-SPECIFIC ANTIGEN (PSA), FREE PSA, DIGITAL RECTAL EXAMINATION AND PROSTATE VOLUME REDUCE FALSE-POSTITIVE PSA RESULTS IN PROSTATE CANCER SCREENING Patrik FINNE 1,2 *, Ralf FINNE 3, Chris BANGMA 4, Jonas HUGOSSON 5, Matti HAKAMA 2,6, Anssi AUVINEN 2 and Ulf-Håkan STENMAN 1 1 Department of Clinical Chemistry, University of Helsinki, Helsinki, Finland 2 School of Public Health, University of Tampere, Tampere, Finland 3 Swedish Polytechnic, Vaasa, Finland 4 Department of Urology, Erasmus Medical Center, Rotterdam, The Netherlands 5 Department of Urology, Göteborg University, Göteborg, Sweden 6 Finnish Cancer Registry, Helsinki, Finland Our objective was to determine whether multivariate algorithms based on serum total PSA, the free proportion of PSA, age, digital rectal examination and prostate volume can reduce the rate of false-positive PSA results in prostate screening more effectively than the proportion of free PSA alone at 95% sensitivity. A total of 1,775 consecutive 55- to 67-year-old men with a serum PSA of 4 10 g/l in the European Randomized Study of Screening for Prostate Cancer were included. To predict the presence of, multivariate algorithms were constructed using logistic regression (LR) and a multilayer perceptron neural network with Bayesian regularization (BR-MLP). A prospective setting was simulated by dividing the data set chronologically into one set for training and validation (67%, n 1,183) and one test set (33%, n 592). The diagnostic models were calibrated using the training set to obtain 95% sensitivity. When applied to the test set, the LR model, the BR-MLP model and the proportion of free PSA reached 92%, 87% and 94% sensitivity and reduced 29%, 36% and 22% of the false-positive PSA results, respectively. At a fixed sensitivity of 95% in the test set, the LR model eliminated more false-positive PSA results (22%) than the proportion of free PSA alone (17%) (p < 0.001), whereas the BR-MLP model did not (19%) (p 0.178). The area under the ROC curve was larger for the LR model (0.764, p 0.030) and the BR-MLP model (0.760, p 0.049) than for the proportion of free PSA (0.718). A multivariate algorithm can be used to reduce unnecessary prostate biopsies in screening more effectively than the proportion of free PSA alone, but the algorithms will require updating when clinical practice develops with time Wiley-Liss, Inc. Key words: prostate ; screening; logistic regression; neural network; prostate-specific antigen Prostate is the with the highest incidence among men in most industrialized countries. 1 PSA is increasingly used for early detection and screening, even though it has not been shown to reduce prostate mortality and may lead to overdiagnosis of the disease. A principal problem with prostate screening is that only approximately a third of men with elevated serum PSA ( 4 g/l) have in prostate biopsies. Most false-positive PSA results are caused by benign prostatic hyperplasia (BPH). False-positive results cause unnecessary prostate biopsies, generating anxiety, discomfort and costs. 2,3 Various approaches have been suggested to identify, among screening positive men, a low-risk group that could be spared unnecessary prostate biopsies. The proportion of free PSA in serum is lower in men with prostate than in those with BPH, 4 7 and determination of the proportion of free PSA has been shown to reduce 30% of false-positive PSA results at 90% sensitivity. 8,9 Prostate volume is usually smaller in prostate cases than in other men with elevated serum PSA, and the ratio between total PSA and prostate volume (PSA density) can also be used to reduce false-positive PSA results. 9 Logistic regression (LR) and neural networks have been applied to produce diagnostic algorithms based mainly on total PSA, free PSA, prostate volume and digital rectal examination (DRE). Combined use of several variables has been shown to reduce the number of false-positive PSA results more efficiently than single use of the proportion of free PSA However, LR or neural network algorithms for prostate detection have not been studied prospectively; i.e., use of algorithm outcomes for making real-life biopsy decisions has not been reported. We simulated a prospective setting by using earlier subjects as training data and later ones for testing of the algorithms. In most earlier studies, randomization was used to divide the subjects into training and test groups, thus ensuring equal distributions of all variables in both groups. Our study design is more demanding for the algorithms as the test set is likely to differ more from the training set. Consequently, the potential of the diagnostic algorithms can be evaluated more realistically. MATERIAL AND METHODS Subjects were identified within the Finnish, Dutch and Swedish sections of the European Randomized Study of Screening for Prostate Cancer In , 32,688 men were screened using serum PSA as an indication for prostate biopsies. Of these, 2,274 (7.0%) had a serum PSA of 4 10 g/l in the first screening round. The following inclusion criteria were met by 1,775 of the men: (i) age years, (ii) DRE performed, (iii) prostate volume measured and (iv) prostate biopsy performed. Initial age ranges of screened men were 55 67, and years in Finland, the Netherlands and Sweden, respectively. The reason for including only 55- to 67-year-old men was to obtain a homogenous age range. The numbers of subjects in Finland, the Netherlands, and Sweden were 1,308, 194 and 273, with detection rates (proportions of subjects with prostate ) of 0.23, 0.23 and 0.21, respectively. Serum samples were kept frozen at 80 C after sampling and thawed for analysis within 1 2 weeks. Total and free PSA were Grant sponsor: Academy of Finland; Grant sponsor: Cancer Society of Finland; Grant sponsor: Europe Against Cancer; Grant sponsor: Beckmann-Hybritech Corporation; Grant sponsor: Perkin-Elmer-Wallac; Grant sponsor: Finska Läkaresällskapet; Grant sponsor: Tampere University Hospital Research Fund. *Correspondence to: Biomedicum-Helsinki A424b, Department of Clinical Chemistry, University of Helsinki, P.O. Box 700, FIN HUS, Finland. Fax: patrik.finne@hus.fi Received 3 July 2003; Revised 26 November 2003, 22 January 2004; Accepted 27 January 2004 DOI /ijc Published online 12 April 2004 in Wiley InterScience ( wiley.com).

2 ALGORITHMS REDUCE FALSE-POSITIVE PSA RESULTS 311 determined by a simultaneous dual-label immunofluorimetric assay (Delfia Prostatus PSA; EG&G-Wallac, Turku, Finland). Prostate volume was determined by transrectal ultrasound using the ellipsoid formula: height width length Ultrasound-guided sextant biopsies were routinely performed, and additional biopsy cores were taken from hypoechoic foci. A DRE finding not suspicious for was defined as negative and a finding suspicious for, as positive. Statistical methods LR. LR models were fitted to predict the probability of detecting prostate in biopsies. Total PSA, the proportion of free PSA, DRE result (0 for nonsuspicious, 1 for suspicious), prostate volume and age were used as explanatory variables. Various transformations (logarithmic, square root, quadratic, quartiles) of the continuous variables were evaluated to decrease the deviance of the equation, to achieve favorable Hosmer-Lemeshow statistics and to reach maximal average reduction of false-positive PSA results at 95% sensitivity in cross-validation within the training set. 23 When including total PSA as a categorical variable (quartiles), the models were more accurate than with the continuous transformations of total PSA. Because of this, other continuous transformations were evaluated and the one producing the highest FIGURE 1 The process of training, validating and testing the diagnostic models. accuracy (highest specificity at 95% sensitivity in cross-validation) was exp[(psa 6.5) 2]/{exp[(PSA 6.5) 2] 1}. No first- (interaction between 2 variables) or higher- (interaction between 3 or more variables) order interactions were observed between any of the explanatory variables. Artificial neural network. The artificial neural network used was a multilayer perceptron with Bayesian regularization (BR-MLP). 15 Preprocessed values of total PSA, the proportion of free PSA, prostate volume, DRE and age were used as input variables. Various models with 2 5 neurons in one hidden layer and one output neuron, were evaluated. The effect of various preprocessing procedures (transformations, normalization to mean 0, SD 1, orthogonalization by principal components analysis) was assessed. For each BR-MLP training, 30 models were constructed with random initial values of all parameters. The model reducing the most false-positive PSA results at 95% sensitivity in training was selected for cross-validation within the training set. Bayesian regularization was used to avoid overtraining and improve generalization. 24,25 In addition to minimizing the sum of squares of errors, this optimization method minimizes the number of model parameters and the sum of squares of parameter values. BR-MLP models were constructed with the MATLAB Neural Network Toolbox, version 3 (Mathworks, Natick, MA). Training, validation and testing of diagnostic models. To simulate a prospective setting, the data set was divided chronologically on the basis of PSA sampling date into a training set with the first 1,183 (2/3) men having a PSA test and a test set with the 592 subsequent men (Fig. 1). The division was made so that similar proportions of training and test data were maintained for each country. The training set was randomly divided into 12 validation sets. Various LR and BR-MLP model structures were evaluated within the training set using 12-fold cross-validation. The model structure (one LR and one BR-MLP) giving the highest mean reduction of false-positive PSA results at 95% sensitivity was selected for final training on the whole training set. The final trained model was tested on the test set. In cross-validation, age did not improve accuracy and was therefore left out from final training and testing. The BR-MLP model structure selected for final training and testing included total PSA concentration, proportion of free PSA, prostate volume and DRE. Total PSA and prostate volume were logarithmically transformed, and all variables were normalized and orthogonalized (orthogonalization is a method which transforms the explanatory variables so that they will be uncorrelated). The model included 3 hidden neurons. In the final training, 100 models with random initial parameter values were constructed, and the one reducing the most false-positive PSA results at 95% sensitivity in the training set was selected for testing on the test set. The final trained LR model is shown in Table I. The mathematical formula of the final BR-MLP model is available on the internet, where the algorithms also can be downloaded as Excel applications ( Outcome measures. Prior to testing, it was decided that the main outcome measure would be proportion of reduced false-positive PSA results at 95% sensitivity. Cut-off values producing 95% sensitivity in the training set were used to calculate sensitivities and proportions of reduced false-positive PSA results in the test set. Sensitivity was defined as the proportion of men with TABLE I LR EQUATION BASED ON THE TRAINING SET OF 1,183 SUBJECTS IN THE FINNISH, DUTCH AND SWEDISH CENTERS OF THE EUROPEAN RANDOMIZED STUDY OF SCREENING FOR PROSTATE CANCER Variables Coefficient SE Wald Significance OR (95% CI) 1 PSA ( ) Proportion of free PSA (%) ( ) ln(prostate volume) ( ) DRE ( / ) ( ) Intercept Odds ratios (and 95% confidence intervals) of total PSA and prostate volume represent a 1-unit increase on the transformed scale, i.e., a PSA increase from 4 to 10 q/l. 2 PSA was transformed as follows: exp [(PSA 6.5) 2]/{exp[(PSA 6.5) 2] 1}. 3 In, natural logarithm.

3 312 FINNE ET AL. being correctly diagnosed by the algorithms. Binomial confidence intervals were calculated for sensitivities and proportions of reduced false-positive PSA results. The abilities of the diagnostic models and the proportion of free PSA to reduce false-positive PSA results at 95% sensitivity were compared using the McNemar test. For receiver-operating characteristic (ROC) analysis, continuous model outputs were used. The area under the ROC curve was calculated according to the method of Hanley and McNeil. 26 Differences in the distributions of diagnostic variables between groups were assessed using the Mann-Whitney U-test for continuous variables and the 2 test for DRE and biopsy results. Twotailed p 0.05 was considered statistically significant. RESULTS Accuracy of diagnostic models in the test set When the diagnostic models were calibrated using the training set to obtain 95% sensitivity, the LR model, the BR-MLP model and the proportion of free PSA reached actual sensitivities of 92%, 87% and 94%, reducing 29%, 36% and 22% of the false-positive PSA results, respectively, in the test set (Table II). To compare proportions of false-positive PSA results, cut-off values were chosen to produce 95% sensitivity in the test set: the LR model reduced more false-positive PSA results (22%, p 0.001) than the proportion of free PSA alone (17%), whereas the BR-MLP model did not (19%, p 0.178). The areas under the ROC curves of the LR model (0.764, p 0.030) and the BR-MLP model (0.760, p 0.049) were larger than that of the proportion of free PSA (0.718) (Fig. 2, Table II). LR and BR-MLP models including only total PSA and the proportion of free PSA were also evaluated, but the diagnostic performance of these did not exceed that of the proportion of free PSA alone. The areas under the ROC curves of the LR and BR-MLP models tended to be larger among Dutch than among Finnish and Swedish patients in the test set, although the differences were not statistically significant (p and p 0.077, respectively) (Table II). The difference may be explained by the fact that, among men without prostate in biopsy, Dutch men had larger prostate volume than other men (median 50 vs. 42 ml, p 0.001). This improves the specificity (reduction of false-positive PSA results) of algorithms that are based on prostate volume. Accuracy of cross-validation within training set and differences between training and test sets When aiming at 95% sensitivity in cross-validation within the training set, the LR model, the BR-MLP model and the proportion of free PSA reached actual sensitivities of 95%, 94% and 96%, reducing 29%, 34% and 22% of the false-positive PSA results, respectively. Thus, the accuracy of the diagnostic models was considerably higher in cross-validation within the training set than when testing on the prospective test set. Therefore, we compared the distributions of all model variables between the training and test sets. The proportion of prostate cases was larger in the test group (0.260) than in the training group (0.214, p 0.029), but there were no differences in the distributions of serum PSA, the proportion of free PSA, prostate volume, suspicious DRE findings and age (p 0.1 for all comparisons). Furthermore, the distributions of serum PSA, the proportion of free PSA, prostate volume and the proportion of suspicious DRE findings differed between men with and without prostate in both the training and test FIGURE 2 ROC curves showing the diagnostic performance of the proportion of free PSA (F/T) and the LR and BR-MLP models in the test set (n 592). TABLE II PERFORMANCE OF THE DIAGNOSTIC MODELS IN THE TEST SET Sensitivity (95% CI) 1 Reduction of false-positive PSA results (95% CI) 1 AUC (95% CI)2 Finland (n 436) Proportion of free PSA 93 (86 97) 23 (19 28) ( ) LR model 91 (84 96) 27 (22 32) ( ) BR-MLP model 87 (80 93) 34 (29 39) ( ) The Netherlands (n 65) Proportion of free PSA 100 (84 100) 18 (8 33) ( ) LR model 100 (84 100) 36 (22 52) ( ) BR-MLP model 95 (76 100) 50 (35 65) ( ) Sweden (n 91) Proportion of free PSA 91 (71 99) 16 (8 27) ( ) LR model 91 (71 99) 32 (21 44) ( ) BR-MLP model 77 (55 92) 36 (25 49) ( ) Entire test set (n 592) Proportion of free PSA 94 (88 97) 22 (18 26) ( ) LR model 92 (87 96) 29 (24 33) ( ) BR-MLP model 87 (81 92) 36 (31 41) ( ) 1 Sensitivities and proportions of reduced PSA results (percentage) achieved when aiming at 95% sensitivity by calibrating the models using the training set. CI, confidence interval. 2 AUC, area under the curve.

4 ALGORITHMS REDUCE FALSE-POSITIVE PSA RESULTS 313 TABLE III DISTRIBUTION OF THE EXPLANATORY VARIABLES IN TRAINING AND TEST SETS Variable Prostate Training set No prostate Prostate Test set No prostate Serum PSA ( g/l, median) Proportion of free PSA (%, mean) Prostate volume (ml, median) Suspicious DRE findings (%) Age (years, mean) Number TABLE IV DISTRIBUTION OF T STAGE AND GLEASON SCORE AMONG PROSTATE CANCER PATIENTS IN THE TEST SET BEING DIAGNOSED CORRECTLY OR INCORRECTLY BY THE MODELS set. Proportion of free PSA LR model BR-MLP model Negative 1 Positive 1 Negative Positive Negative Positive T stage T T T T4 0 Unknown Gleason score Unknown Total Test negative (incorrect) or positive (correct) using a cut-off that produces 95% sensitivity in the test Total sets (p 0.001) (Table III). The age distribution did not differ between men with and without prostate. Clinical stage and histopathologic grade To evaluate the clinical characteristics of the prostate s missed using the LR and BR-MLP models, clinical stage and histopathologic grade were assessed. All missed patients had clinically localized s with a Gleason score of 6 or lower (Table IV). Figure 3 shows the distribution of prostate probabilities calculated by LR grouped according to stage and Gleason score. The probability was higher among men with a Gleason score of 7 10 than among those with a Gleason score of 2 6 (p 0.001). Men with stage T2/3 had a higher calculated probability than those with stage T1 (p 0.001). A possible way of further reducing the number of biopsies would be to avoid diagnosing s with low grade and stage. We calculated cut-off values of the diagnostic models based on 95% sensitivity for cases with a Gleason score of 7 or higher or a T stage of 2 or higher (n 73). This resulted in a reduction of 33% of the false-positive PSA results for the LR model, 30% for the BR-MLP model and 21% for the proportion of free PSA. DISCUSSION Earlier studies showed that algorithms based on total PSA, free PSA, DRE and prostate volume can be used to reduce the number of false-positive PSA results in screening for prostate more efficiently than the proportion of free PSA alone Despite this, there are no reports on the use of such algorithms in a prospective setting. Our study included 1,775 screen-positive men from the 3 largest participating centers in the European Randomized Study of Screening for Prostate Cancer, making it the largest study on predicting prostate biopsy results in the PSA range 4 10 g/l using LR and neural networks. In most previous studies, all subjects were randomized into training, validation and test sets, thus ensuring similarity between FIGURE 3 Distributions of prostate probabilities calculated by LR are shown according to Gleason score and T stage. the sets. However, when a diagnostic model is applied in real life, the test data are always of more recent date than the training data, and this may cause differences in the distributions of diagnostic variables. In our study, positive biopsies became more frequent, indicating a learning curve with gained experience. Apparently as a result of this, our neural network model performed considerably worse on the test set than in cross-validation within the training set, in which the neural network was superior to LR. The LR model, which included fewer parameters (although both methods used the

5 314 FINNE ET AL. same 4 diagnostic variables) and therefore can be considered more robust, performed slightly better in the test set. We also evaluated a neural network model with only 2 hidden neurons (and thus fewer parameters), and it would have performed better on the test set than the model with 3 hidden neurons. However, the more complex model performed better in the cross-validation within the training set, and because this reflects a prospective setting, it was selected for final testing. We also assessed the performance of the LR and BR-MLP models using randomized division of the whole material into training and test sets. Both models performed better in the randomized than in the prospective test set (data not shown). These results show that diagnostic algorithms need to be evaluated prospectively. For prospective estimation of the sensitivity of the diagnostic models in the test set, cut-off values of the model outputs were chosen on the basis of 95% sensitivity in the training set. Surprisingly, the sensitivity of the BR-MLP model dropped to 87% and that of the LR model, to 92%. The corresponding sensitivity of the proportion of free PSA was virtually unchanged, 94%. This further supports the notion that simpler models are more robust. The loss of sensitivity (proportion of cases being correctly diagnosed) was apparently caused by slight differences in the distributions of the explanatory variables in patients between the training and test sets. Those in the test set had a somewhat lower proportion of suspicious DRE findings and slightly larger prostate volumes. Although these differences were not statistically significant, they caused a reduction in the risk calculated by LR and BR-MLP for some patients. The diagnostic models aim of 95% sensitivity, thus missing 5% of cases, was considered acceptable. All s missed by the LR and BR-MLP models and most of those missed by the proportion of free PSA were clinically localized and well-differentiated tumors, which may not need to be detected. 27 To further reduce the number of prostate biopsies, we calculated cut-off values for the algorithm outputs based on 95% sensitivity among patients with of high grade or stage. In this way, the ability of the LR model to reduce false-positive PSA results increased from 22% to 33%, that of the BR-MLP model from 19% to 30% and that of the proportion of free PSA from 17% to 21%. The outcome of the diagnostic models correlated with clinical stage and histopathologic grade of the prostate. This was expected as negative outcomes of the diagnostic models are associated with lower concentration of serum total PSA, higher proportion of free PSA, larger prostate volume and reduced risk of suspicious DRE finding. These results show that the algorithms can be used to identify high- and low-risk s. Reducing the number of unnecessary prostate biopsies is important. Although prostate biopsies are generally well tolerated, they cause at least mild complications in up to 60% of cases. 28,29 In the Finnish study on prostate screening, 69% considered the biopsy procedure unpleasant and 52% reported moderate pain. 3 Multivariate methods based on total PSA, proportion of free PSA, DRE finding and prostate volume enable reduction of the number of biopsies among men with a total PSA of 4 10 l in prostate screening by up to 20%, without sacrificing sensitivity. Yet, this requires information obtained from clinical examination, which reduces feasibility. In our analyses, LR performed slightly better than the artificial neural network. Prospective validation of the diagnostic models showed that the algorithms need to be updated when clinical practice develops with time. REFERENCES 1. Parkin DM, Pisani P, Ferlay J. Global statistics. CA Cancer J Clin 1999;49: Madalinska JB, Essink-Bot ML, de Koning HJ, Kirkels WJ, van der Maas PJ, Schroder FH. Health-related quality of life in patients with screen-detected versus clinically diagnosed prostate preceding primary treatment. Prostate 2001;46: Mäkinen T, Auvinen A, Hakama M, Stenman UH, Tammela TL. Acceptability and complications of prostate biopsy in populationbased PSA screening versus routine clinical practice: a prospective, controlled study. Urology 2002;60: Stenman U-H, Leinonen J, Alfthan H, Rannikko S, Tuhkanen K, Alfthan O. A complex between prostate-specific antigen and alpha 1-antichymotrypsin is the major form of prostate-specific antigen in serum of patients with prostatic : assay of the complex improves clinical sensitivity for. Cancer Res 1991;51: Lilja H, Christensson A, Dahlen U, Matikainen MT, Nilsson O, Pettersson K, Lövgren T. Prostate-specific antigen in serum occurs predominantly in complex with alpha 1-antichymotrypsin. Clin Chem 1991;37: Christensson A, Björk T, Nilsson O, Dahlen U, Matikainen MT, Cockett AT, Abrahamsson PA, Lilja H. Serum prostate specific antigen complexed to alpha 1-antichymotrypsin as an indicator of prostate. J Urol 1993;150: Leinonen J, Lövgren T, Vornanen T, Stenman U-H. Double-label time-resolved immunofluorometric assay of prostate-specific antigen and of its complex with alpha 1-antichymotrypsin. Clin Chem 1993; 39: Catalona WJ, Smith DS, Wolfert RL, Wang TJ, Rittenhouse HG, Ratliff TL, Nadler RB. Evaluation of percentage of free serum prostate-specific antigen to improve specificity of prostate screening. JAMA 1995;274: Bangma CH, Rietbergen JB, Kranse R, Blijenberg BG, Petterson K, Schröder FH. The free-to-total prostate specific antigen ratio improves the specificity of prostate specific antigen in screening for prostate in the general population. J Urol 1997;157: Bangma CH, Kranse R, Blijenberg BG, Schröder FH. The value of screening tests in the detection of prostate. Part II: Retrospective analysis of free/total prostate-specific analysis ratio, age-specific reference ranges, and PSA density. Urology 1995;46: Standaert B, Alwan A, Nelen V, Denis L. Prostate volume and in screening programs. Prostate 1997;33: Kranse R, Beemsterboer P, Rietbergen J, Habbema D, Hugosson J, Schröder FH. Predictors for biopsy outcome in the European Randomized Study of Screening for Prostate Cancer (Rotterdam region). Prostate 1999;39: Gomari M, Finne P, Järvi T, Stenman U-H, Hugosson J. Learning vector quantization, multilayer perceptron, neurofuzzy network and logistic regression in the diagnosis of prostate. In: Arabnia H, ed. Proceedings of the 1998 International Conference on Parallel and Distributed Processing Techniques and Applications, Las Vegas: CSREA Press Virtanen A, Gomari M, Kranse R, Stenman UH. Estimation of prostate probability by logistic regression: free and total prostatespecific antigen, digital rectal examination, and heredity are significant variables. Clin Chem 1999;45: Finne P, Finne R, Auvinen A, Juusela H, Aro J, Määttänen L, Hakama M, Rannikko S, Tammela LJ, Stenman U-H. Predicting the outcome of prostate biopsy in screen-positive men by a multilayer perceptron network. Urology 2000;56: Horninger W, Bartsch G, Snow PB, Brandt JM, Partin AW. The problem of cutoff levels in a screened population: appropriateness of informing screenees about their risk of having prostate carcinoma. Cancer 2001;91: Djavan B, Remzi M, Zlotta A, Seitz C, Snow P, Marberger M. Novel artificial neural network for early detection of prostate. J Clin Oncol 2002;20: Stephan C, Cammann H, Semjonow A, Diamandis EP, Wymenga LF, Lein M, Sinha P, Loening SA, Jung K. Multicenter evaluation of an artificial neural network to increase the prostate detection rate and reduce unnecessary biopsies. Clin Chem 2002;48: Stephan C, Jung K, Cammann H, Vogel B, Brux B, Kristiansen G, Rudolph B, Hauptmann S, Lein M, Schnorr D, Sinha P, Loening SA. An artificial neural network considerably improves the diagnostic power of percent free prostate-specific antigen in prostate diagnosis: results of a 5-year investigation. Int J Cancer 2002;99: Määttänen L, Auvinen A, Stenman UH, Tammela T, Rannikko S, Aro J, Juusela H, Hakama M. Three-year results of the Finnish prostate screening trial. J Natl Cancer Inst 2001;93: Schröder FH, Kranse R, Rietbergen J, Hoedemaeke R, Kirkels W. The European Randomized Study of Screening for Prostate Cancer (ER- SPC): an update. Members of the ERSPC, section Rotterdam. Eur Urol 1999;35: Becker C, Piironen T, Pettersson K, Hugosson J, Lilja H. Clinical value of human glandular kallikrein 2 and free and total prostate-

6 ALGORITHMS REDUCE FALSE-POSITIVE PSA RESULTS 315 specific antigen in serum from a population of men with prostatespecific antigen levels 3.0 ng/ml or greater. Urology 2000;55: Hosmer DW, Lemeshow S. Applied logistic regression. New York: John Wiley & Sons, MacKay DJC. Bayesian interpolation. Neural Comput 1992;4: Foresee FD, Hagan MT. Gauss-Newton approximation to bayesian regularization. In: Proceedings of the 1997 International Joint Conference on Neural Networks 1997: Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 1982;143: Chodak GW, Thisted RA, Gerber GS, Johansson JE, Adolfsson J, Jones GW, Chisholm GD, Moskovitz B, Livne PM, Warner J. Results of conservative management of clinically localized prostate. N Engl J Med 1994;330: Rietbergen JB, Kruger AE, Kranse R, Schröder FH. Complications of transrectal ultrasound-guided systematic sextant biopsies of the prostate: evaluation of complication rates and risk factors within a population-based screening program. Urology 1997;49: Djavan B, Waldert M, Zlotta A, Dobronski P, Seitz C, Remzi M, Borkowski A, Schulman C, Marberger M. Safety and morbidity of first and repeat transrectal ultrasound guided prostate needle biopsies: results of a prospective European prostate detection study. J Urol 2001;166: