|
Advertisement |
|
|
||
 |
|
|||||
|
|||||
|
 |
||||||
|
||||||
|
||||||
|
© 2002 American Society for Clinical Oncology Novel Artificial Neural Network for Early Detection of Prostate CancerByFrom the Department of Urology, University of Vienna, Austria; Department of Urology, Erasme University Clinics of Brussels, Belgium; and Xaim, Inc, Colorado Springs, CO. Address reprint requests to Bob Djavan, MD, PhD, Department of Urology, University of Vienna, Wahringer Gurtel 18-20, A-1090 Vienna, Austria; email: bdjavan{at}hotmail.com
PURPOSE: Two artificial neural networks (ANN) for the early detection of prostate cancer in men with total prostate-specific antigen (PSA) levels from 2.5 to 4 ng/mL and from 4 to 10 ng/mL were prospectively developed. The predictive accuracy of the ANN was compared with that obtained by use of conventional statistical analysis of standard PSA parameters. PATIENTS AND METHODS: Consecutive men with a serum total PSA level between 4 and 10 ng/mL (n = 974) and between 2.5 and 4 ng/mL (n = 272) were analyzed. A separate ANN model was developed for each group of patients. Analyses were performed to determine the presence of prostate cancer. RESULTS: The area under the receiver operator characteristic (ROC) curve (AUC) was 87.6% and 91.3% for the 2.5 to 4 ng/mL and 4 to 10 ng/mL ANN models, respectively. For the latter model, the AUC generated by the ANN was significantly higher than that produced by the single variables of total PSA, percentage of free PSA, PSA density of the transition zone (TZ), and TZ volume (P < .01), but not significantly higher compared with multivariate analysis. For the 2.5 to 4 ng/mL model, the AUC of the ANN ROC curve was significantly higher than the AUCs for percentage of free PSA (P = .0239), PSA-TZ (P = .0204), and PSA density and total prostate volume (P < .01 for both). CONCLUSION: The predictive accuracy of the ANN was superior to that of conventional PSA parameters. ANN models might change the way patients referred for early prostate cancer detection are counseled regarding the need for prostate biopsy.
PROSTATE CANCER IS the most commonly diagnosed cancer in men in many developed countries and is second only to lung cancer as a cause of cancer mortality in the United States. Risk factors for prostate cancer include age, familial history of cancer, and ethnicity. Measurement of prostate-specific antigen (PSA) levels is regarded widely as the most clinically useful tool for the early diagnosis of prostate cancer.1 However, elevated PSA levels are not unique to patients with prostate cancer because they can also occur in patients with benign prostatic disease (eg, prostatitis, benign prostatic hyperplasia [BPH], or prostatic infarct). In particular, substantial overlap of prostate cancer and BPH diagnoses occur in men with serum PSA levels in the diagnostic "gray zone" of 4 to 10 ng/mL.2,3 This lack of specificity can lead to unnecessary prognostic procedures in men with benign disease. Although prostate biopsy is required for a definitive diagnosis of prostate cancer, this invasive and costly procedure should be avoided in men with a low probability of having prostate cancer. An optimal method for early cancer detection will correctly identify men who should undergo tissue sampling (ie, those with a high probability of having prostate cancer). Various strategies have been investigated to improve the sensitivity and specificity of prostate cancer detection in patients with PSA levels between 4 and 10 mg/mL. These strategies include the ratio of the serum PSA level to the volume of the prostate gland (PSA density, or PSAD), the ratio of the serum PSA level to the volume of the prostatic transition zone (TZ; PSAD of the TZ), the rate of change of serum PSA levels over time (PSA velocity, or PSAV), evaluation of the PSA level with regard to patient age (age-adjusted PSA reference ranges), and the percentage of unbound (free) PSA to total PSA. Although the percentage of free PSA has emerged as the most clinically useful method of improving the sensitivity and specificity of prostate cancer detection,4-6 the value of these other methods is still the subject of considerable debate.7-11 Although most methods for early detection of prostate cancer focus on men with a serum total PSA level of 4 to 10 ng/mL, more than 20% of men with prostate cancer have a total PSA level below 4 ng/mL.3 Within 3 to 5 years, however, prostate cancer will develop in 13% to 20% of men with a total PSA level between 2.5 and 4 ng/mL.12-14 Because early detection of prostate cancer greatly improves patient survival and the potential for curative treatment, a detection method is needed that will identify the small number of men with a total PSA level between 2.5 and 4 ng/mL who should undergo a prostatic tissue sampling procedure. Although some studies have focused on the utility of various cancer detection methods for men with PSA levels below 4 ng/mL, more progress is needed to enhance the sensitivity and specificity of detection within this group of patients.5,15,16 Most studies on the early detection of prostate cancer focus on the need for an initial biopsy. After patients undergo this procedure, however, doctors must decide how to manage patients with negative biopsy results. In men with a negative initial biopsy result, prostate cancer will be detected in as few as 10% to 12%17,18 to as many as 16% to 30%19-21 of those who undergo a repeat biopsy. Thus, a method is clearly needed that not only will aid in the decision to undergo an initial tissue sampling procedure, but will also identify those patients who should undergo a second procedure when cancer is not detected the first time. Data from prostate cancer detection studies are typically analyzed by univariate and multivariate regression models to predict the presence of cancer. An innovative new model that uses an artificial neural network (ANN) has been developed to enhance the sensitivity and specificity of prostate cancer detection. This new method is a simulated modeling approach based loosely on the function of a biologic neuron and its relationship to a neural network. The ANN is able to model complex biologic systems by recognizing relationships among input data that cannot always be discerned by conventional analysis. Because the ANN is designed for data input via the Internet, doctors—from those in small practices to university-based researchers—will have access to cutting-edge technology that can greatly enhance the process of prostate cancer detection. In addition, the ANN can be used for men with relatively low total PSA levels (ie, < 4 ng/mL); thus, prostate cancer could be diagnosed at an earlier stage than would be possible via conventional methods of analysis. The purpose of the study presented here was to prospectively develop an ANN to predict the presence of prostate cancer or benign prostatic tissue in men with serum total PSA levels from 2.5 to 4 ng/mL and from 4 to 10 ng/mL. We then compared the predictive accuracy of this ANN with that obtained by conventional univariate statistical analysis of total PSA, percentage of free PSA, PSAD, PSA-TZ, total prostate volume, TZ volume, and PSAV.
From January 1997 to January 2000, 1,246 white men aged 31 to 89 years were enrolled onto this prospective European Prostate Cancer Detection Study. This Vienna-based multicenter European referral database was used in this study. The study population consisted of consecutive referrals of men for lower urinary tract symptoms or for early detection of prostate cancer. Patients with a family history of prostate cancer were excluded from the study because—in the opinion of the members of European Prostate Cancer Detection Study—such patients needed a more thorough PSA screening and needed to undergo early tissue sampling and analysis.
We analyzed 1,246 consecutive men, 974 with a serum total PSA level between 4 and 10 ng/mL and 272 men with a total PSA level between 2.5 and 4 ng/mL. Serum total PSA and free PSA levels were measured from deep-frozen (-70°C) serum samples with the AxSYM Total PSA and AxSYM Free PSA assays (Abbott Laboratories, Abbott Park, IL; intraobserver coefficient of variation range: 1.2% to 2.4% for total PSA of 4 to 10 ng/mL and 0.8% to 1.5% for total PSA of 2.5 to 4 ng/mL). Both assays are in vitro microparticle enzyme immunoassays based on a dual monoclonal antibody design. Of patients with total PSA levels of 4 to 10 ng/mL and 2.5 to 4 ng/mL, 35% and 24% were diagnosed with prostate cancer, and 65% and 76% were diagnosed with benign prostatic tissue defined as normal prostate tissue, chronic inflamed tissue (prostatitis), or benign prostate hyperplasia. The presence of prostate cancer was determined in all patients by means of a transrectal ultrasound (TRUS)–guided sextant biopsy plus two additional TZ biopsies with a systematic repeat biopsy at 6 weeks if no cancer was found and if the total PSA level was The purpose of our analysis was to prospectively develop an ANN to predict the presence of prostate cancer or benign prostatic tissue in prospective validation patients and to compare this predictive accuracy with the accuracy obtained by use of cutoffs for total PSA, percentage of free PSA, PSAD, PSA-TZ, total prostate volume, TZ volume, and PSAV. Patients whose cancer was detected after either the initial biopsy or the repeat biopsy were regarded as having prostate cancer for the development of the neural network. PSAV was measured as three PSA measurements at a 12-month period, all by using the same assay as described by Carter and Pearson.8 The ANN chosen for this analysis was the multilayer perceptron, described elsewhere.22,23 ANNs grew out of attempts to mimic the fault tolerance and capacity to learn of biologic nervous systems. They do this by modeling the low-level structure of the brain. A biologic nervous system is composed of a large number of neuron cells that are extensively interconnected to one another. Each neuron cell is a specialized entity that can propagate an electrochemical signal. Each neuron cell has a branching input structure, known as dendrites, and a branching output structure, known as axons. The axons of one cell are connected to the dendrites of other cells by synapses. Signals are propagated throughout this complex network, regulated primarily by the synapses. Donald Hebb, one of the foremost researchers in neurologic systems, has postulated that learning consists primarily of altering the strength of various synaptic connections. Likewise, a typical ANN consists of computational neurons or processing elements connected together by weighted signal pathways. ANNs typically have a much simpler architecture, with many fewer neurons and connections than does a biologic nervous system. An artificial neuron receives a number of inputs, either from data entering the network or as output from other neurons. Each input comes via a pathway connection that has strength—or, in terms of ANNs, weight. These weights correspond to synaptic strength in biologic systems. Each neuron also has a single threshold value. The activation of this artificial neuron is composed of the weighted sum of its inputs minus the threshold value. This activation signal is then transformed through an activation or transfer function to produce the output of the neuron. The transfer function is generally a nonlinear, continuously differentiable function and may not have a direct biologic equivalent. ANNs consist of input elements that bring in signals from the outside world in a manner similar to biologic sensory nerves from, for example, the eye. The input signals are fed to one or more layers of neurons through the weighted pathway connections. These "hidden" neurons process the input signals and produce signals that are sent to an output layer of neurons through weighted pathway connections. The output neurons generate an output signal to the outside world that is similar to biologic motor nerves connected, for example, to the hands. Training of the ANN takes place by a "teacher" program that loads training cases from a database and adjusts the weights and threshold values of the network to minimize the error between the real-world outputs and the neural network generated outputs for the training case inputs. Testing cases from the database are used an as independent check on training progress. Neural network training is stopped when the testing set no longer indicates a decrease in error between the real world outputs and the neural network generated outputs for the testing case inputs. This procedure greatly reduces the likelihood of overfitting the data. To ensure that overfitting does not occur, it is the usual practice to hold back a third set of data from the database. This third set is a validation data set and is never seen by the network for either training or testing, but this data set provides prospective data to the neural network. Statistical measures of network accuracy are usually quoted on the basis of the validations. The 974 patients with a total PSA of 4 to 10 ng/mL were randomly divided into three groups: a training group (487 patients), a test group (243 patients), and a validation or prospective group (244 patients). The training group was used by the training algorithm to adjust the weights and biases of the ANN. The training algorithm was then applied to the test group. Training with the test group ceased when errors in the test group were substantially higher than in the training group, thus ensuring that the algorithm would not overfit the training data. The validation group, which was not used in the development of the network, was then used to determine the predictive accuracy of the trained network. Likewise, the 272 patients with a total PSA of 2.5 to 4 ng/mL were divided randomly into a training group (136 patients), a test group (68 patients), and a validation or prospective group (68 patients). As was performed with the 4 to 10 ng/mL patients, these groups were then used to generate an ANN model for patients with a total PSA level of 2.5 to 4 ng/mL. A genetic algorithm24 was used in conjunction with the training algorithm to determine a near-optimum network architecture and to identify the input variables of most importance. This algorithm initially creates random populations of binary strings representing architectures and inputs. The algorithm then conducts a number of experiments with different configurations of the multilayer perceptron neural network, retaining a selection of the best performing networks in each generation. Configurations tested involve pre- and postprocessing options, number of hidden layers, number of neurons in each hidden layer, classification confidence thresholds, hidden layer transfer transaction, and feature selections determination of the optimum set of input variables. This process is repeated for several generations of data. The optimal tested network architecture/input combination is selected at the end of this process. In this analysis, more than 5,000 architecture/input combinations were modeled.
ANN Ranking of Input Variables
Cutoff Analysis for ANN Models
Receiver Operator Characteristic Curve Analysis
Statistical Analysis
The variables selected from the database for patients with a serum total PSA level of 4 to 10 ng/mL or 2.5 to 4 ng/mL were as follows: serum total PSA level; percentage of free PSA; PSAV; the TRUS variables PSAD, PSA-TZ, prostate volume, and TZ volume; and the DRE result. Values for these variables are listed in Table 1 for each group of patients.
In the final ANN selected for total PSA levels from 4 to 10 ng/mL, three variables—age, DRE result, and total prostate volume—were excluded from the model because they did not contribute to the predictive ability of the ANN. Variables chosen for inclusion (input variables) were, in order of importance, percentage of free PSA, PSA-TZ, PSAV, free PSA, TZ volume, total PSA, and PSAD. In the final ANN selected for total PSA levels from 2.5 to 4 ng/mL, the variables of age, total PSA, TZ volume, PSAV, and DRE result were excluded from the model because they did not contribute to the predictive ability of the network. Contributing variables, in order of importance, were PSA-TZ, percentage of free PSA, PSAD, and total volume of the prostate. Figure 1 (total PSA, 4 to 10 ng/mL) and Fig 2 (total PSA, 2.5 to 4 ng/mL) indicate the relationship between percentage of free PSA and PSA-TZ as modeled by the respective ANNs. The two figures were generated by using the mean values of the ANN input variables not depicted in the graph. The rather steep slope in each figure is indicative of the importance of PSA-TZ and percentage of free PSA in predicting prostate cancer versus benign prostatic tissue.
The ROC curve for the validation patient set for the ANN is indicated in Fig 3 for patients with a total PSA of 4 to 10 ng/mL and in Fig 4 for patients with a total PSA of 2.5 to 4 ng/mL. The areas under the ROC curve were 91.3% and 87.6%, respectively, which is indicative of each ANN model’s power for predicting the presence of prostate cancer.
Figure 3 also indicates ROC curves of conventional PSA parameters used as input variables in our ANN model for the patients with a total PSA of 4 to 10 ng/mL. The AUC for selected parameters are listed in Table 2. For patients with a serum total PSA level of 4 to 10 ng/mL, the AUC generated by the ANN is significantly higher than that produced by any of the single PSA parameters (P < .01).
Figure 4 indicates the ROC curves for conventional PSA parameters for the patients with a total PSA of 2.5 to 4 ng/mL. The AUC of the ANN ROC curve is significantly higher than the AUC for the single PSA parameters used as inputs to the ANN (P = .0239 for percentage of free PSA, P = .0204 for PSA-TZ, and P < .01 for both PSAD and total prostate volume; Table 3).
A predictive model needs to identify patients with cancer with a high degree of accuracy (high sensitivity), yet minimize the number of unnecessary tissue sampling procedures (high specificity). A sensitivity of 95% means that, on average, the predictive model will correctly identify 95% of patients with prostate cancer. The specificity value indicates the percentage of unnecessary tissue sampling procedures that would be eliminated by accurately identifying patients with benign disease. Table 2 lists the specificity, PPV, and NPV at 95% sensitivity for our ANN and various predictive models in patients with a total PSA of 4 to 10 ng/mL. Table 3 lists the sensitivity, PPV, and NPV at 95% specificity for our ANN and other predictive models in patients with a total PSA of 2.5 to 4 ng/mL. In both groups of patients, the ANN model was superior to the other models with respect to specificity or sensitivity, and PPV and NPV. In addition, both ANN models were superior with respect to specificity or sensitivity when compared with a multivariate logistic regression of the parameters used as input variables to the ANN.
PSA is considered the best serum marker of prostate cancer, but the measurement of PSA is associated with reduced diagnostic sensitivity and specificity in patients with PSA levels between 4 and 10 ng/mL. Approximately 80% of men with prostate cancer and 30% to 50% of those with BPH have a serum PSA level exceeding 4 ng/mL.26 Such patients may be subjected to unnecessary prostatic tissue sampling with attendant unnecessary cost and patient discomfort. Many techniques have been advocated and tested in an attempt to improve the sensitivity and specificity of prostate cancer detection. These techniques include correcting for increases in serum PSA levels resulting from an increase in the size of the prostate gland or the prostate gland TZ (PSAD and PSA-TZ, respectively). PSAV uses a change in the serum total PSA level of more than 0.75 ng/mL per year to identify patients who are likely to have prostate cancer. In addition, percentage of free PSA testing differentiates the percentage of free (unbound) PSA in serum relative to the percentage that is bound to various protease inhibitors. The relative proportion of these forms of PSA is known to differ in patients with prostate cancer relative to those with BPH.27 Each of these techniques, however, has limitations. PSAD and PSA-TZ require the patient to undergo TRUS and are thus subject to the technical limitations of the TRUS procedure. The calculation of PSAV is limited by the inherent variability of the PSA assay and the need to use consistent materials and methodology for each yearly measurement. The percentage of free PSA method requires the use of a total PSA assay and free PSA assay from the same manufacturer.28,29 We investigated the use of an ANN to improve the sensitivity and specificity of PSA-based early prostate cancer detection. An ANN uses sample (training) input and output data to define (learn) the interrelationships among the data. Once the ANN has been trained, it can then predict outcomes from new sets of input data. The processing performed by an ANN is different from that used by conventional statistical analysis. An ANN comprises an interconnected assembly of data processing units that emulate some of the higher-level functions of the architecture of the brain. As in the human brain, neurons and synapses are modeled, with various synaptic connection strengths (referred to as weights) for each connected pair of neurons. However, like many computer programs (and unlike the brain), a specific set of input and output neurons exist for each problem and each net. These input and output neurons correspond to the inputs to, and outputs from, a traditional computer program. The other (so-called hidden) neurons, along with the synapses and weights, correspond to the instructions in a traditional program. The ANN for the PSA range of 2.5 to 4 ng/mL, for example, has 4 input neurons, 1 output neuron, 2 hidden layers, 16 hidden neurons, and 100 synapses/weights. The synaptic weights contain the "intelligence" of the system. Percentage of free PSA, PSA-TZ, PSAV, free PSA, TZ volume, total PSA, and PSAD were selected as inputs into the ANN model for patients with a total serum PSA level of 4 to 10 ng/mL. Similarly, PSA-TZ, percentage of free PSA, PSAD, and total prostate volume were used as inputs into the ANN for the patients with a total PSA level of 2.5 to 4 ng/mL. These inputs were then used to calculate a composite output that considered the relative contribution of each input. Although others have reported on the use of ANNs to enhance prostate cancer screening and to better predict outcome in patients diagnosed with prostate cancer,30-35 this is one of the first reports of the prospective use of an ANN for the early detection of prostate cancer in the PSA range of 4 to 10 ng/mL. Our study of 1,246 men uses most of the leading methods of prostate cancer detection to produce sensitivity, specificity, PPV, NPV, and ROC curve AUC values that heretofore have not been possible. In the present study, at 95% sensitivity, the ANN model for patients with total PSA levels from 4 to 10 ng/mL produced specificity, PPV, NPV, and ROC curve AUC values that were superior to any of the other comparative parameters. In patients with total PSA levels from 2.5 to 4 ng/mL, our ANN model yielded sensitivity, PPV, NPV, and ROC curve AUC values at 95% specificity that were superior to any of the other comparative parameters. Elsewhere, we reported the results of a study designed to evaluate the ability of percentage of free PSA, PSA-TZ, PSAD, PSAV, and their combination to predict the presence of prostate cancer in men with serum total PSA levels between 4 and 10 ng/mL.11 Multivariate analysis indicated that the percentage of free PSA and PSA-TZ values were the most accurate predictors of prostate cancer. A cutoff of 37% for percentage of free PSA and of 0.25 ng/mL/cc for PSA-TZ had a sensitivity of 95%, with a corresponding specificity of 36% for percentage of free PSA and 47% for PSA-TZ. In another study, we compared the ability of total PSA, percentage of free PSA, PSAD, and PSA-TZ to predict the outcome of repeat prostatic biopsy in a prospective study of men with a total serum PSA between 4 and 10 ng/mL who were diagnosed with benign prostatic tissue after initial tests.17 Prostate cancer was detected on repeat testing in 10% of subjects. Again, percentage of free PSA and PSA-TZ were the best prognostic indicators of prostate cancer. At a cutoff of 38%, percentage of free PSA had a sensitivity of 95% and a specificity of 34%. The sensitivity and specificity of PSA-TZ at a cutoff of 0.19 ng/mL was 95% and 21%, respectively. In a study of patients with a serum total PSA level of 2.5 to 4 ng/mL, we found that a percentage of free PSA cutoff of 41% yielded a sensitivity of 95% with a specificity of 29%.15 In this same study, PSA-TZ had a 95% sensitivity and a 17% specificity at a cutoff of 0.095 ng/mL/cc. The sensitivity and specificity values for the ANN models we report here are far superior to any of our past results. Thus, our ANN model can identify virtually all cases of prostate cancer (high sensitivity) and would result in significantly fewer unnecessary tissue sampling procedures (high specificity) compared with all previously reported methods for the early detection of prostate cancer. This ANN model also has the advantage of accurately identifying benign disease in patients with low PSA levels. A 95% specificity means that a predictive model will, on average, accurately identify 95% of cases of benign disease. At 95% specificity in patients within the total PSA range of 2.5 to 4 ng/mL, our ANN model produced a PPV and NPV of 80% and 88%, respectively. The AUC of the ROC curve is a particularly useful measure of the predictive accuracy of a model. The AUC is an indication of the accuracy of a diagnostic test (which in the present study is the ability to distinguish between benign prostatic tissue and prostate cancer). AUC values range from 50% (no discriminatory power) to 100% (perfect predictive accuracy). The AUC of the ROC curve for the validation patient set processed by the ANN was 91.3% for patients with total PSA levels from 4 to 10 ng/mL. This value is superior to the results of the ROC curve analysis for each of the input variables in our current study (P < .01), as well as the analyses from our recent studies. In the combination and multivariate analysis study, the AUC of the ROC curve was 77.8% for percentage of free PSA and 82.7% for PSA-TZ.11 In this same study, the combination of percentage of free PSA and PSA-TZ significantly increased the AUC of the ROC curve (to 88.1%) compared with each of the other single parameters analyzed as well as their combination (P = .020, McNemar test). In the repeat biopsy study, the AUC of the ROC curve was 74.2% for percentage of free PSA and 82.7% for PSA-TZ when the results of both initial and repeat biopsies were considered.17 In our previous study of patients with serum total PSA levels between 2.5 and 4 ng/mL, the AUC of the ROC curve was 74.9% and 70.1% for percentage of free PSA and PSA-TZ, respectively.15 In comparison, the AUC of the ROC curve for the validation patient set processed by the ANN was 87.6% for patients with total PSA levels from 2.5 to 4 ng/mL. As was the case with the sensitivity and specificity values, the AUCs of the ROC curve for our ANN models were superior to that from all of our previous studies—including the relatively high AUC for the combination of percentage of free PSA and PSA-TZ (P < .05). In conclusion, the most important variables identified by our ANN analysis for predicting the presence of prostate cancer or benign prostatic tissue were percentage of free PSA and PSA-TZ. When combined in the ANN, these variables produce a highly accurate diagnostic tool for early prostate cancer detection, even when the patient has a low PSA range of 2.5 to 10 ng/mL. At 95% sensitivity for detecting prostate cancer, the optimized ANN is approximately twice as specific as percentage of free PSA, PSAD, PSA-TZ, and PSAV when used alone. At this sensitivity, the network gives a specificity of 67% for prospective validation patients with a total PSA level of 4 to 10 ng/mL. Even in the low PSA range from 2.5 to 4 ng/mL, the ANN elicited a sensitivity of 59% for prospective validation patients. The ANN was able to improve sensitivity (at 95% specificity) by 80% as compared with percentage of free PSA, PSA-TZ, PSAD, and PSAV in this latter group of patients. Thus, the ANN will allow a lowering of the total PSA cutoff value for prostate cancer detection to 2.5 ng/mL without markedly reducing specificity, as occurs with the use of biochemical markers only. To our knowledge, this is the most accurate predictive result yet reported for the diagnosis of prostate cancer and is due, in part, to the advanced analysis methods that are able to evolve from optimized ANNs, as well as to the care and expertise taken in gathering the multicenter patient data. In comparing the neural network to logistic regression, ROC curve areas indicate that the ANN is slightly (1.5% to 2.5%) more accurate than logistic regression. However, the ROC curve shapes favor the neural network. For a PSA level of more than 4 ng/mL, at a sensitivity for finding cancer of 95%, the specificity of the ANN is 7% greater than the logistic regression. For PSA levels less than 4 ng/mL, at a specificity for finding noncancer (benign prostatic tissue) of 95%, the ANN sensitivity is 26% greater than the logistic regression. When trying to find cancers (high sensitivity) or preventing unnecessary tissue sampling procedures (high specificity), the ANN demonstrates a significant improvement in predictive performance over logistic regression. The novel and highly accurate ANNs presented herein represent a significant advancement in the early detection of prostate cancer that will permit individualized counseling of patients referred for early prostate cancer detection. The potential improvement in sensitivity and specificity afforded by the ANN, relative to the conventional PSA parameters included in our study, will improve patient quality of life by eliminating unnecessary tissue sampling procedures. The availability of this ANN at no charge through the Internet (available at www.Uroservice.com and www.Urohealth .com) permits widespread access to this technology to urologists at all levels of practice.
1. Partin AW, Oesterling JE: The clinical usefulness of prostate specific antigen: Update 1994. J Urol 152: 1358-1368, 1994[Medline] 2. Stamey TA, Yang N, Hay AR, et al: Prostate-specific antigen as a serum marker for adenocarcinoma of the prostate. N Engl J Med 317: 909-916, 1987[Abstract] 3. Catalona WJ, Smith DS, Ratliff TL, et al: Measurement of prostate-specific antigen in serum as a screening test for prostate cancer. N Engl J Med 324: 1156-1161, 1991[Abstract] 4. Partin AW, Oesterling JE (eds): The clinical usefulness of percent free prostate-specific antigen. Urology 48:1-87, 1996 (suppl 6A) 5. Vashi AR, Wojno KJ, Henricks W, et al: Determination of the "reflex range" and appropriate cutpoints for percent free prostate-specific antigen in 413 men referred for prostatic evaluation using the AxSYM system. Urology 49: 19-27, 1997[CrossRef][Medline]
6. Catalona WJ, Partin AW, Slawin KM, et al: Use of the percentage of free prostate-specific antigen to enhance differentiation of prostate cancer from benign prostatic disease: A prospective multicenter clinical trial. JAMA 279: 1542-1547, 1998 7. Beduschi MC, Oesterling JE: Prostate-specific antigen density. Urol Clin North Am 24: 323-332, 1997[Medline] 8. Carter HB, Pearson JD: Prostate-specific antigen velocity and repeated measures of prostate-specific antigen. Urol Clin North Am 24: 333-338, 1997[CrossRef][Medline] 9. Richardson TD, Oesterling JE: Age-specific reference ranges for serum prostate-specific antigen. Urol Clin North Am 24: 339-351, 1997[CrossRef][Medline] 10. Djavan B, Zlotta AR, Byttebier G, et al: Prostate specific antigen density of the transition zone for early detection of prostate cancer. J Urol 160: 411-419, 1998[Medline] 11. Djavan B, Remzi M, Zlotta AR, et al: Combination and multivariate analysis of PSA-based parameters for prostate cancer prediction. Tech Urol 5: 71-76, 1999[Medline] 12. Stenman UH, Hakama M, Knekt P, et al: Serum concentrations of prostate specific antigen and its complex with alpha1-antichymotrypsin before diagnosis of prostate cancer. Lancet 344: 1594-1598, 1994[CrossRef][Medline] 13. Gann PH, Hennekens CH, Stampfer MJ: A prospective evaluation of plasma prostate-specific antigen for detection of prostatic cancer. JAMA 273: 289-294, 1995[Abstract] 14. Smith DS, Catalona WJ, Herschman JD: Longitudinal screening for prostate cancer with prostate-specific antigen. JAMA 276: 1309-1315, 1996[Abstract] 15. Djavan B, Zlotta A, Kratzik C, et al: PSA, PSA density, PSA density of transition zone, free/total PSA ratio, and PSA velocity for early detection of prostate cancer in men with serum PSA 2.5 to 4.0 ng/mL. Urology 54: 517-522, 1999[CrossRef][Medline] 16. Catalona WJ, Smith DS, Ornstein DK: Prostate cancer detection in men with serum PSA concentrations of 2.6 to 4.0 ng/mL and benign prostate examination. JAMA 277: 1452-1455, 1997[Abstract] 17. Djavan B, Zlotta A, Remzi M, et al: Optimal predictors of prostate cancer on repeat prostate biopsy: A prospective study of 1,051 men. J Urol 163: 1144-1149, 2000[CrossRef][Medline] 18. Rietbergen JBW, Boeken Kruger AE, Hoedemaeker RF, et al: Repeat screening for prostate cancer after 1-year followup in 984 biopsied men: Clinical and pathological features of detected cancer. J Urol 160: 2121-2125, 1998[CrossRef][Medline] 19. Morgan TO, McLeod DG, Leifer ES, et al: Prospective use of free prostate-specific antigen to avoid repeat prostate biopsies in men with elevated total prostate-specific antigen. Urology 48: 76-80, 1996 (suppl 6A)[Medline] 20. Fleshner NE, O’Sullivan M, Fair WR: Prevalence and predictors of a positive repeat transrectal ultrasound guided needle biopsy of the prostate. J Urol 158: 505-509, 1997[CrossRef][Medline] 21. Catalona WJ, Beiser JA, Smith DS: Serum free prostate specific antigen and prostate specific antigen density measurements for predicting cancer in men with prior negative prostatic biopsies. J Urol 158: 2162-2167, 1997[CrossRef][Medline] 22. Rumelhart DE, McClelland J (eds): Parallel Distributed Processing, Vol 1. Cambridge, MA, MIT Press, 1998 23. Bishop C: Neural Networks for Pattern Recognition. Oxford, United Kingdom, Oxford University Press, 1995 24. Goldberg DE: Genetic Algorithms. Reading, MA, Addison-Wesley, 1989 25. DeLong ER, DeLong DM, Clarke-Pearson DL: Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach. Biometrics 44: 837-845, 1988[CrossRef][Medline] 26. Beduschi MC, Oesterling JE: Percent free prostate-specific antigen: The next frontier in prostate-specific antigen testing. Urology 51: 98-109, 1998[Medline] 27. Christensson A, Björk T, Nilsson O, et al: Serum prostate specific antigen complexed to alpha1-antichymotrypsin as an indicator of prostate cancer. J Urol 150: 100-105, 1993[Medline] 28. Semjonow A, Oberpenning F, Brandt B, et al: Impact of free prostate-specific antigen on discordant measurement results of assays for total prostate-specific antigen. Urology 48: 10-15, 1996 (suppl 6A)[Medline] 29. Chan DW, Sokoll LJ, Jones KA, et al: Clinical evaluation of two free PSA assays in combination with different total PSA assays. J Urol 157: 112, 1997 (suppl) 30. Murphy GP, Snow PB, Brandt J, et al: Evaluation of prostate cancer patients receiving multiple staging tests, including ProstaScint scintiscans. Prostate 42: 145-149, 2000[Medline] 31. Finne P, Finne R, Auvinen A, et al: Predicting the outcome of prostate biopsy in screen-positive men by a multilayer perceptron network. Urology 56: 418-422, 2000[CrossRef][Medline] 32. Babaian RJ, Fritsche H, Ayala A, 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 56: 1000-1006, 2000[CrossRef][Medline] 33. Potter SR, Miller MC, Mangold LA, et al: Genetically engineered neural networks for predicting prostate cancer progression after radical prostatectomy. Urology 54: 791-795, 1999[CrossRef][Medline] 34. Loch T, Leuschner I, Genberg C, et al: Artificial neural network analysis (ANNA) of prostatic transrectal ultrasound. Prostate 39: 198-204, 1999[CrossRef][Medline] 35. Snow PB, Smith DS, Catalona WJ: Artificial neural networks in the diagnosis and prognosis of prostate cancer: A pilot study. J Urol 152: 1923-1926, 1994[Medline] Submitted January 25, 2001; accepted October 2, 2001. This article has been cited by other articles:
|
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
|||||||||||
|
|||||||||||
|
|
Copyright © 2002 by the American Society of Clinical Oncology, Online ISSN: 1527-7755. Print ISSN: 0732-183X
|
TOP
ABSTRACT
INTRODUCTION
4 ng/mL.
