Genetically engineered neural networks for predicting prostate cancer progression after radical prostatectomy

A systematic review of the applications of Expert Systems (ES) and machine learning (ML) in clinical urology

BACKGROUND

Testing a hypothesis for 'factors-outcome effect' is a common quest, but standard statistical regression analysis tools are rendered ineffective by data contaminated with too many noisy variables. Expert Systems (ES) can provide an alternative methodology in analysing data to identify variables with the highest correlation to the outcome. By applying their effective machine learning (ML) abilities, significant research time and costs can be saved. The study aims to systematically review the applications of ES in urological research and their methodological models for effective multi-variate analysis. Their domains, development and validity will be identified.

METHODS

The PRISMA methodology was applied to formulate an effective method for data gathering and analysis. This study search included seven most relevant information sources: WEB OF SCIENCE, EMBASE, BIOSIS CITATION INDEX, SCOPUS, PUBMED, Google Scholar and MEDLINE. Eligible articles were included if they applied one of the known ML models for a clear urological research question involving multivariate analysis. Only articles with pertinent research methods in ES models were included. The analysed data included the system model, applications, input/output variables, target user, validation, and outcomes. Both ML models and the variable analysis were comparatively reported for each system.

RESULTS

The search identified n = 1087 articles from all databases and n = 712 were eligible for examination against inclusion criteria. A total of 168 systems were finally included and systematically analysed demonstrating a recent increase in uptake of ES in academic urology in particular artificial neural networks with 31 systems. Most of the systems were applied in urological oncology (prostate cancer = 15, bladder cancer = 13) where diagnostic, prognostic and survival predictor markers were investigated. Due to the heterogeneity of models and their statistical tests, a meta-analysis was not feasible.

CONCLUSION

ES utility offers an effective ML potential and their applications in research have demonstrated a valid model for multi-variate analysis. The complexity of their development can challenge their uptake in urological clinics whilst the limitation of the statistical tools in this domain has created a gap for further research studies. Integration of computer scientists in academic units has promoted the use of ES in clinical urological research.

Quantification of large scale DNA organization for predicting prostate cancer recurrence

This study investigates whether Genomic Organization at Large Scales (which we propose to call GOALS) as quantified via nuclear phenotype characteristics and cell sociology features (describing cell organization within tissue) collected from prostate tissue microarrays (TMAs) can separate biochemical failure from biochemical nonevidence of disease (BNED) after radical prostatectomy (RP). Of the 78 prostate cancer tissue cores collected from patients treated with RP, 16 who developed biochemical relapse (failure group) and 16 who were BNED patients (nonfailure group) were included in the analyses (36 cores from 32 patients). A section from this TMA was stained stoichiometrically for DNA using the Feulgen-Thionin methodology, and scanned with a Pannoramic MIDI scanner. Approximately 110 nuclear phenotypic features, predominately quantifying large scale DNA organization (GOALS), were extracted from each segmented nuclei. In addition, the centers of these segmented nuclei defined a Voronoi tessellation and subsequent architectural analysis. Prostate TMA core classification as biochemical failure or BNED after RP using GOALS features was conducted (a) based on cell type and cell position within the epithelium (all cells, all epithelial cells, epithelial >2 cell layers away from basement membrane) from all cores, and (b) based on epithelial cells more than two cell layers from the basement membrane using a Classifier trained on Gleason 6, 8, 9 (16 cores) only and applied to a Test set consisting of the Gleason 7 cores (20 cores). Successful core classification as biochemical failure or BNED after RP by a linear classifier was 75% using all cells, 83% using all epithelial cells, and 86% using epithelial >2 layers. Overall success of predicted classification by the linear Classifier of (b) was 87.5% using the Training Set and 80% using the Test Set. Overall success of predicted progression using Gleason score alone was 75% for Gleason >7 as failures and 69% for Gleason >6 as failures. © 2017 International Society for Advancement of Cytometry.

Applications of Machine Learning in Cancer Prediction and Prognosis

Machine learning is a branch of artificial intelligence that employs a variety of statistical, probabilistic and optimization techniques that allows computers to “learn” from past examples and to detect hard-to-discern patterns from large, noisy or complex data sets. This capability is particularly well-suited to medical applications, especially those that depend on complex proteomic and genomic measurements. As a result, machine learning is frequently used in cancer diagnosis and detection. More recently machine learning has been applied to cancer prognosis and prediction. This latter approach is particularly interesting as it is part of a growing trend towards personalized, predictive medicine. In assembling this review we conducted a broad survey of the different types of machine learning methods being used, the types of data being integrated and the performance of these methods in cancer prediction and prognosis. A number of trends are noted, including a growing dependence on protein biomarkers and microarray data, a strong bias towards applications in prostate and breast cancer, and a heavy reliance on “older” technologies such artificial neural networks (ANNs) instead of more recently developed or more easily interpretable machine learning methods. A number of published studies also appear to lack an appropriate level of validation or testing. Among the better designed and validated studies it is clear that machine learning methods can be used to substantially (15–25%) improve the accuracy of predicting cancer susceptibility, recurrence and mortality. At a more fundamental level, it is also evident that machine learning is also helping to improve our basic understanding of cancer development and progression.

Chromatin changes predict recurrence after radical prostatectomy

Background:

Pathological evaluations give the best prognostic markers for prostate cancer patients after radical prostatectomy, but the observer variance is substantial. These risk assessments should be supported and supplemented by objective methods for identifying patients at increased risk of recurrence. Markers of epigenetic aberrations have shown promising results in several cancer types and can be assessed by automatic analysis of chromatin organisation in tumour cell nuclei.

Methods:

A consecutive series of 317 prostate cancer patients treated with radical prostatectomy at a national hospital between 1987 and 2005 were followed for a median of 10 years (interquartile range, 7–14). On average three tumour block samples from each patient were included to account for tumour heterogeneity. We developed a novel marker, termed Nucleotyping, based on automatic assessment of disordered chromatin organisation, and validated its ability to predict recurrence after radical prostatectomy.

Results:

Nucleotyping predicted recurrence with a hazard ratio (HR) of 3.3 (95% confidence interval (CI), 2.1–5.1). With adjustment for clinical and pathological characteristics, the HR was 2.5 (95% CI, 1.5–4.1). An updated stratification into three risk groups significantly improved the concordance with patient outcome compared with a state-of-the-art risk-stratification tool (P<0.001). The prognostic impact was most evident for the patients who were high-risk by clinical and pathological characteristics and for patients with Gleason score 7.

Conclusion:

A novel assessment of epigenetic aberrations was capable of improving risk stratification after radical prostatectomy.

Risk prediction models for biochemical recurrence after radical prostatectomy using prostate-specific antigen and Gleason score

Many computer models for predicting the risk of prostate cancer have been developed including for prediction of biochemical recurrence (BCR). However, models for individual BCR free probability at individual time-points after a BCR free period are rare. Follow-up data from 1656 patients who underwent laparoscopic radical prostatectomy (LRP) were used to develop an artificial neural network (ANN) to predict BCR and to compare it with a logistic regression (LR) model using clinical and pathologic parameters, prostate-specific antigen (PSA), margin status (R0/1), pathological stage (pT), and Gleason Score (GS). For individual BCR prediction at any given time after operation, additional ANN, and LR models were calculated every 6 months for up to 7.5 years of follow-up. The areas under the receiver operating characteristic (ROC) curve (AUC) for the ANN (0.754) and LR models (0.755) calculated immediately following LRP, were larger than that for GS (AUC: 0.715; P = 0.0015 and 0.001), pT or PSA (AUC: 0.619; P always <0.0001) alone. The GS predicted the BCR better than PSA (P = 0.0001), but there was no difference between the ANN and LR models (P = 0.39). Our ANN and LR models predicted individual BCR risk from radical prostatectomy for up to 10 years postoperative. ANN and LR models equally and significantly improved the prediction of BCR compared with PSA and GS alone. When the GS and ANN output values are combined, a more accurate BCR prediction is possible, especially in high-risk patients with GS ≥7.

Early diagnosis of complex diseases by molecular biomarkers, network biomarkers, and dynamical network biomarkers

Many studies have been carried out for early diagnosis of complex diseases by finding accurate and robust biomarkers specific to respective diseases. In particular, recent rapid advance of high-throughput technologies provides unprecedented rich information to characterize various disease genotypes and phenotypes in a global and also dynamical manner, which significantly accelerates the study of biomarkers from both theoretical and clinical perspectives. Traditionally, molecular biomarkers that distinguish disease samples from normal samples are widely adopted in clinical practices due to their ease of data measurement. However, many of them suffer from low coverage and high false-positive rates or high false-negative rates, which seriously limit their further clinical applications. To overcome those difficulties, network biomarkers (or module biomarkers) attract much attention and also achieve better performance because a network (or subnetwork) is considered to be a more robust form to characterize diseases than individual molecules. But, both molecular biomarkers and network biomarkers mainly distinguish disease samples from normal samples, and they generally cannot ensure to identify predisease samples due to their static nature, thereby lacking ability to early diagnosis. Based on nonlinear dynamical theory and complex network theory, a new concept of dynamical network biomarkers (DNBs, or a dynamical network of biomarkers) has been developed, which is different from traditional static approaches, and the DNB is able to distinguish a predisease state from normal and disease states by even a small number of samples, and therefore has great potential to achieve "real" early diagnosis of complex diseases. In this paper, we comprehensively review the recent advances and developments on molecular biomarkers, network biomarkers, and DNBs in particular, focusing on the biomarkers for early diagnosis of complex diseases considering a small number of samples and high-throughput data (or big data). Detailed comparisons of various types of biomarkers as well as their applications are also discussed.

Artificial neural networks and prostate cancer--tools for diagnosis and management

Artificial neural networks (ANNs) are mathematical models that are based on biological neural networks and are composed of interconnected groups of artificial neurons. ANNs are used to map and predict outcomes in complex relationships between given 'inputs' and sought-after 'outputs' and can also be used find patterns in datasets. In medicine, ANN applications have been used in cancer diagnosis, staging and recurrence prediction since the mid-1990s, when an enormous effort was initiated, especially in prostate cancer detection. Modern ANNs can incorporate new biomarkers and imaging data to improve their predictive power and can offer a number of advantages as clinical decision making tools, such as easy handling of distribution-free input parameters. Most importantly, ANNs consider nonlinear relationships among input data that cannot always be recognized by conventional analyses. In the future, complex medical diagnostic and treatment decisions will be increasingly based on ANNs and other multivariate models.

ARTIFICIAL INTELLIGENCE FOR THE PREDICTION OF BLADDER CANCER

New techniques for the prediction of tumour behaviour are needed as statistical analysis has a poor accuracy and is not applicable to the individual. Artificial intelligence (AI) may provide these suitable methods. We have previously shown that the predictive accuracies of neuro-fuzzy modelling (NFM) and artificial neural networks (ANN), two methods of AI, are superior to traditional statistical methods for the behaviour of bladder cancer (Catto et al, 2003). In this paper, we explain the AI techniques required to produce these predictive models. We used 9 parameters, which were a combination of experimental molecular biomarkers and conventional clinicopathological data, to predict the risk of tumour progression in a population of 109 patients with bladder cancer, NFM, using fuzzy logic to model data, achieved similar or superior predictive accuracy to ANN, which required cross-validation. However, unlike the impenetrable opaque structure of neural networks, the rules of NFM are transparent, enabling validation from clinical knowledge and the manipulation of input variables to allow exploratory predictions.

Nuclear morphometry, nucleomics and prostate cancer progression

Prostate cancer (PCa) results from a multistep process. This process includes initiation, which occurs through various aging events and multiple insults (such as chronic infection, inflammation and genetic instability through reactive oxygen species causing DNA double-strand breaks), followed by a multistep process of progression. These steps include several genetic and epigenetic alterations, as well as alterations to the chromatin structure, which occur in response to the carcinogenic stress-related events that sustain proliferative signaling. Events such as evading growth suppressors, resisting cell death, enabling replicative immortality, inducing angiogenesis, and activating invasion and metastasis are readily observed. In addition, in conjunction with these critical drivers of carcinogenesis, other factors related to the etiopathogenesis of PCa, involving energy metabolism and evasion of the immune surveillance system, appear to be involved. In addition, when cancer spread and metastasis occur, the 'tumor microenvironment' in the bone of PCa patients may provide a way to sustain dormancy or senescence and eventually establish a 'seed and soil' site where PCa proliferation and growth may occur over time. When PCa is initiated and progression ensues, significant alterations in nuclear size, shape and heterochromatin (DNA transcription) organization are found, and key nuclear transcriptional and structural proteins, as well as multiple nuclear bodies can lead to precancerous and malignant changes. These series of cellular and tissue-related malignancy-associated events can be quantified to assess disease progression and management.

Use of machine learning methods for prediction of acute toxicity in organs at risk following prostate radiotherapy

PURPOSE

The goal of this study is to investigate the advantages of large scale optimization methods vs conventional classification techniques in predicting acute toxicity for urinary bladder and rectum due to prostate irradiation.

METHODS

Clinical and dosimetric data of 321 patients undergoing prostate conformal radiotherapy were recorded. Gastro-intestinal and genito-urinary acute toxicities were scored according to the Radiation Therapy Oncology Group/European Organization for Research and Treatment of Cancer (RTOG/EORTC) scale. Patients were classified in two categories to separate mild (Grade < 2) from severe toxicity levels (Grade > 2). Machine learning methods at different complexity were implemented to predict toxicity as a function of multiple variables. The first approach consisted of a large scale optimization method, based on genetic algorithms (GAs) and artificial neural networks (ANN). The second approach was a binary classifier based on support vector machines (SVM).

RESULTS

The ANN and SVM-based solutions showed comparable prediction accuracy, exhibiting an area under the receiver operating characteristic (ROC) curve of 0.7. Different sensitivity and specificity features were measured for the two approaches. The ANN algorithm showed enhanced sensitivity if combined with appropriate classification criteria.

CONCLUSIONS

The results demonstrate that high sensitivity in toxicity prediction can be achieved with optimized ANNs, that are put forward to represent a valuable support in medical decisions. Future studies will be focused on enlarging the available patient database to increase the reliability of toxicity prediction algorithms and to define optimal classification criteria.

Theranostic and prognostic biomarkers: genomic applications in urological malignancies

Compared to other solid tumours such as breast, colon, and lung, the current clinical management of urological malignancies is lagging behind in terms of utilisation of clinically robust molecular tests that can identify patients that are more likely to respond to a given targeted agent, or even those in need of a more aggressive treatment approach based on well-validated molecular prognosticators. Several promising biomarkers for detection, prognosis, and targeted therapeutics are now under evaluation. The following review discusses some of the candidate biomarkers that may soon make their transition into clinically applicable assays in urological oncology patients.

Critical review of prostate cancer predictive tools

Prostate cancer is a very complex disease, and the decision-making process requires the clinician to balance clinical benefits, life expectancy, comorbidities and potential treatment-related side effects. Accurate prediction of clinical outcomes may help in the difficult process of making decisions related to prostate cancer. In this review, we discuss attributes of predictive tools and systematically review those available for prostate cancer. Types of tools include probability formulas, look-up and propensity scoring tables, risk-class stratification prediction tools, classification and regression tree analysis, nomograms and artificial neural networks. Criteria to evaluate tools include discrimination, calibration, generalizability, level of complexity, decision analysis and ability to account for competing risks and conditional probabilities. The available predictive tools and their features, with a focus on nomograms, are described. While some tools are well-calibrated, few have been externally validated or directly compared with other tools. In addition, the clinical consequences of applying predictive tools need thorough assessment. Nevertheless, predictive tools can facilitate medical decision-making by showing patients tailored predictions of their outcomes with various alternatives. Additionally, accurate tools may improve clinical trial design.

Inventory of prostate cancer predictive tools

PURPOSE OF REVIEW

We created an inventory of current predictive tools available for prostate cancer. This review may serve as an initial step toward a comprehensive reference guide for physicians to locate published nomograms that apply to the clinical decision in question. Using MEDLINE a literature search was performed on prostate cancer predictive tools from January 1966 to November 2007. We describe the patient populations to which they apply and the outcomes predicted, and record their individual characteristics.

RECENT FINDINGS

The literature search generated 111 published prediction tools that may be applied to patients in various clinical stages of disease. Of the 111 prediction tools, only 69 had undergone validation. We present an inventory of models with input variables, prediction form, number of patients used to develop the prediction tools, the outcome being predicted, prediction tool-specific features, predictive accuracy, and whether validation was performed.

SUMMARY

Decision rules, such as nomograms, provide evidence-based and at the same time individualized predictions of the outcome of interest. Such predictions have been repeatedly shown to be more accurate than those of clinicians, regardless of their level of expertise. Accurate risk estimates are also required for clinical trial design, to ensure homogeneous high-risk patient groups for whom new cancer therapeutics will be investigated.

Using nuclear morphometry to predict the need for treatment among men with low grade, low stage prostate cancer enrolled in a program of expectant management with curative intent

PURPOSE

We assessed the use of quantitative clinical and pathologic information to predict which patients would eventually require treatment for prostate cancer (CaP) in an expectant management (EM) cohort.

EXPERIMENTAL DESIGN

We identified 75 men having prostate cancer with favorable initial biopsy characteristics; 30 developed an unfavorable biopsy (Gleason grade >6, >2 cores with cancer, >50% of a core with cancer, or a palpable nodule) requiring treatment and 45 maintained favorable biopsies throughout a median follow-up of 2.7 years. Demographic, clinical data and quantitative tissue histomorphometry determined by digital image analysis were analyzed.

RESULTS

Logistic regression (LR) modeling generated a quantitative nuclear grade (QNG) signature based on the enrollment biopsy for differentiation of Favorable and Unfavorable groups using a variable LR selection criteria of P(z)<0.05. The QNG signature utilized 12 nuclear morphometric descriptors (NMDs) and had an area under the receiver operator characteristic curve (ROC-AUC) of 87% with a sensitivity of 82%, specificity of 70% and accuracy of 75%. A multivariable LR model combining QNG signature with clinical and pathological variables yielded an AUC-ROC of 88% and a sensitivity of 81%, specificity of 78% and accuracy of 79%. A LR model using prostate volume, PSA density, and number of pre-diagnosis biopsies resulted in an AUC-ROC of 68% and a sensitivity of 85%, specificity of 37% and accuracy of 56%.

CONCLUSIONS

QNG using EM prostate biopsies improves the predictive accuracy of LR models based on traditional clinicopathologic variables in determining which patients will ultimately develop an unfavorable biopsy. Our QNG-based model must be rigorously, prospectively validated prior to use in the clinical arena.

Is the GPSM Scoring Algorithm for Patients With Prostate Cancer Valid in the Contemporary Era?

PURPOSE

The GPSM (Gleason, prostate specific antigen, seminal vesicle and margin status) scoring algorithm is a user friendly model for predicting biochemical recurrence following radical retropubic prostatectomy. It was developed from patients who underwent radical retropubic prostatectomy from 1990 to 1993. We investigated the predictive ability of GPSM in the contemporary era.

MATERIALS AND METHODS

We identified 2,728 patients who underwent radical retropubic prostatectomy for prostate cancer from 1997 to 2000 at our institution. Cox proportional hazard regression models were used to develop multivariate scoring algorithms. Harrell's measure of concordance was used to compare the competing models.

RESULTS

In the contemporary era each GPSM feature remained significantly associated with biochemical recurrence in a multivariate model (each p <0.001). Harrell's measure of concordance for the algorithm was 0.706 vs 0.718 in the original study. After adjusting for GPSM on multivariate analysis Gleason primary 4/5 (p <0.001), DNA ploidy (p = 0.018) and tumor size (p <0.001) were associated with biochemical recurrence. However, none of these features increased Harrell's measure of concordance greater than 0.01 when added to the GPSM model. In addition, using the original 1990 to 1993 cohort, 495 patients with a GPSM score of 10 or greater were significantly more likely to die of prostate cancer compared with 2,169 with a GPSM score of less than 10 (at 15 years 13% vs 2%, HR 6.5, p <0.001).

CONCLUSIONS

The GPSM scoring algorithm is a simple predictive model that remains associated with biochemical recurrence in the contemporary era. In addition, to our knowledge the GPSM algorithm is the first nomogram associated with survival in patients with prostate cancer.

Predictive Ability of Partin Tables 2001 in a Welsh Population

OBJECTIVES

Partin tables are widely used to select and counsel patients prior to radical surgery for prostate cancer. However, Partin tables have been developed in the USA which has a different ethnic mixture from that of North Wales. We aimed to assess Partin tables' predictive ability in a Welsh population.

MATERIALS AND METHODS

193 patients underwent radical retropubic prostatectomy for clinically localized carcinoma of the prostate between April 1993 and July 2004 in a single institution in North Wales. Complete preoperative clinical staging information was available in 177 patients. Receiver operating characteristic curve analysis was used.

RESULTS

The mean patient age was 64 (48-73) years. Preoperative clinical staging distribution was: T1c 46.6% and T2 53.4%. 75% had organ-confined disease (TNM 1992). Extracapsular extension without seminal vesicle or lymph node involvement was seen in 13.5%. Nine percent had seminal vesicle invasion without lymph node involvement. Lymph node metastasis was found in 2.2%. The predictive effectiveness of the Partin table was high with an area under ROC curve of 0.733 for organ confinement, 0.738 for seminal vesicle invasion and 0.780 for lymph node involvement (CI 95%).

CONCLUSION

Our study demonstrated that the predictive ability of Partin tables for prostate cancer is also applicable to a Welsh population.

Prognostic factors and reporting of prostate carcinoma in radical prostatectomy and pelvic lymphadenectomy specimens

This paper, based on the activity of the Morphology-Based Prognostic Factors Committee of the 2004 World Health Organization-sponsored International Consultation, describes various methods of handling radical prostatectomy specimens for both routine clinical use and research purposes. The correlation between radical prostatectomy findings and postoperative failure is discussed in detail. This includes issues relating to pelvic lymph node involvement, detected both at the time of frozen section and in permanent sections. Issues of seminal vesicle invasion, including its definition, routes of invasion and relationship to prognosis, are covered in detail. The definition, terminology and incidence of extra-prostatic extension are elucidated, along with its prognostic significance relating to location and extent. Margins of resection are covered in terms of their definition, the etiology, incidence and sites of positive margins, the use of frozen sections to assess the margins and the relationship between margin positivity and prognosis. Issues relating to grade within the radical prostatectomy specimen are covered in depth, including novel ways of reporting Gleason grade and the concept of tertiary Gleason patterns. Tumor volume, tumor location, vascular invasion and perineural invasion are the final variables discussed relating to the prognosis of radical prostatectomy specimens. The use of multivariate analysis to predict progression is discussed, together with proposed modifications to the TNM system. Finally, biomarkers to predict progression following radical prostatectomy are described, including DNA ploidy, microvessel density, Ki-67, neuroendocrine differentiation, p53, p21, p27, Bcl-2, Her-2/neu, E-cadherin, CD44, retinoblastoma proteins, apoptotic index, androgen receptor status, expression of prostate-specific antigen and prostatic-specific acid phosphatase and nuclear morphometry.

Combining neural network and genetic algorithm for prediction of lung sounds

Recognition of lung sounds is an important goal in pulmonary medicine. In this work, we present a study for neural networks-genetic algorithm approach intended to aid in lung sound classification. Lung sound was captured from the chest wall of The subjects with different pulmonary diseases and also from the healthy subjects. Sound intervals with duration of 15-20 s were sampled from subjects. From each interval, full breath cycles were selected. Of each selected breath cycle, a 256-point Fourier Power Spectrum Density (PSD) was calculated. Total of 129 data values calculated by the spectral analysis are selected by genetic algorithm and applied to neural network. Multilayer perceptron (MLP) neural network employing backpropagation training algorithm was used to predict the presence or absence of adventitious sounds (wheeze and crackle). We used genetic algorithms to search for optimal structure and training parameters of neural network for a better predicting of lung sounds. This application resulted in designing of optimum network structure and, hence reducing the processing load and time.

Ability to predict metastasis based on pathology findings and alterations in nuclear structure of normal-appearing and cancer peripheral zone epithelium in the prostate

PURPOSE

Malignant transformation in the prostate produces significant alterations in glandular architecture (Gleason grade) and nuclear structure that provide valuable prognostic information. Normal-appearing nuclei (NN) adjacent to cancer may also have altered functions in response to malignancy. We studied NN adjacent to peripheral zone (PZ) prostate cancer (PCa), as well as the PZ cancer nuclei (CaN) using quantitative image cytometry. The nuclear structure information was combined with routine pathological findings to predict metastatic PCa progression and/or death.

EXPERIMENTAL DESIGN

Tissue microarrays of normal-appearing and cancer areas were prepared from 182 pathologist-selected paraffin blocks. Feulgen-stained CaN and NN were captured from the tissue microarrays using the AutoCyte Pathology Workstation. Multivariate logistic regression was used to calculate quantitative nuclear grade (QNG) solutions based on nuclear morphometric descriptors determined from NN and CaN. Multivariate logistic regression and Kaplan-Meier plots were also used to predict risk for distant metastasis and/or PCa-specific death using QNG solutions and routine pathology.

RESULTS

The pathology model yielded an area under the receiver operator characteristic curve of 72.5%. The QNG-NN and QNG-CaN solutions yielded an area under the receiver operator characteristic curve of 81.6 and 79.9%, respectively, but used different sets of nuclear morphometric descriptors. Kaplan-Meier plots for the pathology variables, the QNG-NN and QNG-CaN solutions, were combined with pathology to defined three statistically significantly distinct risk groups for distant metastasis and/or death (P < 0.0001).

CONCLUSIONS

Alterations in cancer or normal-appearing nuclei adjacent to peripheral zone cancer areas can predict PCa progression and/or death. The QNG-NN and QNG-CA solutions could be combined with pathology variables to improve the prediction of distant metastasis.

Biochemical recurrence and survival prediction models for the management of clinically localized prostate cancer

A number of new predictive modeling techniques have emerged in the past several years. These methods, which have been developed in fields such as artificial intelligence research, engineering, and meteorology, are now being applied to problems in medicine with promising results. This review outlines our recent work with use of selected advanced techniques such as artificial neural networks, genetic algorithms, and propensity scoring to develop useful models for estimating the risk of biochemical recurrence and long-term survival in men with clinically localized prostate cancer. In addition, we include a description of our efforts to develop a comprehensive prostate cancer database that, along with these novel modeling techniques, provides a powerful research tool that allows for the stratification of risk for treatment failure and survival by such factors as age, race, and comorbidities. Clinical and pathologic data from 1400 patients were used to develop the biochemical recurrence model. The area under the receiver operating characteristic curve for this model was 0.83, with a sensitivity of 85% and specificity of 74%. For the survival model, data from 6149 men were used. Our analysis indicated that age, income, and comorbidities had a statistically significant impact on survival. The effect of race did not reach statistical significance in this regard. The C index value for the model was 0.69 for overall survival. We conclude that these methods, along with a comprehensive database, allow for the development of models that provide estimates of treatment failure risk and survival probability that are more meaningful and clinically useful than those previously developed.

Artificial neural networks for predictive modeling in prostate cancer

Artificial neural networks (ANNs) represent a relatively new methodology for predictive modeling in medicine. ANNs, a form of artificial intelligence loosely based on the brain, have a demonstrated ability to learn complex and subtle relationships between variables in medical applications. In contrast with traditional statistical techniques, ANNs are capable of automatically resolving these relationships without the need for a priori assumptions about the nature of the interactions between variables. As with any technique, ANNs have limitations and potential drawbacks. This article provides an overview of the theoretical basis of ANNs, how they function, their strengths and limitations, and examples of how ANNs have been used to develop predictive models for the management of prostate cancer.

Human prostate cancer risk factors

Prostate cancer has the highest prevalence of any nonskin cancer in the human body, with similar likelihood of neoplastic foci found within the prostates of men around the world regardless of diet, occupation, lifestyle, or other factors. Essentially all men with circulating androgens will develop microscopic prostate cancer if they live long enough. This review is a contemporary and comprehensive, literature-based analysis of the putative risk factors for human prostate cancer, and the results were presented at a multidisciplinary consensus conference held in Crystal City, Virginia, in the fall of 2002. The objectives were to evaluate known environmental factors and mechanisms of prostatic carcinogenesis and to identify existing data gaps and future research needs. The review is divided into four sections, including 1) epidemiology (endogenous factors [family history, hormones, race, aging and oxidative stress] and exogenous factors [diet, environmental agents, occupation and other factors, including lifestyle factors]); 2) animal and cell culture models for prediction of human risk (rodent models, transgenic models, mouse reconstitution models, severe combined immunodeficiency syndrome mouse models, canine models, xenograft models, and cell culture models); 3) biomarkers in prostate cancer, most of which have been tested only as predictive factors for patient outcome after treatment rather than as risk factors; and 4) genotoxic and nongenotoxic mechanisms of carcinogenesis. The authors conclude that most of the data regarding risk relies, of necessity, on epidemiologic studies, but animal and cell culture models offer promise in confirming some important findings. The current understanding of biomarkers of disease and risk factors is limited. An understanding of the risk factors for prostate cancer has practical importance for public health research and policy, genetic and nutritional education and chemoprevention, and prevention strategies.

Use of genetic algorithms for neural networks to predict community-acquired pneumonia

BACKGROUND

Genetic algorithms have been used to solve optimization problems for artificial neural networks (ANN) in several domains. We used genetic algorithms to search for optimal hidden-layer architectures, connectivity, and training parameters for ANN for predicting community-acquired pneumonia among patients with respiratory complaints.

METHODS

Feed-forward back-propagation ANN were trained on sociodemographic, symptom, sign, comorbidity, and radiographic outcome data among 1044 patients from the University of Illinois (the training cohort), and were applied to 116 patients from the University of Nebraska (the testing cohort). Binary chromosomes with genes representing network attributes, including the number of nodes in the hidden layers, learning rate and momentum parameters, and the presence or absence of implicit within-layer connectivity using a competition algorithm, were operated on by various combinations of crossover, mutation, and probabilistic selection based on network mean-square error (MSE), and separately on average cross entropy (ENT). Predictive accuracy was measured as the area under a receiver-operating characteristic (ROC) curve.

RESULTS

Over 50 generations, the baseline genetic algorithm evolved an optimized ANN with nine nodes in the first hidden layer, zero nodes in the second hidden layer, learning rate and momentum parameters of 0.5, and no within-layer competition connectivity. This ANN had an ROC area in the training cohort of 0.872 and in the testing cohort of 0.934 (P-value for difference, 0.181). Algorithms based on cross-generational selection, Gray coding of genes prior to mutation, and crossover recombination at different genetic levels, evolved optimized ANN identical to the baseline genetic strategy. Algorithms based on other strategies, including elite selection within generations (training ROC area 0.819), and inversions of genetic material during recombination (training ROC area 0.812), evolved less accurate ANN.

CONCLUSION

ANN optimized by genetic algorithms accurately discriminated pneumonia within a training cohort, and within a testing cohort consisting of cases on which the networks had not been trained. Genetic algorithms can be used to implement efficient search strategies for optimal ANN to predict pneumonia.

Use of algorithms as determinants for individual patient decision making: national comprehensive cancer network versus artificial neural networks

The National Comprehensive Cancer Network (NCCN) developed a series of algorithms based on expert opinion to guide the treatment of patients with prostate cancer. These algorithms define acceptable treatment options according to the risk of disease recurrence and the life expectancy of the patient. However, practicing clinicians are expected to use medical judgment when making actual treatment decisions. Many clinical and pathologic variables affect patient prognosis, which, in turn, influences the treatment and surveillance of patients. Artificial neural networks (ANNs) offer promise for improving the predictive value of traditional statistical modeling. ANN models have been designed that predict risk of lymph node spread and capsular involvement during disease staging, risk of disease recurrence after prostatectomy, and overall and cause-specific survival. This article provides a review of guidelines, such as NCCN and ANN, used for the management of prostate cancer and suggests that group-level recommendations based on these algorithms or other decision trees may misrepresent individual patient preferences for treatment. Patients and their clinicians need to consider available prognostic information, including clinical status, pathologic variables, and comorbidities, and then select a reasonable treatment approach that maximizes outcome and quality of life according to the preferences of each patient.

Artificial neural networks for decision-making in urologic oncology

The authors are presenting a thorough introduction in Artificial Neural Networks (ANNs) and their contribution to modern Urologic Oncology. The article covers a description of Artificial Neural Network methodology and points out the differences of Artificial Intelligence to traditional statistic models in terms of serving patients and clinicians, in a different way than current statistical analysis. Since Artificial Intelligence is not yet fully understood by many practicing clinicians, the authors have reviewed a careful selection of articles in order to explore the clinical benefit of Artificial Intelligence applications in modern Urology questions and decision-making. The data are from real patients and reflect attempts to achieve more accurate diagnosis and prognosis, especially in prostate cancer that stands as a good example of difficult decision-making in everyday practice. Experience from current use of Artificial Intelligence is also being discussed, and the authors address future developments as well as potential problems such as medical record quality, precautions in using ANNs or resistance to system use, in an attempt to point out future demands and the need for common standards. The authors conclude that both methods should continue to be used in a complementary manner. ANNs still do not prove always better as to replace standard statistical analysis as the method of choice in interpreting medical data.

Can predictive models for prostate cancer patients derived in the United States of America be utilized in European patients? A validation study of the Partin tables

OBJECTIVES

Prostate cancer patients in the US and Europe differ due to selection and treatment differences. Accuracy of predictive tools derived in the US might therefore suffer when applied to European patients. We tested the validity of the widely accepted Partin tables for their ability to predict pathologic stage in German patients.

METHODS

Clinical and pathological characteristics were obtained from 1,298 consecutive men with clinically localized prostate cancer undergoing radical prostatectomy at the University Hospital Hamburg between January 1992 and February 2000. Receiver operating characteristic (ROC) curve analysis was performed to compare observed and predicted Partin rates for each pathologic stage.

RESULTS

The rate for organ confinement was 56% in Hamburg patients compared to 48% in the Partin study. The rates of Hamburg patients for extracapsular extension without seminal vesicle or lymph node involvement were 25%, for seminal vesicle without lymph node involvement 14% and for lymph node metastases 5%. The corresponding rates of the Partin study were 40, 7 and 5%, respectively. The accuracy of Partin table derived probability was high with an area under the ROC curve of 0.817 (95% CI, 0.757-0.876) for organ confinement and 0.807 (95% CI, 0.781-0.833) for lymph node involvement.

CONCLUSION

Our study demonstrated that predictive tools for prostate cancer developed in the US could be applied to European patients with comparable accuracy to that reported for validation studies performed with US patients.

Saw palmetto alters nuclear measurements reflecting DNA content in men with symptomatic BPH: evidence for a possible molecular mechanism

OBJECTIVES

To examine the nuclear chromatin characteristics of epithelial cells, looking for an SPHB-mediated effect on nuclear DNA structure and organization. Saw palmetto herbal blend (SPHB) causes contraction of prostate epithelial cells and suppression of tissue dihydrotestosterone levels in men with symptomatic benign prostatic hyperplasia, but a fundamental mechanism remains unknown.

METHODS

A 6-month randomized trial, comparing prostatic tissue of men treated with SPHB (n = 20) or placebo (n = 20), was performed. At baseline, the two groups were similar in age (65 versus 64 years), symptoms (International Prostate Symptom Score 18 versus 17), uroflow (maximal urinary flow rate 10 versus 11 mL/s), prostate volume (59 versus 58 cm(3)), prostate-specific antigen (4.2 versus 2.7 ng/mL), and percentage of epithelium (17% versus 16%). Prostatic tissue was obtained by sextant biopsy before and after treatment. Five-micron sections were Feulgen stained and quantitatively analyzed using the AutoCyte QUIC-DNA imaging system. Images were captured from 200 randomly selected epithelial cell nuclei, and 60 nuclear morphometric descriptors (NMDs) (eg, size, shape, DNA content, and textural features) were determined for each nucleus. Logistic regression analysis was used to assess the differences in the variances of the NMDs between the treated and untreated prostate epithelial cells.

RESULTS

At baseline, the SPHB and placebo groups had similar NMD values. After 6 months of placebo, no significant change from baseline was found in the NMDs. However, after 6 months of SPHB, 25 of the 60 NMDs were significantly different compared with baseline, and a multivariate model for predicting treatment effect using 4 of the 25 was created (P <0.001). The multivariate model had an area under the receiver operating characteristic curve of 94% and an accuracy of 85%.

CONCLUSIONS

Six months of SPHB treatment appears to alter the DNA chromatin structure and organization in prostate epithelial cells. Thus, a possible molecular basis for tissue changes and therapeutic effect of the compound is suggested.

Comparison of Logistic Regression and Neural Net Modeling for Prediction of Prostate Cancer Pathologic Stage

Background: Prostate cancer (PCa) pathologic staging remains a challenge for the physician using individual pretreatment variables. We have previously reported that UroScoreTM, a logistic regression (LR)-derived algorithm, can correctly predict organ-confined (OC) disease state with &gt;90% accuracy. This study compares statistical and neural network (NN) approaches to predict PCa stage.

Methods: A subset (756 of 817) of radical prostatectomy patients was assessed: 434 with OC disease, 173 with capsular penetration (NOC-CP), and 149 with metastases (NOC-AD) in the training sample. Additionally, an OC + NOC-CP (n = 607) vs NOC-AD (n = 149) two-outcome model was prepared. Validation sets included 120 or 397 cases not used for modeling. Input variables included clinical and several quantitative biopsy pathology variables. The classification accuracies achieved with a NN with an error back-propagation architecture were compared with those of LR statistical modeling.

Results: We demonstrated &gt;95% detection of OC PCa in three-outcome models, using both computational approaches. For training patient samples that were equally distributed for the three-outcome models, NNs gave a significantly higher overall classification accuracy than the LR approach (40% vs 96%, respectively). In the two-outcome models using either unequal or equal case distribution, the NNs had only a marginal advantage in classification accuracy over LR.

Conclusions: The strength of a mathematics-based disease-outcome model depends on the quality of the input variables, quantity of cases, case sample input distribution, and computational methods of data processing of inputs and outputs. We identified specific advantages for NNs, especially in the prediction of multiple-outcome models, related to the ability to pre- and postprocess inputs and outputs.

Prognostic value of DNA ploidy and nuclear morphometry in prostate cancer treated with androgen deprivation

OBJECTIVES

To assess the prognostic value of flow cytometry and nuclear morphometry in prostate cancer after androgen deprivation treatment.

METHODS

A total of 127 patients with a prostate cancer diagnosis who had undergone androgen suppression were retrospectively studied. The DNA content by flow cytometry and nuclear morphometry was studied from biopsy specimens. In the patients with Stage M0, two multivariate analyses by the Cox proportional regression model were performed to determine whether the experimental variables (DNA content and nuclear area) added independent information to the classic prognostic factors (Gleason score and stage). Using the statistical analysis results, risk groups were created.

RESULTS

T and M categories, Gleason score, DNA ploidy, and mean nuclear area proved to have prognostic value in the univariate analysis. For the group of patients free of metastasis (M0), it was possible to create low, intermediate, and high-risk groups using stage and Gleason score with statistically significant differences in survival. Multivariate analysis, combining the classic and experimental variables, selected Gleason score and DNA content as prognostic independent factors. Also, risk groups with statistically significant differences in survival were created. However, the net result of combining both kinds of factors was at least as valuable as the combination of stage and Gleason score in predicting survival.

CONCLUSIONS

The determination of DNA ploidy and mean nuclear area do not add enough independent information to improve the predictive value to justify their use in this group of patients treated with hormonal therapy.

Contemporary identification of patients at high risk of early prostate cancer recurrence after radical retropubic prostatectomy

OBJECTIVES

To develop a model that will identify a contemporary cohort of patients at high risk of early prostate cancer recurrence (greater than 50% at 36 months) after radical retropubic prostatectomy for clinically localized disease. Data from this model will provide important information for patient selection and the design of prospective randomized trials of adjuvant therapies.

METHODS

Proportional hazards regression analysis was applied to two patient cohorts to develop and cross-validate a multifactorial predictive model to identify men with the highest risk of early prostate cancer recurrence. The model and validation cohorts contained 904 and 901 men, respectively, who underwent radical retropubic prostatectomy at Johns Hopkins Hospital. This model was then externally validated using a cohort of patients from the Mayo Clinic.

RESULTS

A model for weighted risk of recurrence was developed: R(W)'=lymph node involvement (0/1)x1.43+surgical margin status (0/1)x1.15+modified Gleason score (0 to 4)x0.71+seminal vesicle involvement (0/1)x0.51. Men with an R(W)' greater than 2.84 (9%) demonstrated a 50% biochemical recurrence rate (prostrate-specific antigen level greater than 0.2 ng/mL) at 3 years and thus were placed in the high-risk group. Kaplan-Meier analyses of biochemical recurrence-free survival demonstrated rapid deviation of the curves based on the R(W)'. This model was cross-validated in the second group of patients and performed with similar results. Furthermore, similar trends were apparent when the model was externally validated on patients treated at the Mayo Clinic.

CONCLUSIONS

We have developed a multivariate Cox proportional hazards model that successfully stratifies patients on the basis of their risk of early prostate cancer recurrence.

Use of artificial neural networks in prostate cancer

Artificial neural networks (ANNs) are a type of artificial intelligence software inspired by biological neuronal systems that can be used for nonlinear statistical modeling. In recent years, these applications have played an increasing role in predictive and classification modeling in medical research. We review the basic concepts behind ANNs and examine the role of this technology in selected applications in prostate cancer research.

Predictive modeling techniques in prostate cancer

A number of new predictive modeling techniques have emerged in the past several years. These methods can be used independently or in combination with traditional modeling techniques to produce useful tools for the management of prostate cancer. Investigators should be aware of these techniques and avail themselves of their potentially useful properties. This review outlines selected predictive methods that can be used to develop models that may be useful to patients and clinicians for prostate cancer management.

Novel artificial neural network for early detection of prostate cancer

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.

Artificial neural network model to predict biochemical failure after radical prostatectomy

BACKGROUND

Biochemical failure, defined here as a rise in the serum prostate specific antigen (PSA) concentration to >0.3 ng/mL or the initiation of adjuvant therapy, is thought to be an adverse prognostic factor for men who undergo radical prostatectomy (RP) as definitive treatment for clinically localized cancer of the prostate (CAP). We have developed an artificial neural network (ANN) to predict biochemical failure that may benefit clinicians and patients choosing among the definitive treatment options for CAP.

MATERIALS AND METHODS

Clinical and pathologic data from 196 patients who had undergone RP at one institution between 1988 and 1999 were utilized. Twenty-one records were deleted because of missing outcome, Gleason sum, PSA, or clinical stage data. The variables from the 175 remaining records were analyzed for input variable selection using principal component analysis, decision tree analysis, and stepped logistic regression. The selected variables were age, PSA, primary and secondary Gleason grade, and Gleason sum. The records were randomized and split into three bootstrap training and validation sets of 140 records (80%) and 35 records (20%), respectively.

RESULTS

Forty-four percent of the patients suffered biochemical failure. The average duration of follow up was 2.5 years (range 0-11.5 years). Forty-two percent of the patients had pathologic evidence of non-organ-confined disease. The average area under the receiver operator characteristic (ROC) curve for the validation sets was 0.75 +/- 0.07. The ANN with the highest area under the ROC curve (0.80) was used for prediction and had a sensitivity of 0.74, a specificity of 0.78, a positive predictive value of 0.71, and a negative predictive value of 0.81.

CONCLUSION

These results suggest that ANN models can predict PSA failure using readily available preoperative variables. Such predictive models may offer assistance to patients and physicians deciding on definitive therapy for CaP.

Validation study of the accuracy of a postoperative nomogram for recurrence after radical prostatectomy for localized prostate cancer

PURPOSE

A postoperative nomogram for prostate cancer was developed at Baylor College of Medicine. This nomogram uses readily available clinical and pathologic variables to predict 7-year freedom from recurrence after radical prostatectomy. We evaluated the predictive accuracy of the nomogram when applied to patients of four international institutions.

PATIENTS AND METHODS

Clinical and pathologic data of 2,908 patients were supplied for validation, and 2,465 complete records were used. Nomogram-predicted probabilities of 7-year freedom from recurrence were compared with actual follow-up in two ways. First, the area under the receiver operating characteristic curve (AUC) was calculated for all patients and stratified by the time period of surgery. Second, calibration of the nomogram was achieved by comparing the predicted freedom from recurrence with that of an ideal nomogram. For patients in whom the pathologic report does not distinguish between focal and established extracapsular extension (an input variable of the nomogram), two separate calculations were performed assuming one or the other.

RESULTS

The overall AUC was 0.80 when applied to the validation data set, with individual institution AUCs ranging from 0.77 to 0.82. The predictive accuracy of the nomogram was apparently higher in patients who were operated on between 1997 and 2000 (AUC, 0.83) compared with those treated between 1987 and 1996 (AUC, 0.78). Nomogram predictions of 7-year freedom from recurrence were within 10% of an ideal nomogram.

CONCLUSION

The postoperative Baylor nomogram was accurate when applied at international treatment institutions. Our results suggest that accurate predictions may be expected when using this nomogram across different patient populations.

Quantitative nuclear grade (QNG): a new image analysis-based biomarker of clinically relevant nuclear structure alterations

This review addresses the potential clinical value of using quantitative nuclear morphometry information derived from computer-assisted image analysis for cancer detection and predicting outcomes such as tumor stage, recurrence, and progression. Today's imaging technology uses sophisticated hardware platforms coupled with powerful and user-friendly software packages that are commercially available as complete image analysis systems. There are many different mathematically derived nuclear morphometric descriptors (NMD's) (i.e. texture features) that can be calculated by these image analysis systems, but for the most part, these NMD's quantify nuclear size, shape, DNA content (ploidy), and chromatin organization (i.e. texture, both Markovian and non-Markovian) parameters. We have utilized commercially available image analysis systems and the NMD's calculated by these systems to create a mathematical solution, termed quantitative nuclear grade (QNG), for making clinical, diagnostic, and prognostic outcome predictions in both prostate and bladder cancer. A separate computational model is calculated for each outcome of interest using well-characterized and robust training, testing, and validation patient sample sets that adequately represent the selected population and clinical dilemma. A specific QNG solution may be calculated either by non-parametric statistical methods or non-linear mathematics employed by artificial neural networks (ANNs). The QNG solution, a measure of genomic instability, provides a unique independent variable to be used alone or to be included in an algorithm to assess a specific clinical outcome. This approach of customization of the nuclear morphometric descriptor (NMD) information through the calculation of a QNG solution mathematically adjusts for redundancy of features and reduces the complexity of the inputs used to create decision support tools for patient disease management. J. Cell. Biochem. Suppl. 35:151-157, 2000.