Search

Prediction of ground reaction force (GRF) magnitudes during running-based sports has several important applications, including optimal load prescription and injury prevention in athletes. Existing methods typically require information from multiple body-worn sensors, limiting their ecological validity, or aim to estimate discrete force parameters, limiting their ability to assess overall biomechanical load. This paper presents a neural network method to predict GRF time series from a single, commonly used, trunk-mounted accelerometer. The presented method uses a principal component analysis and multilayer perceptron (MLP) to obtain predictions. Time-series r2 coefficients with test data averaged around 0.9 for each impact, comparing favourably with alternative approaches which require additional sensors. For the impact peak, r2 was 0.74 across activities, comparing favourably with correlation analysis approaches. Several modifications, such as subject-specific training of the MLP, may help to improve results further, but the presented method can accurately predict GRF from trunk accelerometry data without requiring additional information. Results demonstrate the scope of machine learning to exploit common wearable technologies to estimate GRF in sport-specific environments. ispartof: MEDICAL ENGINEERING & PHYSICS vol:78 pages:82-89 ispartof: location:England status: published

This paper discusses the concepts of interpretability and explainability and outlines desiderata for robust interpretability. It then describes a neural network model that meets all criteria, with the addition of global faithfulness.

Neural networks are frequently applied to medical data. We describe how complex and imbalanced data can be modelled with simple but accurate neural networks that are transparent to the user. In the case of a data set on cervical cancer with 753 observations excluding, missing values, and 32 covariates, with a prevalence of 73 cases (9.69%), we explain how model selection can be applied to the Multi-Layer Perceptron (MLP) by deriving a representation using a General Additive Neural Network.

This case study benchmarks a range of statistical and machine learning methods relevant to computer-based decision support in clinical medicine, focusing on the diagnosis of knee osteoarthritis at first presentation. The methods, comprising logistic regression, Multilayer Perceptron (MLP), Chi-square Automatic Interaction Detector (CHAID) and Classification and Regression Trees (CART), are applied to a public domain database, the Osteoarthritis Initiative (OAI), a 10 year longitudinal study starting in 2002 (n = 4,796). In this real-world application, it is shown that logistic regression is comparable with the neural networks and decision trees for discrimination of positive diagnosis on this data set. This is likely because of weak non-linearities among high levels of noise. After comparing the explanations provided by the different methods, it is concluded that the interpretability of the risk score index provided by logistic regression is expressed in a form that most naturally integrates with clinical reasoning. The reason for this is that it gives a statistical assessment of the weight of evidence for making the diagnosis, so providing a direction for future research to improve explanation of generic non-linear models.

This paper presents a robust methodology to find biomarkers that are predictive of any given clinical outcome, by combining three critical steps: Adjustment for correlated biomarkers, through Linkage Disequilibrium pre-processing; False Detection Rate (FD) control with q-values; multivariate predictive modelling with neural networks. The results show that neural network modelling with pre-processing using p-values can be misleading. In particular, the interpretation of the neural network through calculation of the conditional probabilities P(x|c) where x represents covariates and c the classes, haw an important role in elucidating the predictive power (or lack of it) of the biomarkers. The methodology is generally applicable to p>n modelling where the initial pool of potential predictive parameters p, e.g. biomarkers, is greater than the sample size n.

Clustering issues are fundamental to exploratory analysis of bioinformatics data. This process may follow algorithms that are reproducible but make assumptions about, for instance, the ability to estimate the global structure by successful local agglomeration or alternatively, they use pattern recognition methods that are sensitive to the initial conditions. This paper reviews two clustering methodologies and highlights the differences that result from the changes in data representation, applied to a protein expression data set for breast cancer (n = 1,076). The two clustering methodologies are a reproducible approach to model-free clustering and a probabilistic competitive neural network. The results from the two methods are compared with existing studies of the same data set, and the preferred clustering solutions are profiled for clinical interpretation.

This paper proposes a novel type of higher-order pipelined neural network: the polynomial pipelined neural network. The proposed network is constructed from a number of higher-order neural networks concatenated with each other to predict highly nonlinear and nonstationary signals based on the engineering concept of divide and conquer. The polynomial pipelined neural network is used to predict the exchange rate between the US dollar and three other currencies. In this application, two sets of experiments are carried out. In the first set, the input data are pre-processed between 0 and 1 and passed to the neural networks as nonstationary data. In the second set of experiments, the nonstationary input signals are transformed into one step relative increase in price. The network demonstrates more accurate forecasting and an improvement in the signal to noise ratio over a number of benchmarked neural networks.

This paper proposes a methodology to optimise the future accuracy of a collaborative recommender application in a citizen Web portal. There are four stages namely, user modelling, benchmarking of clustering algorithms, prediction analysis and recommendation. The first stage is to develop analytical models of common characteristics of Web-user data. These artificial data sets are then used to evaluate the performance of clustering algorithms, in particular benchmarking the ART2 neural network with K-means clustering. Afterwards, it is evaluated the predictive accuracy of the clusters applied to a real-world data set derived from access logs to the citizen Web portal Infoville XXI (http://www.infoville.es). The results favour ART2 algorithms for cluster-based collaborative filtering on this Web portal. Finally, a recommender based on ART2 is developed. The follow-up of real recommendations will allow to improve recommendations by including new behaviours that are observed when users interact with the recommender system.

PurposeTo investigate the feasibility of using artificial neural networks for conceptual design of bracings systems for tall steel buildings.Design/methodology/approachDatabase of 234 design examples has been developed using commercially available detailed design software. These examples represent building up to 20 storeys. Feed forward back‐propagation neural network is trained on these examples. The results obtained from the artificial neural network are evaluated by re‐substitution, hold‐out and ten‐fold cross‐validation techniques.FindingsResults indicate that artificial neural network would give a performance of 97.91 percent (ten‐fold cross‐validation). The performance of this system is benchmarked by developing a binary logistic regression model from the same data. Performance of the two models has been compared using McNemar's test and receiver operation characteristics curves. Artificial neural network shows a better performance. The difference is found to be statically significant.Research limitations/implicationsThe developed model is applicable only to steel building up to 20 storeys. The feasibility of using artificial neural networks for conceptual design of bracings systems for tall steel buildings more than 20 storeys has not been investigated.Practical implicationsImplementation of the broad methodology outlined for the use of neural networks can be accomplished by conducting short training courses. This will provide personnel with flexibility in addressing buildings‐specifics bracing conditions and limitations.Originality/valueIn tall building design a lot of progress has been made in the development of software tools for numerical intensive tasks of analysis, design and optimization, however, professional software tools are not available to help the designer to choose an optimum building configuration at the conceptual design stage. The presented research provides a methodology to investigate the feasibility of using artificial neural networks for conceptual design of bracings systems for tall buildings. It is found that this approach for the selection of bracings in tall buildings is a better and cost effective option compared with database generated on the basis of expert opinion. It also correctly classifies and recommends the type of trussed bracing system.

In this study, the challenge to maximise the potential of gait analysis by employing advanced methods was addressed by using self-organising neural networks to quantify the deviation of patients' gait from normal. Data including three-dimensional joint angles, moments and powers of the two lower limbs and the pelvis were used to train Kohonen artificial neural networks to learn an abstract definition of normal gait. Subsequently, data from patients with gait problems were presented to the network which quantified the quality of gait in the form of a single curve by calculating the quantisation error during the gait cycle. A sensitivity analysis involving the manipulation of gait variables' weighting was able to highlight specific causes of the deviation including the anatomical location and the timing of wrong gait patterns. Use of the quantisation error can be regarded as an extension of previously described gait indices because it measures the goodness of gait and additionally provides information related to the causes underlying gait deviations.

Self-organising artificial neural networks were used to reduce the complexity of joint kinematic and kinetic data, which form part of a typical instrumented gait assessment. Three-dimensional joint angles, moments and powers during the gait cycle were projected from the multi-dimensional data space onto a topological neural map, which thereby identified gait stem-patterns. Patients were positioned on the map in relation to each other and this enabled them to be compared from their gait patterns. The visualisation of large amounts of complex data in a two-dimensional map labelled with gait patterns is a step towards more objective analysis protocols which may enhance decision making.

Artificial neural networks have featured in a wide range of medical journals, often with promising results. This paper reports on a systematic review that was conducted to assess the benefit of artificial neural networks (ANNs) as decision making tools in the field of cancer. The number of clinical trials (CTs) and randomised controlled trials (RCTs) involving the use of ANNs in diagnosis and prognosis increased from 1 to 38 in the last decade. However, out of 396 studies involving the use of ANNs in cancer, only 27 were either CTs or RCTs. Out of these trials, 21 showed an increase in benefit to healthcare provision and 6 did not. None of these studies however showed a decrease in benefit. This paper reviews the clinical fields where neural network methods figure most prominently, the main algorithms featured, methodologies for model selection and the need for rigorous evaluation of results.

Fitting the continuous valid logic of the exclusive OR (XOR) problem requires more than a minimal neural network configuration. This letter shows that the simplest logic to fit the XOR problem involves boundaries that cannot be mapped by a multi-layer perceptron with just two hidden nodes. This observation calls into question rule extraction methodologies based on pruning, since this practice can lock the network into a subspace of achievable continuous valued logic functions that prevent it from mapping the simplest logic to explain the data.

Automatic Relevance Determination (ARD), with multiple regularisation terms. A real-world classification problem, concerning the discrimination of online purchasers and non-purchasers using Internet’s WWW users’ opinions, is the test-bed for this assessment.

Realizing the full potential of the on-line consumer market requires careful identification of customer needs and expectations. As research on Internet consumer behavior is still in its infancy, a quantitative framework to characterize user profiles for e-commerce has not yet been established. This study proposes a quantitative framework that uses factor analysis to identify latent factor descriptors of Internet users' opinions on Web vendors and on-line shopping. Predictive models based on logistic discrimination and neural networks then select the factors most predictive of the propensity to buy on-line and classify Internet users accordingly. The application of this framework shows that the obtained latent factors agree in general with the major indicators identified in previous qualitative research. A small subset of the obtained factors is shown to retain the predictive power of the whole set. Neural networks are found to perform only marginally better than logistic discrimination in the task of classification.

The characterization and analysis of on-line customers' needs and expectations, regarding the Internet as a new marketing channel, is considered a prerequisite to the realization of the expected growth of the consumer-oriented electronic commerce market. The aim of the present study is twofold: to carry out an exploratory segmentation of this market that can throw some light upon its structure, and to characterize the on-line shopping adoption process. The Self-Organizing Map (SOM), an unsupervised neural network model devised by Kohonen (Kohonen, T. (1982). Self-organized formation of topologically correct feature maps. Biological Cybernetics, 43(1), 59–69; Kohonen, T., (1995). Self-organizing maps. Berlin: Springer) will be used as part of a tandem approach to segmentation, which involves the factor analysis of the observable variables in the data to be analyzed, prior to clustering. The SOM is shown to be a powerful data visualization tool, able to assist the data analysis, providing supervised methods with useful explanatory capabilities. It is also applied, in a completely unsupervised mode, to discover the clusters or segments that naturally occur in the data. The SOM is proposed as a flexible clustering model able to accommodate both Finer Segmentation and Normative Segmentation approaches. Within the latter, a cluster-partition is proposed and analysed, and high-level customer profiles, of potential interest to on-line marketers, are derived and described in marketing terms.

Despite its widespread use, plasma etching remains a poorly understood operation. Using chromatic monitor to measure key plasma parameters, the etch process was characterised using response surface methodology. In order to develop control-oriented models, open-loop dynamic testing was carried out. Dynamic models of a RIE system were developed using both standard system identification algorithms and neural network techniques.

Magnetic resonance spectroscopy opens a window into the biochemistry of living tissue. However, spectra acquired from different tissue types in vivo or in vitro and from body fluids contain a large number of peaks from a range of metabolites, whose relative intensities vary substantially and in complicated ways even between successive samples from the same category. The realization of the full clinical potential of NMR spectroscopy relies, in part, on our ability to interpret and quantify the role of individual metabolites in characterizing specific tissue and tissue conditions. This paper addresses the problem of tissue classification by analysing NMR spectra using statistical and neural network methods. It assesses the performance of classification models from a range of statistical methods and compares them with the performance of artificial neural network models. The paper also assesses the consistency of the models in selecting, directly from the spectra, the subsets of metabolites most relevant for differentiating between tissue types. The analysis techniques are examined using in vitro spectra from eight classes of normal tissue and tumours obtained from rats. We show that, for the given data set, the performance of linear and non-linear methods is comparable, possibly due to the small sample size per class. We also show that using a subset of metabolites selected by linear discriminant analysis for further analysis by neural networks improves the classification accuracy, and reduces the number of metabolites necessary for correct classification.

A PC-based neural-network processing system is described for the interpretation of chromatic sensor information monitored remotely using a CCD camera. Chromaticity based monitoring provides a means of simple data pre-processing to reduce system noise due to total intensity variation. However, sensing elements using chromaticity usually show complex variations in chromaticity with measurand changes. This contribution shows how these complex variations may be processed to yield high resolution values of a measurand. The paper presents a novel application of neural networks for monitoring temperatures using thermochromic materials and to perform 2-D pressure analysis using photoelastic materials. The inherent complex mapping and generalisation abilities of multi-layered perceptrons (MLP) make them ideal for processing the detected signals. Results are presented showing that the neural network can provide levels of resolution and performance for remotely addressing chromatic transducers, which are acceptable for detailed metrological applications.

Real-time predictive control requires a forward model that is both accurate and fast. This paper introduces two nonlinear internal memory network architectures and compares their performance with a Multi-layer Perceptron (MLP) augmented with the use of spread encoding. The test plant is a single component from an Underwater Robotic Vehicle (URV), comprising a thruster encased in a steel frame and provided with buoyancy. This assemblv is free to move under water and is controlled for depth. The internal memory networks are of comparable accuracy to the MLP but more parsimonious, resulting, in a faster response which makes them better suited for on-line control. Although a particular case study is presented as the focus of this paper, the algorithms and methods developed have generic applicability.

Neural networks are currently finding practical applications ranging from ‘soft’ regulatory control in consumer products to accurate control of non-linear plant in the process industries. This paper describes the application of neural networks to modelling and control of a prototype underwater vehicle, as an example of a system containing severe non-linearities. The most common implementation strategy for neural control is model predictive control, where a model of the process is developed first and is used off-line to design an appropriate compensator. The accuracy and robustness of this control strategy relies on the quality of the non-linear process model, in particular its ability to predict the plant response accurately multiple-steps ahead. In this paper, several neural network architectures are used to evaluate a long-range model predictive control structure, both in simulation and for on-line control of vehicle depth, achieving accurate control with a smooth actuator signal.

This paper addresses the problem of compensation mechanisms which can be used by Bayesian Neural Networks (BNNs) when dealing with skewed training data. The compensation mechanisms are used to balance the training data towards a mean value so that to be able to calculate the marginalized neural network predictions. There are presented 2 compensation mechanisms and each of them is applied to a BNN: a local compensation mechanism and a global mechanism. There is presented a third BNN model which does not use a compensation mechanism. It is shown that in the absence of a compensation mechanism, the marginalized network outputs can still be calculated through a scaling of the Jacobian and Hessian matrixes involved in the respective calculations. The standard BNN is a Partial Logistic Artificial Neural Network with Automatic Relevance Determination, which has multiple competing network outputs which corresponds to the Competing Risks (CRs) type of analysis specific to the medical domain of survival analysis. The resulted model is entitled the PLANN-CR-ARD model. The three versions of the PLANN-CR-ARD model are tested on a very demanding medical dataset taken from the survival analysis. The ARD framework implements the calculation of the network outputs, the marginalization of the network outputs and the model selection. The numerical results show that the neural network model based on the global compensation is very effective.

A benchmark medical study is realized for a Primary Biliary Cirrhosis (PBC) dataset by using two different versions of a Bayesian Neural Network (BNN) entitled Partial Logistic Artificial Neural Network for Competing Risks with Automatic Relevance Determination (PLANN-CR-ARD). The two BNN versions are based on two different compensation mechanisms which are designed to preserve the numerical stability of the PLANN-CR-ARD model and to calculate the marginalized network results. The predictions of the PLANN-CR-ARD models are comparable to the non-parametric estimates obtained through the survival analysis of the PBC dataset. The input variables from the PBC dataset which can have a strong influence on the outcome of the disease are determined. The PLANN-CR-ARD models can be used to investigate the non-linear inter-dependencies between the predicted outputs and the input data which consist of the characteristics of the PBC patients.

This paper focuses on the use of the standard multi-layer perceptron (MLP) neural network to provide accurate multi-step-ahead predictions of non-linear dynamical systems. A spread encoding method of representing continuous variables in a form suitable for presentation to an MLP is investigated. With this technique each numerical value is spread over the activity of several nodes at the inputs and outputs of the network. The main purpose of using spread encoding in this application is to form representations with sufficient accuracy to allow a neural network, trained using conventional feed-forward algorithms, to be used recursively. In this mode the network is required to predict the time evolution of the process output multiple time steps into the future, thus acting as a process model which has potential for improving control strategies that rely on a model of the plant and enhancing the performance of neural networks when used as simulation tools. The spread encoding form of data representation is compared to the conventional scaling method in an application of the MLP to modelling the response of a non-linear process. Results demonstrate that significant improvements in the neural network model prediction accuracy can be achieved using the spread encoding technique. The ability of the network model to capture the process dynamics is further illustrated by examining the localised frequency response of the network, in a novel application of spectral analysis techniques. The paper also includes introductory material on using neural networks for multi-step and single-step prediction.

This paper investigates the performance of the PLANN-CR-ARD network predictions through a comparison with the confidence intervals and the non-parametric estimates obtained from the survival analysis of a Primary Billiary Cirrhosis (PBC) dataset. The predictions of the PLANN-CR-ARD model are marginalized using two methods: approximation of the integral of marginalization and the Monte Carlo method. The numerical results show that the PLANN-CR-ARD predicts very well, the results being situated within the confidence intervals of the non-parametric estimates. The Model Selection is also performed on the same dataset. The PLANN-CR-ARD can be used to explore the non-linear interdependencies between the predicted outputs and the input data which in survival analysis describes the characteristics of the patients.

This paper presents a new compensation mechanism to be used with a Partial Logistic Artificial Neural Network for Competing Risks with Automatic Relevance Determination (PLANN-CR-ARD) and tested comprehensibly on a real breast cancer dataset with excellent convergence properties and numerical stability for the non-linear model. The Model Selection is implemented for the PLANNCRARD model, benefiting from a scaling of the prior error term which together with the data error term forms the total error function that is optimized. The PLANN-CR-ARD proves to be an excellent prognostic tool that can be used in regression analysis tasks such as the survival analysis of cancer datasets.

International audience; This paper proposes a Web clinical decision support system for clinical oncologists and for breast cancer patients making prognostic assessments, using the particular characteristics of the individual patient. This system comprises three different prognostic modelling methodologies: the clinically widely used Nottingham prognostic index (NPI); the Cox regression modelling and a partial logistic artificial neural network with automatic relevance determination (PLANN-ARD). All three models yield a different prognostic index that can be analysed together in order to obtain a more accurate prognostic assessment of the patient. Missing data is incorporated in the mentioned models, a common issue in medical data that was overcome using multiple imputation techniques. Risk group assignments are also provided through a methodology based on regression trees, where Boolean rules can be obtained expressed with patient characteristics.

Clinical characterization of breast cancer patients related to their risk and profiles is an important part for making their correct prognostic assessments. This paper first proposes a prognostic index obtained when it is applied a flexible non-linear time-to-event model and compares it to a widely used linear survival estimator. This index underpins different stratification methodologies including informed clustering utilising the principle of learning metrics, regression trees and recursive application of the log-rank test. Missing data issue was overcome using multiple imputation, which was applied to a neural network model of survival fitted to a data set for breast cancer (n=743). It was found the three methodologies broadly agree, having however important differences.

The classification and recognition of two-dimensional patterns independently of their position, orientation, and size by using high-order networks are discussed. A method is introduced for reducing and controlling the number of weights of a third-order network used for invariant pattern recognition. The method leads to economical networks that exhibit high recognition rates for translated, rotated, and scaled, as well as locally distorted, patterns. The performance of these networks at recognizing types and handwritten numerals independently of their position, size, and orientation is compared with and found superior to the performance of a layered feedforward network to which image features extracted by the method of moments are presented as input. >

Novel higher order polynomial neural network architecture is presented in this paper. The new proposed neural network is called Dynamic Ridge Polynomial neural network that combines the properties of higher order and recurrent neural networks. The advantage of this type of network is that it exploits the properties of higher-order neural networks by functionally extending the input space into a higher dimensional space, where linear separability is possible, without suffering from the combinatorial explosion in the number of weights. Furthermore, the network has a regular structure, since the order can be suitably augmented by additional sigma units. Finally, the presence of the recurrent link expands the network’s ability for attractor dynamics and storing information for later use. The performance of the network is tested for the prediction of nonlinear and nonstationary time series. Two popular time series, the Lorenz attractor and the mean value of the AE index, are used in our studies. The simulation results showed better results in terms of the signal to noise ratio in comparison to a number of higher order and feedforward networks.

Time-to-event analysis is important in a wide range of applications from clinical prognosis to risk modeling for credit scoring and insurance. In risk modeling, it is sometimes required to make a simultaneous assessment of the hazard arising from two or more mutually exclusive factors. This paper applies to an existing neural network model for competing risks (PLANNCR), a Bayesian regularization with the standard approximation of the evidence to implement automatic relevance determination (PLANNCR-ARD). The theoretical framework for the model is described and its application is illustrated with reference to local and distal recurrence of breast cancer, using the data set of Veronesi (1995).

Clinical management often relies on stratification of patients by outcome. The application of flexible non-linear time-to-event models to stratification of patient populations into different and clinically meaningful risk groups is currently an important area of research. This paper proposes a definition of prognostic index for neural network models of survival. This index underpins different stratification strategies including k-means clustering, regression trees and recursive application of the log-rank test. It was obtained with multiple imputation applied to a neural network model of survival fitted to a substantial data set for breast cancer (n=931) and was evaluated with a large out of sample data set (n=4,083). It was found that the constraint imposed by regression trees on the form of the permitted rules makes it less specific than stratifying directly from the prognostic index and deriving unconstrained low-order rules with Orthogonal Search Rule Extraction.

Early characterization of patients with respect to their predicted response to treatment is a fundamental step towards the delivery of effective, personalized care. Starting from the results of a time-to-event model with competing risks using the framework of partial logistic artificial neural networks with automatic relevance determination (PLANNCR-ARD), we discuss an effective semi-supervised approach to patient stratification with application to Acute Myeloid Leukaemia (AML) data (n = 509) acquired prospectively by the GIMEMA consortium. Multiple prognostic indices provided by the survival model are exploited to build a metric based on the Fisher information matrix. Cluster number estimation is then performed in the Fisher-induced affine space, yielding to the discovery of a stratification of the patients into groups characterized by significantly different mortality risks following induction therapy in AML. The proposed model is shown to be able to cluster the input data, while promoting specificity of both target outcomes, namely Complete Remission (CR) and Induction Death (ID). This generic clustering methodology generates an affine transformation of the data space that is coherent with the prognostic information predicted by the PLANNCR-ARD model.

This work is intended for providing a review of reallife practical applications of Artificial Intelligence (AI) methods. We focus on the use of Machine Learning (ML) methods applied to rather real problems than synthetic problems with standard and controlled environment. In particular, we will describe the following problems in next sections: • Optimization of Erythropoietin (EPO) dosages in anaemic patients undergoing Chronic Renal Failure (CRF). • Optimization of a recommender system for citizen web portal users. • Optimization of a marketing campaign. The choice of these problems is due to their relevance and their heterogeneity. This heterogeneity shows the capabilities and versatility of ML methods to solve real-life problems in very different fields of knowledge. The following methods will be mentioned during this work: • Artificial Neural Networks (ANNs): Multilayer Perceptron (MLP), Finite Impulse Response (FIR) Neural Network, Elman Network, Self-Oganizing Maps (SOMs) and Adaptive Resonance Theory (ART). • Other clustering algorithms: K-Means, Expectation- Maximization (EM) algorithm, Fuzzy C-Means (FCM), Hierarchical Clustering Algorithms (HCA). • Generalized Auto-Regressive Conditional Heteroskedasticity (GARCH). • Support Vector Regression (SVR). • Collaborative filtering techniques. • Reinforcement Learning (RL) methods.

A complete solution of the excitation values which may occur at the local minima of the XOR problem is obtained analytically for two-layered networks in the two most commonly quoted configurations, using the gradient backpropagation algorithm. The role of direct connections which bypass the two-layered system is discussed in connection to the XOR problem and other related training tasks.

A novel neural network clustering algorithm, CoRe, is benchmarked against previously published results on a breast cancer data set and applying the method of Partition Around Medoids (PAM). The data serve to compare the samples partitions obtained with the neural network, PAM and model-based algorithms, namely Gaussian Mixture Model (GMM), Variational Bayesian Gaussian Mixture (VBG) and Variational Bayesian Mixtures with Splitting (VBS). It is found that CoRe, on the one hand, agrees with the previously published partitions; on the other hand, it supports the existence of a supplementary cluster that we hypothesize to be an additional tumor subgroup with respect to those previously identified by PAM.

This paper describes the evaluation of a regularized Bayesian neural network model in prognostic applications. A total sample size of 5442 subjects treated with ocular melanoma in two centers; Liverpool and Paris was used to carry out external validation analysis of the model. The performance of the model was benchmarked against the traditional Cox regression model and a clinically accepted TNM staging system. The cumulative hazards curve for the neural network model was much closer to the empirical curve in the test data than the one produced by the Cox model. The neural network model showed equal performance to Cox's model in terms of discrimination. However, the neural network model was better than Cox's model in terms of calibration. The paper proposes an alternative staging system based on the model, which takes into account histopathological information. The new system has many advantages over the existing staging system.

The combination of an Artificial Neural Network classifier, a feature selection process, and a novel linear dimensionality reduction technique that provides a data projection for visualization and which preserves completely the class discrimination achieved by the classifier, is applied in this study to the analysis of an international, multi-centre 1H-MRS database of brain tumors. This combination yields results that are both intuitively interpretable and very accurate. The method as a whole remains simple enough as to allow its easy integration in existing medical decision support systems.

Missing values are common in medical datasets and may be amenable to data imputation when modelling a given data set or validating on an external cohort. This paper discusses model averaging over samples of the imputed distribution and extends this approach to generic non-linear modelling with the Partial Logistic Artificial Neural Network (PLANN) regularised within the evidence-based framework with Automatic Relevance Determination (ARD). The study then applies the imputation to external validation over new patient cohorts, considering also the case of predictions made for individual patients. A prognostic index is defined for the non-linear model and validation results show that 4 statistically significant risk groups identified at the 95% level of confidence from the modelling data, from Christie Hospital (n=931), retain good separation during external validation with data from the British Columbia Cancer Agency (n=4,083).

The authors describe two types of associative memory which are obtained by iterating a multilayered perceptron recursively after training it to auto-associate. The convergence of the associative memories with real-valued output and additional weights connecting the input and output layers directly is demonstrated analytically. Computational results are presented to illustrate the key concepts used in the proof, and also to characterise the capacity of these networks, comparing them with the Hopfield model and an alternative formulation of the multilayered auto-associative memory with thresholding to binary outputs. The capacity of the real-valued memory exceeds that of the alternative formulations when training with noise is used.

Evidence-based medicine has grown in stature over the last three decades and is now regarded a key development of modern medicine. The evidence base can be heterogeneous, involving both qualitative knowledge and measured quantitative data. Machine Learning (ML) methods have also begun to establish themselves as an alternative and promising approach to computer-based data analysis in oncology, as this field moves gradually away from being the preserve of traditional statistical analysis. In this paper, we describe the main areas of cancer research in which ML methods are currently being applied, and briefly discuss some of the advantages and disadvantages of their application.

International audience; This paper presents an analysis of censored survival data for breast cancer specific mortality and disease-free survival. There are three stages to the process, namely time-to-event modelling, risk stratification by predicted outcome and model interpretation using rule extraction. Model selection was carried out using the benchmark linear model, Cox regression but risk staging was derived with Cox regression and with Partial Logistic Regression Artificial Neural Networks regularised with Automatic Relevance Determination (PLANN-ARD). This analysis compares the two approaches showing the benefit of using the neural network framework especially for patients at high risk. The neural network model also has results in a smooth model of the hazard without the need for limiting assumptions of proportionality. The model predictions were verified using out-of-sample testing with the mortality model also compared with two other prognostic models called TNG and the NPI rule model. Further verification was carried out by comparing marginal estimates of the predicted and actual cumulative hazards. It was also observed that doctors seem to treat mortality and disease-free models as equivalent, so a further analysis was performed to observe if this was the case. The analysis was extended with automatic rule generation using Orthogonal Search Rule Extraction (OSRE). This methodology translates analytical risk scores into the language of the clinical domain, enabling direct validation of the operation of the Cox or neural network model. This paper extends the existing OSRE methodology to data sets that include continuous-valued variables.

In this paper we describe and compare two neural network models aimed at survival analysis modeling, based on formulations in continuous and discrete time. Learning in both models is approached in a Bayesian inference framework. We test the models on a real survival analysis problem, and we show that both models exhibit good discrimination and calibration capabilities. The C index of discrimination varied from 0.8 (SE=0.093) at year 1, to 0.75 (SE=0.034) at year 7 for the continuous time model; from 0.81 (SE=0.07) at year 1, to 0.75 (SE=0.033) at year 7 for the discrete time model. For both models the calibration was good (p

International audience; A three stage development process for the production of a hierarchical rule based prognosis tool is described. The application for this tool is specific to breast cancer patients that have a positive expression of the HER 2 gene. The first stage is the development of a Bayesian classification neural network to classify for cancer specific mortality. Secondly, low-order Boolean rules are extracted form this model using an orthogonal search based rule extraction (OSRE) algorithm. Further to these rules additional information is gathered from the Kaplan-Meier survival estimates of the population, stratified by the categorizations of the input variables. Finally, expert knowledge is used to further simplify the rules and to rank them hierarchically in the form of a decision tree. The resulting decision tree groups all observations into specific categories by clinical profile and by event rate. The practical clinical value of this decision support tool will in future be tested by external validation with additional data from other clinical centres.

Artificial neural networks have been shown to be effective as general non-linear models with applications to medical diagnosis, prognosis and survival analysis. This chapter begins with a review of artificial neural networks used as non-linear regression models in the survival analysis of breast cancer patients. These techniques are of much interest because they allow modelling of time-dependent hazards in the presence of complex non-linear and non-additive effects between covariates. First, the role of neural networks is introduced within the context of statistical methods and parametric techniques for prognosis of survival in breast cancer. Second, these methods are applied in a study comprising node-negative breast cancer patients in order to evaluate the evidence for improved models or combination of prognostic indices to be used in a clinical environment. In particular, node-negative breast cancer is an early form of breast cancer in which cancer cells have not yet spread to the regional lymph nodes. There is much interest in determining the relevant prognostic factors that can allocate node-negative patients into prognostic groups correlating with the risk of disease relapse and mortality following surgery. This risk index can then be used to inform the choice of therapy. The Cox regression model and Artificial Neural Networks (ANN), a Partial Logistic Artificial Neural Network with Automatic Relevance Determination (PLANN-ARD) are used in order to identify and interpret the prognostic group allocation. A monthly retrospective cohort study with 5-year follow-up is conducted in pathologically node-negative patients selected from two datasets collected from Manchester Christie Hospital, UK.

This paper describes an integrated clinical support system combining data entry and access with prognostic modelling, for use by the clinician, complemented by a patient information system tailored to the particular characteristics of the individual patient. The core of the system comprises a modelling methodology based on the PLANN-ARD neural network which combines risk staging with automatic rule generation to derive an explanation facility for the risk group allocation of each patient. The aim of the system is to promote better informed decision making on the part of both the clinician and the patient, exploiting the combined potential of analytical methodologies and the internet.

This chapter reviews classical training methods for multilayer neural networks. These methods are widely used for classification and function modelling tasks. Nevertheless, they show a number of flaws or drawbacks that should be addressed in the development of such systems. They work by searching the minimum of an error function which defines the optimal behaviour of the neural network. Different standard problems are used to show the capabilities of these models; in particular, we have benchmarked the algorithms in a nonlinear classification problem and in three function modelling problems.

There is much interest in rule extraction from neural networks and a plethora of different methods have been proposed for this purpose. We discuss the merits of pedagogical and decompositional approaches to rule extraction from trained neural networks, and show that some currently used methods for binary data comply with a theoretical formalism for extraction of Boolean rules from continuously valued logic. This formalism is extended into a generic methodology for rule extraction from smooth decision surfaces fitted to discrete or quantized continuous variables independently of the analytical structure of the underlying model, and in a manner that is efficient even for high input dimensions. This methodology is then tested with Monks' data, for which exact rules are obtained and to Wisconsin's breast cancer data, where a small number of high-order rules are identified whose discriminatory performance can be directly visualized.

This paper presents a Bayesian neural network for the analysis of competing risk (CR) data model. Based on a previously developed non-linear model namely partial logistic artificial neural network (PLANN) with automatic relevance determination (ARD), this paper proposes an extension for the flexible joint estimation of cause-specific hazards depending on both discrete and continuous covariates (PLANN-CR-ARD) and for censored data. The Bayesian analysis uses Gaussian priors for the neural network parameters and the likelihood function based on the competing risk data is identified as the cross-entropy function. The PLANN-CR-ARD model is illustrated with analyses of an intra-ocular melanoma dataset and comparison with the non-parametric Nelson-Alien estimates of the cause-specific cumulative hazards functions. (4 pages)

The performance of the non-linear prognostic model 'partial logistic artificial neural network with automatic relevance determination' (PLANN-ARD) is observed here and compared to a variant of the model that is trained using a compensating mechanism to account for the skewed distribution in the neural network target vector. The application dataset in this survival analysis is cancer patient information namely diagnosed with intra-ocular melanoma. The outcomes of the two models are compared with the empirical cumulative hazard curve derived from the Kaplan-Meier survival function for a particular population. Three forms of outcome from both the compensated and non compensated models are obtained: the network output itself, marginalised network output and the median of the network output distribution. (4 pages)