Search

In this paper we propose to use partial responses derived from an initial multilayer perceptron (MLP) to build an explanatory risk prediction model of in-hospital mortality in intensive care units (ICU). Traditionally, MLPs deliver higher performance than linear models such as multivariate logistic regression (MLR). However, MLPs interlink input variables in such a complex way that is not straightforward to explain how the outcome is influenced by inputs and/or input interactions. In this paper, we hypothesized that in some scenarios, such as when the data noise is significant or when the data is just marginally non-linear, we could find slightly more complex associations by obtaining MLP partial responses. That is, by letting change one variable at the time, while keeping constant the rest. Overall, we found that, although the MLR and MLP in-hospital mortality model performances were equivalent, the MLP could explain non-linear associations that otherwise the MLR had considered non-significant. We considered that, although deeming higher-other interactions as disposable noise could be a strong assumption, building explanatory models based on the MLP partial responses could still be more informative than on MLR.

This paper focuses on the scalability of the Fisher Information manifold by applying techniques of distributed computing. The main objective is to investigate methodologies to improve two bottlenecks associated with the measurement of distances in a Riemannian manifold formed by the Fisher Information metric. The first bottleneck is the quadratic increase in the number of pairwise distances. The second is the computation of global distances, approximated through a fully connected network of the observed pairwise distances, where the challenge is the computation of the all sources shortest path (ASSP). The scalable implementation for the pairwise distances is performed in Spark. The scalable global distance computations are addressed by applying the Dijkstra algorithm (SSSP) sequentially on two proposed networks based on prototypes that approximate the manifold. The proposed solutions are compared with a single-machine implementation in Matlab with experiments showing the first bottleneck solution is faster in Spark, but the distributed solutions for the second bottleneck is slower. Therefore, our conclusion is that the most appropriate method is a hybrid approach, where in terms of runtime and scalability a hybrid approach performs best; running the distributed method and the single-machine approach to solve bottleneck one then two, respectively.

Within genomic studies, a considerable amount of publications have reported SNP variants associated with COPD with little to no reproducibility. In this paper, we present a robust methodology which analyses a COPD cohort dataset using a genome-wide association study, additionally an investigation of the associated results using a variety of machine learning (ML) methods is performed. We use a logistic regression model to provide preliminary results and for further analysis we use machine learning models, RF, MLP, GLM and SVM. Within this study, indications of well established SNPs in previous publications occur in the preliminary results but fail to provide further indication of associative relationship when using ML methods for classification purposes. Results within this study show little to no predictive power after performing a robust methodology. These results indicate that a standardization of practice should be implemented to ensure the publication of false positive results is reduced and deterred. Further investigation of associative features should be considered a standard practice given the resulting information that can be provided with its' use.

In this paper, a novel application of machine learning algorithms including Neural Network architecture is presented for the prediction of flood severity. Floods are considered natural disasters that cause wide scale devastation to areas affected. The phenomenon of flooding is commonly caused by runoff from rivers and precipitation, specifically during periods of extremely high rainfall. Due to the concerns surrounding global warming and extreme ecological effects, flooding is considered a serious problem that has a negative impact on infrastructure and humankind. This paper attempts to address the issue of flood mitigation through the presentation of a new flood dataset, comprising 2000 annotated flood events, where the severity of the outcome is categorised according to 3 target classes, demonstrating the respective severities of floods. The paper also presents various types of machine learning algorithms for predicting flood severity and classifying outcomes into three classes, normal, abnormal, and high-risk floods. Extensive research indicates that artificial intelligence algorithms could produce enhancement when utilised for the pre-processing of flood data. These approaches helped in acquiring better accuracy in the classification techniques. Neural network architectures generally produce good outcomes in many applications, however, our experiments results illustrated that random forest classifier yields the optimal results in comparison with the benchmarked models.

Repeated experiences of negative emotions, such as stress, anger or anxiety, can have long-term consequences for health. These episodes of negative emotion can be associated with inflammatory changes in the body, which are clinically relevant for the development of disease in the long-term. However, the development of effective coping strategies can mediate this causal chain. The proliferation of ubiquitous and unobtrusive sensor technology supports an increased awareness of those physiological states associated with negative emotion and supports the development of effective coping strategies. Smartphone and wearable devices utilise multiple on-board sensors that are capable of capturing daily behaviours in a permanent and comprehensive manner, which can be used as the basis for self-reflection and insight. However, there are a number of inherent challenges in this application, including unobtrusive monitoring, data processing, and analysis. This paper posits a mobile Iifelogging platform that utilises wearable technology to monitor and classify levels of stress. A pilot study has been undertaken with six participants, who completed up to ten days of data collection. During this time, they wore a wearable device on the wrist during waking hours to collect instances of heart rate (HR) and Galvanic Skin Resistance (GSR). Preliminary data analysis was undertaken using three supervised machine learning algorithms: Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA) and Decision Tree (DT). An accuracy of 70% was achieved using the Decision Tree algorithm.

Glyphosate is a commonly used herbicide which carries some risks of ecotoxicity and has been shown to be harmful to human beings with high levels of exposure. Existing methods of glyphosate detection often struggle to achieve the level of sensitivity required to meet regulatory requirements without the use of complicated analytical methods with multiple intermediary steps. We propose the use of microwave spectroscopy to determine the concentration of glyphosate in aqueous solutions, using machine learning methodology to identify a minimum feature set for our model. The resulting model had a limit of detection of roughly 10−3mg/L, fitted values were significantly (Pearson's R = 0.8833, P

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.

Glioblastomas are one the most aggressive brain tumors. Their usual bad prognosis is due to the heterogeneity of their response to treatment and the lack of early and robust biomarkers to decide whether the tumor is responding to therapy. In this work, we propose the use of a semi-supervised methodology for source extraction to identify the sources representing tumor response to therapy, untreated/unresponsive tumor, and normal brain; and create nosological images of the response to therapy based on those sources. Fourteen mice were used to calculate the sources, and an independent test set of eight mice was used to further evaluate the proposed approach. The preliminary results obtained indicate that was possible to discriminate response and untreated/unresponsive areas of the tumor, and that the color-coded images allowed convenient tracking of response, especially throughout the course of therapy.

Support vector machines are highly flexible and generalizing mathematical models that can be used to build prediction models. Their success on a mathematical field is not followed by their application in practice due to their black-box nature. The RBF kernel is often used but the good performance cannot be accompanied by an interpretation of the results. We present a method to visualize the different components of an RBF kernel and propose a method to select the relevant ones. The proposed method is able to automatically detect important main and two-way interaction effects while still obtaining interpretable prediction models. The method is illustrated on a large dataset to predict the viability of pregnancies at the end of the first trimester based on initial scan findings. © 2013 IEEE. ispartof: pages:26-31 ispartof: Proc. of the IEEE symposium on computational intelligence and data mining pages:26-31 ispartof: IEEE-CIDM location:Singapore date:Apr - Apr 2013 status: published

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.

Survival analysis is an important part of medical statistics or for the study of failures in mechanical systems. The latter is called reliability analysis in engineering. This paper addresses the survival analysis in medical research. In this context of medical research, the survival analysis is frequently used to define prognostic indices for survival or recurrence of a disease, and to study outcome of treatment. Within the broader domain of medical survival analysis or more specifically for the study of patients with cancer disease, there are a number of clinical/prognostic techniques developed over the last half of century in many research groups, universities and hospitals around the globe. Despite numerous publications on each of these techniques, there is still a need of evaluating the numerical outputs of the different prognostic algorithms on identical medical datasets. Hence the need to benchmark the mathematical algorithms against the available medical and other cancer datasets held by various research groups from universities or hospitals. The main scope of a benchmark process is to specifically inform and to provide prognostic advice and choice of post-operative adjuvant therapy. It is also essential to define data acquisition protocols, to create the necessary databases on which the benchmarked methodologies can be tested. In case there are several research groups which are taking part in a benchmark study, then they form a multicentre study. This paper presents the steps to be followed in order to realize such a multicentre study.

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.

The challenge of sustainable healthcare requires significant changes to the structure of healthcare delivery, with greater emphasis on prevention and pro-active, personalised care. This, in turn, relies for its effectiveness on detailed analysis of the evidence contained in large data bases. This paper describes a powerful and general analytical approach to derive insights from complex health and behavioural data, and to underpin evidence-based scenario analysis for commissioning of clinically efficient and cost effective health interventions.

Pattern Recognition and Data Mining can provide invaluable insights in the field of neuro oncology. This is because the clinical analysis of brain tumors requires the use of non-invasive methods that generate complex data in electronic format. Magnetic resonance, in the modalities of imaging and spectroscopy, is one of these methods that has been widely applied to this purpose. The heterogeneity of the tissue in the brain volumes analyzed by magnetic resonance remains a challenge in terms of pathological area delimitation. In this brief paper, we show that the Convex-Nonnegative Matrix Factorization technique can be used to extract MRS signal sources that are extremely tissue type-specific and that can be used to delimit these pathological areas with great accuracy.

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.

The clinical assessment of human brain tumors requires the use of non-invasive information measurement technologies, usually from the modalities of imaging or spectroscopy. The latter may provide insight into the tumor metabolism. The Magnetic Resonance Spectroscopy (MRS) signal is the result of the combination of heterogeneous signal sources. In this study, we investigate the use of two spectral decomposition techniques for the identification of such sources in MRS from brain tumors collected in an international database, and for a number of different diagnostic problems.

Clustering to identify inherent structure is an important first step in data exploration. The k-means algorithm is a popular choice, but K-means is not generally appropriate for categorical data. A specific extension of k-means for categorical data is the k-modes algorithm. Both of these partition clustering methods are sensitive to the initialization of prototypes, which creates the difficulty of selecting the best solution for a given problem. In addition, selecting the number of clusters can be an issue. Further, the k-modes method is especially prone to instability when presented with ‘noisy’ data, since the calculation of the mode lacks the smoothing effect inherent in the calculation of the mean. This is often the case with real-world datasets, for instance in the domain of Public Health, resulting in solutions that can be radically different depending on the initialization and therefore lead to different interpretations. This paper presents two methodologies. The first addresses sensitivity to initializations using a generic landscape mapping of k-mode solutions. The second methodology utilizes the landscape map to stabilize the partition clusters for discrete data, by drawing a consensus sample in order to separate signal from noise components. Results are presented for the benchmark soybean disease dataset, an artificially generated dataset and a case study involving Public Health data.

A key question in medical decision support is how best to visualise a patient database, with especial reference to cohort labelling, whether this is an indicator function for classification or a cluster index. We propose the use of the kernel trick to visualise complete patient databases, in low-dimensional projections, with class labelling, given a non-linear classifier of choice. The results show that this method is useful both to see how individual patient cases relate to each other with reference to the classification boundary, and also to obtain a visual indication of the separation that can be obtained with difference choices of kernel functions.

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.

A hybrid lossy image compression technique using classified vector quantiser and singular value decomposition is presented for the efficient representation of medical magnetic resonance – brain images. The proposed method is called hybrid classified vector quantisation. It involves a simple yet efficient classifier based gradient method in the spatial domain which employs only one threshold to determine the class of the input image block, and uses three AC coefficients of the discrete cosine transform coefficients to determine the orientation of the block without employing any threshold that results in high fidelity medical compressed images. Singular value decomposition was used to generate the classified codebooks. The proposed technique was benchmarked with JPEG-2000 standard. Simulation results indicate that the proposed approach can reconstruct high visual quality images with higher Peak Signal-to Noise-Ratio than the benchmarked technique and also meet the legal requirement of medical images archiving.

An efficient adaptive lossy image compression technique using classified vector quantiser and singular value decomposition for compression of medical magnetic resonance-brain images is presented. The proposed method is called adaptive hybrid classified vector quantisation. A simple but efficient classifier based gradient method without employing any threshold to determine the class of the input image block in the spatial domain that results in a high- fidelity medical compressed image was utilised. The proposed technique was benchmarked with JPEG-2000 standard. Simulation results indicated that the proposed approach can reconstruct high visual quality images with higher Peak Signal-to Noise-Ratio than the benchmarked technique, and also meet the legal requirement of medical image archiving.

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).

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

The evaluation of generic non-linear models for censored data needs to address the two complementary requirements in the software development life-cycle, of validation and verification. The former involves making a rigorous assessment of predictive accuracy in prognostic modelling and the latter is interpreted in this paper as comprising two different stages, namely model selection and rule-based interpretation of the composition of prognostic risk groups. With reference to prognostic performance is survival modelling the well-known ROC framework has recently been extended to a threshold independent, time-dependent performance index to quantify the predictive accuracy of censored data models, termed the C' index, which is briefly described. The rule-based framework for direct validation of risk group allocation against expert domain knowledge, uses low-order Boolean rules to approximate the response surfaces generated by analytical inference models. In the case of censored data, this approach serves to characterise the allocation of patients into risk groups generated by a risk staging index. Furthermore, the low-order rules define low-dimensional sub-spaces where individual data points can be directly visualised in relation to the decision boundaries for their risk group. Taken together, the quantitative performance index, Boolean explanatory rules and direct visualisation of the data, define a consistent and transparent validation framework based on triangulation of information. This information can be included in decision support systems.

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.

A novel image compression technique using classified vector quantiser and singular value decomposition is presented for the efficient representation of still images. The proposed method is called hybrid classified vector quantisation. It involves a simple, but efficient, classifier based gradient method in the spatial domain which employs only one threshold to determine the class of the input image block, and uses three AC coefficients of the discrete cosine transform coefficients to determine the orientation of the block without employing any threshold. Singular value decomposition was used to generate the classified codebooks. The proposed technique was benchmarked with the standard vector quantiser generated using the k-means algorithm, and JPEG-2000. Simulation results indicated that the proposed approach alleviates edge degradation and can reconstruct good visual quality images with higher peak signal-to noise-ratio than the benchmarked techniques.

This paper describes iShakti, a real-world, Intelligent, Interactive and Adaptive Web application. At present, iShakti is deployed across 1000 rural kiosks in India, covering 5000 villages and reaching 1 million people. Further scale up is underway, expected to cover tens of thousands of villages within the next 2 years. iShakti is a ?virtual information and marketing channel?, deploying leading-edge technology in a developing-world environment. It allows rich interactions with people in previously ?mediadark? regions, with easy access to high-value community development services coupled with engaging and scalable market and brand development activities. The impact is already being felt -- iShakti is giving some of the most deprived and disempowered people more choice and control over their lives, and providing significant independent revenue for the iShakti entrepreneurs. Computational Intelligence is both in the design as well as the personalisation and synchronisation algorithms. The project was nominated as a finalist of the Stockholm Challenge (Economic Development category), an international award for ICT projects in ?underserved? regions of the world [1].

Process control using a model based structure incorporating a neural network is examined by application to the control of a real pilot-scale process exhibiting non-linearities and typical disturbances. Initially, a methodology for identibing an accurate neural network process model from plant data is described and practical aspects of applying the techniques are discussed. It is shown that the approach leads to a neural network description of the process dynamics that is suficiently accurate to be used independently from the process, emulating the process response from only process input information. The main success of the approach is the use of a novel coding technique for representing data in the network. The network model is incorporated into a model predictive control structure and on-line results illustrate the improvements in control performance that can be achieved compared to conventional PI control. Additionally, an insight into the dynamics and stability of the neural control scheme is obtained in a novel application of linear system identification techniques.

Oceanographic exploration is one of the fast emerging applications of robotics. The design of Underwater Robotic Vehicles (URV’s), is as challenging as for land based ones. The dificulties in modelling an URV and its hazardous environment restrict the use of conventional controllers. In this paper the application of Neural networks (NNs) for the modelling and control of a prototype URV, which is an example of a system containing non-linearities, is described. A NN model is developed and then incorporated into a predictive control strategy which it is evaluated both in simulation and on-line. Results are shown for both the modelling and control of the system, including hybrid control strategies which combine neural predictive with conventional three term controllers.

Classij-ication of cancer and normal animal tissues is carried out on the basis of their 'H Nuclear Magnetic Resonance (NMR) spectra with neural networks trained by Back-Error Propagation (BEP), using two direrent costfunctions. A log-likelihood costfinction is shown to result in accurate out-of-sample generalisation with a smaller network than the usual Least Mean Squared (ZMS) error. ntejirst step in the interpretation of the operation of neural networks is to quantiJjr the relevance of the input parameters to the diagnosis of each tissue class. Two techniques for achieving this are investigated, namely the Jacobian method and a logarithmic sensitivity matrix. The latter is demonstrated to result in a clearer signature which is consistent across direrent network architectures and also broadly in agreement with conventional statistical correlations.

Industrial plasma etcher operation in IC manufacture is often carried out through process recipes, with little or no adjustment based on feedback of important process outputs. The recipes created are not transferable between different etch chemistries or from reactor to reactor and the etch profiles resulting from a given recipe may vary with time as conditions within the reactor change. Two major difficulties with implementing closed-loop control are the need for in situ measurements of important process variables and nonlinear modelling of the process responses to the coupled control parameters. This paper concerns the use of chromatic monitoring as a process feedback measurement and compares two methodologies for the reactive ion etching (RIE) process modelling.

Oceanographic exploration is one of the fast emerging applications of robotics, and the design of controllers for Underwater Robotic Vehicles (URVs) is as challenging as for land based ones. The difficulties in modelling an URV and its hazardous environment restrict the use of conventional controllers. This paper presents an approach for control and system identification of a prototype URV, as an example of a system containing severe non-linearities, using neural networks (NNs). NNs models are developed and then incorporated into a predictive control strategy which are evaluated on-line. Results are shown for both the modelling and control of the system including hybrid control strategies which combine neural predictive with conventional three term controllers. >

The prognostic assessment of breast cancer patients following surgery to excise the tumour and other pathological tissue, is an important consideration in determining the most appropriate course of adjuvant therapy for the patient. The paper compares the relative predictive performance of a neural network and an established statistical method, on the basis of their ROC and Kaplan-Meier curves, using information available at the time of surgery.

We present a data set obtained from neurophysiological recordings made from single neurons in the human basal ganglia in relation to movements of the limbs during surgical procedures for Parkinson's disease. Data from both the pre-movement (movement preparation) phase and the post-movement phase were recorded in 25 cells in 9 patients. Using the generative topographic mapping algorithm (GTM), we show that the data corresponding to individual cells cluster in different parts of the map, which suggests that different cells have their own distinct firing pattern. The self-organising map as implemented by GTM appears to be a useful tool for visualising the responses of these neurons in relation to movement.

Nuclear magnetic resonance (NMR) spectroscopy has considerable potential for non-invasive characterisation of tissue biochemistry and the diagnosis of tissue abnormalities, ranging from focal lesions in the brain, to tumours in any area of the body to assessing effect of HIV damage. However, the realisation of the full clinical potential NMR spectroscopy will depend on extracting information from the spectra directly and specifically related to the biochemistry of different tissue types under various normal and pathological circumstances. This paper reviews the progress made in the application of neural network analysis to the automatic characterisation of NMR data, raising some key issues and providing a perspective of the future of this technology.

Artificial neural networks are popularly used as universal nonlinear inference models. However, they suffer from two major drawbacks. Their operation is opaque because of the distributed nature of the representations they form, and this makes it different to interpret what they do. Worse still, there are no clearly accepted models of generality which makes it difficult to demonstrate reliability when applied to future data. In this paper neural networks generate hypotheses concerning interaction terms which are integrated into standard statistical models that are linear in the parameters, where the significance of the nonlinear terms and the generality of the model, can be assured using well established statistical tests.

The paper presents a method for the 3D visualization of the structure of radial basis function networks using traditional and novel methods of dimensionality reduction. This method allows the visualization of basis function characteristics (centers and widths) along with second level weights. To facilitate the interpretation of a wide variety of high dimensional problems, several forms of projections into 20 or 30 spaces can be used interactively. The traditional methods of principal component analysis and Fisher's linear discriminant are used as well as a novel linear projection method.

The authors describe two different approaches to object orientation detection: one method uses first- and second-order moments, while the other uses a multilayer perceptron network trained by back error propagation. A comparison between these methods shows that the neural network is able to generalize the trained orientations for different classes of objects and affords better control in determining the orientation at the pick-up point. Maximum orientation resolution is achieved economically by using a form of coarse coding, in which the output excitations for all orientations are spread out among neighboring output nodes. >

The classification of two-dimensional binary images by artificial neural networks, irrespective of their position, orientation, and size, is investigated using two complementary methods. Third-order networks were used first. Invariance under all three transformations was achieved by grouping triplets of pixels into appropriate equivalent classes. A suitable reduction in the number of weights then resulted in economical networks which exhibit high recognition rates under all transformations simultaneously, together with robustness against local distortions. The performance of these networks was tested on invariant classification of both typed and handwritten numerals. It was found to be superior to that achieved by the more traditional method of preprocessing that image using moments. Zernike moments were used because they are known to be well suited for pattern classification. In both cases, the network was trained by back-error propagation. >

Gynaecologists and obstetricians visually interpret cardiotocography (CTG) traces using the International Federation of Gynaecology and Obstetrics (FIGO) guidelines to assess the wellbeing of the foetus during antenatal care. This approach has raised concerns among professionals with regards to inter- and intra-variability where clinical diagnosis only has a 30\% positive predictive value when classifying pathological outcomes. Machine learning models, trained with FIGO and other user derived features extracted from CTG traces, have been shown to increase positive predictive capacity and minimise variability. This is only possible however when class distributions are equal which is rarely the case in clinical trials where case-control observations are heavily skewed in favour of normal outcomes. Classes can be balanced using either synthetic data derived from resampled case training data or by decreasing the number of control instances. However, this either introduces bias or removes valuable information. Concerns have also been raised regarding machine learning studies and their reliance on manually handcrafted features. While this has led to some interesting results, deriving an optimal set of features is considered to be an art as well as a science and is often an empirical and time consuming process. In this paper, we address both of these issues and propose a novel CTG analysis methodology that a) splits CTG time-series signals into n-size windows with equal class distributions, and b) automatically extracts features from time-series windows using a one dimensional convolutional neural network (1DCNN) and multilayer perceptron (MLP) ensemble. Collectively, the proposed approach normally distributes classes and removes the need to handcrafted features from CTG traces., Comment: 11 Pages, 12 Figures (excluding profile pictures), accepted for publication in IEEE Transactions in Emerging Topics in Computational Intelligence

Algorithms in the Adaptive Resonance Theory (ART) family adapt to structural changes in data as new information presents, making it an exciting candidate for dynamic online clustering of big health data. Its use however has largely been restricted to the signal processing field. In this paper we introduce an refinement of the ART2-A method within an adapted separation and concordance (SeCo) framework which has been shown to identify stable and reproducible solutions from repeated initialisations that also provides evidence for an appropriate number of initial clusters that best calibrates the algorithm with the data presented. The results show stable, reproducible solutions for a mix of real-world heath related datasets and well known benchmark datasets, selecting solutions which better represent the underlying structure of the data than using a single measure of separation. The scalability of the method and it's facility for dynamic online clustering makes it suitable for finding structure in big data.

A major challenge in delivering reliable and trustworthy computational intelligence for practical applications in clinical medicine is interpretability. This aspect of machine learning is a major distinguishing factor compared with traditional statistical models for the stratification of patients, which typically use rules or a risk score identified by logistic regression.\ud We show how functions of one and two variables can be extracted from pre-trained machine learning models using anchored Analysis of Variance (ANOVA) decompositions. This enables complex interaction terms to be filtered out by aggressive regularisation using the Least Absolute Shrinkage and Selection Operator (LASSO) resulting in a sparse model with comparable or even better performance than the original pre-trained black-box. \ud Besides being theoretically well-founded, the decomposition of a black-box multivariate probabilistic binary classifier into a General Additive Model (GAM) comprising a linear combination of non-linear functions of one or two variables provides full interpretability. In effect this extends logistic regression into non-linear modelling without the need for manual intervention by way of variable transformations, using the pre-trained model as a seed.\ud The application of the proposed methodology to existing machine learning models is demonstrated using the Multi-Layer Perceptron (MLP), Support Vector Machine (SVM), Random Forests (RF) and Gradient Boosting Machines (GBM), to model a data frame from a well-known benchmark dataset available from Physionet, the Medical Information Mart for Intensive Care (MIMIC-III). Both the classification performance and plausibility of clinical interpretation compare favourably with other state-of-the-art sparse models namely Sparse Additive Models (SAM) and the Explainable Boosting Machine (EBM).

Type 2 Diabetes condition is a multifactorial disorder involves the convergence of genetics, environment, diet and lifestyle risk factors. This paper investigates genetic and conventional (clinical, sociodemographic) risk factors and their predictive power in classifying Type 2 Diabetes. Six statistically significant Single Nucleotide Polymorphisms (SNPs) associated with Type 2 Diabetes are derived by conducting logistic association analysis. The derived SNPs in addition to conventional risk factors are used to model supervised machine learning algorithms to classify cases and controls in genome wide association studies (GWAS). Models are trained using genetic variable analysis, genetic and conventional variable analysis, and conventional variable analysis. The results demonstrate of the three models, higher predictive capacity is evident when genetic and conventional predictors are combined. Using a Random Forest classifier, the Area Under the Curve=73.96%, Sensitivity=68.42 %, and Specificity=78.67%.