Using data mining for prediction of hospital length of stay: an application of the CRISP-DM Methodology

Hospitals are nowadays collecting vast amounts of data related with patient records. All this data hold valuable knowledge that can be used to improve hospital decision making. Data mining techniques aim precisely at the extraction of useful knowledge from raw data. This work describes an implementation of a medical data mining project approach based on the CRISP-DM methodology. Recent real-world data, from 2000 to 2013, were collected from a Portuguese hospital and related with inpatient hospitalization. The goal was to predict generic hospital Length Of Stay based on indicators that are commonly available at the hospitalization process (e.g., gender, age, episode type, medical specialty). At the data preparation stage, the data were cleaned and variables were selected and transformed, leading to 14 inputs. Next, at the modeling stage, a regression approach was adopted, where six learning methods were compared: Average Prediction, Multiple Regression, Decision Tree, Artificial Neural Network ensemble, Support Vector Machine and Random Forest. The best learning model was obtained by the Random Forest method, which presents a high quality coefficient of determination value (0.81). This model was then opened by using a sensitivity analysis procedure that revealed three influential input attributes: the hospital episode type, the physical service where the patient is hospitalized and the associated medical specialty. Such extracted knowledge confirmed that the obtained predictive model is credible and with potential value for supporting decisions of hospital managers.

Search for the Standard Model Higgs boson produced in association with top quarks and decaying into bb in pp collisions at √s = 8 TeV with the ATLAS detector

A search for the Standard Model Higgs boson produced in association with a pair of top quarks, tt¯H, is presented. The analysis uses 20.3 fb−1 of pp collision data at s√ = 8 TeV, collected with the ATLAS detector at the Large Hadron Collider during 2012. The search is designed for the H to bb¯ decay mode and uses events containing one or two electrons or muons. In order to improve the sensitivity of the search, events are categorised according to their jet and b-tagged jet multiplicities. A neural network is used to discriminate between signal and background events, the latter being dominated by tt¯+jets production. In the single-lepton channel, variables calculated using a matrix element method are included as inputs to the neural network to improve discrimination of the irreducible tt¯+bb¯ background. No significant excess of events above the background expectation is found and an observed (expected) limit of 3.4 (2.2) times the Standard Model cross section is obtained at 95% confidence level. The ratio of the measured tt¯H signal cross section to the Standard Model expectation is found to be μ=1.5±1.1 assuming a Higgs boson mass of 125 GeV.

Identifying the neural correlates of executive functions in early cerebral microangiopathy: a combined VBM and DTI study.

Cerebral microangiopathy (CMA) has been associated with executive dysfunction and fronto-parietal neural network disruption. Advances in magnetic resonance imaging allow more detailed analyses of gray (e.g., voxel-based morphometry-VBM) and white matter (e.g., diffusion tensor imaging-DTI) than traditional visual rating scales. The current study investigated patients with early CMA and healthy control subjects with all three approaches. Neuropsychological assessment focused on executive functions, the cognitive domain most discussed in CMA. The DTI and age-related white matter changes rating scales revealed convergent results showing widespread white matter changes in early CMA. Correlations were found in frontal and parietal areas exclusively with speeded, but not with speed-corrected executive measures. The VBM analyses showed reduced gray matter in frontal areas. All three approaches confirmed the hypothesized fronto-parietal network disruption in early CMA. Innovative methods (DTI) converged with results from conventional methods (visual rating) while allowing greater spatial and tissue accuracy. They are thus valid additions to the analysis of neural correlates of cognitive dysfunction. We found a clear distinction between speeded and nonspeeded executive measures in relationship to imaging parameters. Cognitive slowing is related to disease severity in early CMA and therefore important for early diagnostics.

Objective analysis of gait coordination for total knee arthroplasty outcome

Introduction: Coordination is a strategy chosen by the central nervous system to control the movements and maintain stability during gait. Coordinated multi-joint movements require a complex interaction between nervous outputs, biomechanical constraints, and pro-prioception. Quantitatively understanding and modeling gait coordination still remain a challenge. Surgeons lack a way to model and appreciate the coordination of patients before and after surgery of the lower limbs. Patients alter their gait patterns and their kinematic synergies when they walk faster or slower than normal speed to maintain their stability and minimize the energy cost of locomotion. The goal of this study was to provide a dynamical system approach to quantitatively describe human gait coordination and apply it to patients before and after total knee arthroplasty. Methods: A new method of quantitative analysis of interjoint coordination during gait was designed, providing a general model to capture the whole dynamics and showing the kinematic synergies at various walking speeds. The proposed model imposed a relationship among lower limb joint angles (hips and knees) to parameterize the dynamics of locomotion of each individual. An integration of different analysis tools such as Harmonic analysis, Principal Component Analysis, and Artificial Neural Network helped overcome high-dimensionality, temporal dependence, and non-linear relationships of the gait patterns. Ten patients were studied using an ambulatory gait device (Physilog®). Each participant was asked to perform two walking trials of 30m long at 3 different speeds and to complete an EQ-5D questionnaire, a WOMAC and Knee Society Score. Lower limbs rotations were measured by four miniature angular rate sensors mounted respectively, on each shank and thigh. The outcomes of the eight patients undergoing total knee arthroplasty, recorded pre-operatively and post-operatively at 6 weeks, 3 months, 6 months and 1 year were compared to 2 age-matched healthy subjects. Results: The new method provided coordination scores at various walking speeds, ranged between 0 and 10. It determined the overall coordination of the lower limbs as well as the contribution of each joint to the total coordination. The difference between the pre-operative and post-operative coordination values were correlated with the improvements of the subjective outcome scores. Although the study group was small, the results showed a new way to objectively quantify gait coordination of patients undergoing total knee arthroplasty, using only portable body-fixed sensors. Conclusion: A new method for objective gait coordination analysis has been developed with very encouraging results regarding the objective outcome of lower limb surgery.

An empirical hierarchical Bayesian unification of occupational exposure assessment methods.

In occupational exposure assessment of airborne contaminants, exposure levels can either be estimated through repeated measurements of the pollutant concentration in air, expert judgment or through exposure models that use information on the conditions of exposure as input. In this report, we propose an empirical hierarchical Bayesian model to unify these approaches. Prior to any measurement, the hygienist conducts an assessment to generate prior distributions of exposure determinants. Monte-Carlo samples from these distributions feed two level-2 models: a physical, two-compartment model, and a non-parametric, neural network model trained with existing exposure data. The outputs of these two models are weighted according to the expert's assessment of their relevance to yield predictive distributions of the long-term geometric mean and geometric standard deviation of the worker's exposure profile (level-1 model). Bayesian inferences are then drawn iteratively from subsequent measurements of worker exposure. Any traditional decision strategy based on a comparison with occupational exposure limits (e.g. mean exposure, exceedance strategies) can then be applied. Data on 82 workers exposed to 18 contaminants in 14 companies were used to validate the model with cross-validation techniques. A user-friendly program running the model is available upon request.

Hybrid coordination of reinforcement learning-based behaviors for AUV control

This paper proposes a hybrid coordination method for behavior-based control architectures. The hybrid method takes advantages of the robustness and modularity in competitive approaches as well as optimized trajectories in cooperative ones. This paper shows the feasibility of applying this hybrid method with a 3D-navigation to an autonomous underwater vehicle (AUV). The behaviors are learnt online by means of reinforcement learning. A continuous Q-learning implemented with a feed-forward neural network is employed. Realistic simulations were carried out. The results obtained show the good performance of the hybrid method on behavior coordination as well as the convergence of the behaviors

Efficient learning of reactive robot behaviors with a Neural-Q_learning approach

The purpose of this paper is to propose a Neural-Q_learning approach designed for online learning of simple and reactive robot behaviors. In this approach, the Q_function is generalized by a multi-layer neural network allowing the use of continuous states and actions. The algorithm uses a database of the most recent learning samples to accelerate and guarantee the convergence. Each Neural-Q_learning function represents an independent, reactive and adaptive behavior which maps sensorial states to robot control actions. A group of these behaviors constitutes a reactive control scheme designed to fulfill simple missions. The paper centers on the description of the Neural-Q_learning based behaviors showing their performance with an underwater robot in a target following task. Real experiments demonstrate the convergence and stability of the learning system, pointing out its suitability for online robot learning. Advantages and limitations are discussed

Semi-online neural-Q_leaming for real-time robot learning

Reinforcement learning (RL) is a very suitable technique for robot learning, as it can learn in unknown environments and in real-time computation. The main difficulties in adapting classic RL algorithms to robotic systems are the generalization problem and the correct observation of the Markovian state. This paper attempts to solve the generalization problem by proposing the semi-online neural-Q_learning algorithm (SONQL). The algorithm uses the classic Q_learning technique with two modifications. First, a neural network (NN) approximates the Q_function allowing the use of continuous states and actions. Second, a database of the most representative learning samples accelerates and stabilizes the convergence. The term semi-online is referred to the fact that the algorithm uses the current but also past learning samples. However, the algorithm is able to learn in real-time while the robot is interacting with the environment. The paper shows simulated results with the "mountain-car" benchmark and, also, real results with an underwater robot in a target following behavior

Morphometric shape analysis using learning vector quantization neural networks - an example distinguishing two microtine vole species

Closely related species may be very difficult to distinguish morphologically, yet sometimes morphology is the only reasonable possibility for taxonomic classification. Here we present learning-vector-quantization artificial neural networks as a powerful tool to classify specimens on the basis of geometric morphometric shape measurements. As an example, we trained a neural network to distinguish between field and root voles from Procrustes transformed landmark coordinates on the dorsal side of the skull, which is so similar in these two species that the human eye cannot make this distinction. Properly trained neural networks misclassified only 3% of specimens. Therefore, we conclude that the capacity of learning vector quantization neural networks to analyse spatial coordinates is a powerful tool among the range of pattern recognition procedures that is available to employ the information content of geometric morphometrics.

Identificación de puntos de pesca en los viajes pesqueros de anchoveta monitoreados por satélites mediante redes neuronales artificiales

La sostenibilidad de los recursos marinos y de su ecosistema hace necesario un manejo responsable de las pesquerías. Conocer la distribución espacial del esfuerzo pesquero y en particular de las operaciones de pesca es indispensable para mejorar el monitoreo pesquero y el análisis de la vulnerabilidad de las especies frente a la pesca. Actualmente en la pesquería de anchoveta peruana, se recoge información del esfuerzo y capturas mediante un programa de observadores a bordo, pero esta solo representa una muestra de 2% del total de viajes pesqueros. Por otro lado, se dispone de información por cada hora (en promedio) de la posición de cada barco de la flota gracias al sistema de seguimiento satelital de las embarcaciones (VMS), aunque en estos no se señala cuándo ni dónde ocurrieron las calas. Las redes neuronales artificiales (ANN) podrían ser un método estadístico capaz de inferir esa información, entrenándose en una muestra para la cual sí conocemos las posiciones de calas (el 2% anteriormente referido), estableciendo relaciones analíticas entre las calas y ciertas características geométricas de las trayectorias observadas por el VMS y así, a partir de las últimas, identificar la posición de las operaciones de pesca. La aplicación de la red neuronal requiere un análisis previo que examine la sensibilidad de la red a variaciones en sus parámetros y bases de datos de entrenamiento, y que nos permita desarrollar criterios para definir la estructura de la red e interpretar sus resultados de manera adecuada. La problemática descrita en el párrafo anterior, aplicada específicamente a la anchoveta (Engraulis ringens) es detalllada en el primer capítulo, mientras que en el segundo se hace una revisión teórica de las redes neuronales. Luego se describe el proceso de construcción y pre-tratamiento de la base de datos, y definición de la estructura de la red previa al análisis de sensibilidad. A continuación se presentan los resultados para el análisis en los que obtenemos una estimación del 100% de calas, de las cuales aproximadamente 80% están correctamente ubicadas y 20% poseen un error de ubicación. Finalmente se discuten las fortalezas y debilidades de la técnica empleada, de métodos alternativos potenciales y de las perspectivas abiertas por este trabajo.

Bayesian image reconstruction with space-variant noise suppression

In this paper we present a Bayesian image reconstruction algorithm with entropy prior (FMAPE) that uses a space-variant hyperparameter. The spatial variation of the hyperparameter allows different degrees of resolution in areas of different statistical characteristics, thus avoiding the large residuals resulting from algorithms that use a constant hyperparameter. In the first implementation of the algorithm, we begin by segmenting a Maximum Likelihood Estimator (MLE) reconstruction. The segmentation method is based on using a wavelet decomposition and a self-organizing neural network. The result is a predetermined number of extended regions plus a small region for each star or bright object. To assign a different value of the hyperparameter to each extended region and star, we use either feasibility tests or cross-validation methods. Once the set of hyperparameters is obtained, we carried out the final Bayesian reconstruction, leading to a reconstruction with decreased bias and excellent visual characteristics. The method has been applied to data from the non-refurbished Hubble Space Telescope. The method can be also applied to ground-based images.

Molecular modeling of peptide-MHC class I complexes : applications to peptide based immunotherapy

Summary The specific CD8+ T cell immune response against tumors relies on the recognition by the T cell receptor (TCR) on cytotoxic T lymphocytes (CTL) of antigenic peptides bound to the class I major histocompatibility complex (MHC) molecule. Such tumor associated antigenic peptides are the focus of tumor immunotherapy with peptide vaccines. The strategy for obtaining an improved immune response often involves the design of modified tumor associated antigenic peptides. Such modifications aim at creating higher affinity and/or degradation resistant peptides and require precise structures of the peptide-MHC class I complex. In addition, the modified peptide must be cross-recognized by CTLs specific for the parental peptide, i.e. preserve the structure of the epitope. Detailed structural information on the modified peptide in complex with MHC is necessary for such predictions. In this thesis, the main focus is the development of theoretical in silico methods for prediction of both structure and cross-reactivity of peptide-MHC class I complexes. Applications of these methods in the context of immunotherapy are also presented. First, a theoretical method for structure prediction of peptide-MHC class I complexes is developed and validated. The approach is based on a molecular dynamics protocol to sample the conformational space of the peptide in its MHC environment. The sampled conformers are evaluated using conformational free energy calculations. The method, which is evaluated for its ability to reproduce 41 X-ray crystallographic structures of different peptide-MHC class I complexes, shows an overall prediction success of 83%. Importantly, in the clinically highly relevant subset of peptide-HLAA*0201 complexes, the prediction success is 100%. Based on these structure predictions, a theoretical approach for prediction of cross-reactivity is developed and validated. This method involves the generation of quantitative structure-activity relationships using three-dimensional molecular descriptors and a genetic neural network. The generated relationships are highly predictive as proved by high cross-validated correlation coefficients (0.78-0.79). Together, the here developed theoretical methods open the door for efficient rational design of improved peptides to be used in immunotherapy. Résumé La réponse immunitaire spécifique contre des tumeurs dépend de la reconnaissance par les récepteurs des cellules T CD8+ de peptides antigéniques présentés par les complexes majeurs d'histocompatibilité (CMH) de classe I. Ces peptides sont utilisés comme cible dans l'immunothérapie par vaccins peptidiques. Afin d'augmenter la réponse immunitaire, les peptides sont modifiés de façon à améliorer l'affinité et/ou la résistance à la dégradation. Ceci nécessite de connaître la structure tridimensionnelle des complexes peptide-CMH. De plus, les peptides modifiés doivent être reconnus par des cellules T spécifiques du peptide natif. La structure de l'épitope doit donc être préservée et des structures détaillées des complexes peptide-CMH sont nécessaires. Dans cette thèse, le thème central est le développement des méthodes computationnelles de prédiction des structures des complexes peptide-CMH classe I et de la reconnaissance croisée. Des applications de ces méthodes de prédiction à l'immunothérapie sont également présentées. Premièrement, une méthode théorique de prédiction des structures des complexes peptide-CMH classe I est développée et validée. Cette méthode est basée sur un échantillonnage de l'espace conformationnel du peptide dans le contexte du récepteur CMH classe I par dynamique moléculaire. Les conformations sont évaluées par leurs énergies libres conformationnelles. La méthode est validée par sa capacité à reproduire 41 structures des complexes peptide-CMH classe I obtenues par cristallographie aux rayons X. Le succès prédictif général est de 83%. Pour le sous-groupe HLA-A*0201 de complexes de grande importance pour l'immunothérapie, ce succès est de 100%. Deuxièmement, à partir de ces structures prédites in silico, une méthode théorique de prédiction de la reconnaissance croisée est développée et validée. Celle-ci consiste à générer des relations structure-activité quantitatives en utilisant des descripteurs moléculaires tridimensionnels et un réseau de neurones couplé à un algorithme génétique. Les relations générées montrent une capacité de prédiction remarquable avec des valeurs de coefficients de corrélation de validation croisée élevées (0.78-0.79). Les méthodes théoriques développées dans le cadre de cette thèse ouvrent la voie du design de vaccins peptidiques améliorés.

Model-Free Reconstruction of Excitatory Neuronal Connectivity from Calcium Imaging Signals

A systematic assessment of global neural network connectivity through direct electrophysiological assays has remained technically infeasible, even in simpler systems like dissociated neuronal cultures. We introduce an improved algorithmic approach based on Transfer Entropy to reconstruct structural connectivity from network activity monitored through calcium imaging. We focus in this study on the inference of excitatory synaptic links. Based on information theory, our method requires no prior assumptions on the statistics of neuronal firing and neuronal connections. The performance of our algorithm is benchmarked on surrogate time series of calcium fluorescence generated by the simulated dynamics of a network with known ground-truth topology. We find that the functional network topology revealed by Transfer Entropy depends qualitatively on the time-dependent dynamic state of the network (bursting or non-bursting). Thus by conditioning with respect to the global mean activity, we improve the performance of our method. This allows us to focus the analysis to specific dynamical regimes of the network in which the inferred functional connectivity is shaped by monosynaptic excitatory connections, rather than by collective synchrony. Our method can discriminate between actual causal influences between neurons and spurious non-causal correlations due to light scattering artifacts, which inherently affect the quality of fluorescence imaging. Compared to other reconstruction strategies such as cross-correlation or Granger Causality methods, our method based on improved Transfer Entropy is remarkably more accurate. In particular, it provides a good estimation of the excitatory network clustering coefficient, allowing for discrimination between weakly and strongly clustered topologies. Finally, we demonstrate the applicability of our method to analyses of real recordings of in vitro disinhibited cortical cultures where we suggest that excitatory connections are characterized by an elevated level of clustering compared to a random graph (although not extreme) and can be markedly non-local.

Artificial neural networks applied for soil class prediction in mountainous landscape of the Serra do Mar¹

Soil information is needed for managing the agricultural environment. The aim of this study was to apply artificial neural networks (ANNs) for the prediction of soil classes using orbital remote sensing products, terrain attributes derived from a digital elevation model and local geology information as data sources. This approach to digital soil mapping was evaluated in an area with a high degree of lithologic diversity in the Serra do Mar. The neural network simulator used in this study was JavaNNS and the backpropagation learning algorithm. For soil class prediction, different combinations of the selected discriminant variables were tested: elevation, declivity, aspect, curvature, curvature plan, curvature profile, topographic index, solar radiation, LS topographic factor, local geology information, and clay mineral indices, iron oxides and the normalized difference vegetation index (NDVI) derived from an image of a Landsat-7 Enhanced Thematic Mapper Plus (ETM+) sensor. With the tested sets, best results were obtained when all discriminant variables were associated with geological information (overall accuracy 93.2 - 95.6 %, Kappa index 0.924 - 0.951, for set 13). Excluding the variable profile curvature (set 12), overall accuracy ranged from 93.9 to 95.4 % and the Kappa index from 0.932 to 0.948. The maps based on the neural network classifier were consistent and similar to conventional soil maps drawn for the study area, although with more spatial details. The results show the potential of ANNs for soil class prediction in mountainous areas with lithological diversity.

Data-driven clustering : new methods and applications

Abstract : This work is concerned with the development and application of novel unsupervised learning methods, having in mind two target applications: the analysis of forensic case data and the classification of remote sensing images. First, a method based on a symbolic optimization of the inter-sample distance measure is proposed to improve the flexibility of spectral clustering algorithms, and applied to the problem of forensic case data. This distance is optimized using a loss function related to the preservation of neighborhood structure between the input space and the space of principal components, and solutions are found using genetic programming. Results are compared to a variety of state-of--the-art clustering algorithms. Subsequently, a new large-scale clustering method based on a joint optimization of feature extraction and classification is proposed and applied to various databases, including two hyperspectral remote sensing images. The algorithm makes uses of a functional model (e.g., a neural network) for clustering which is trained by stochastic gradient descent. Results indicate that such a technique can easily scale to huge databases, can avoid the so-called out-of-sample problem, and can compete with or even outperform existing clustering algorithms on both artificial data and real remote sensing images. This is verified on small databases as well as very large problems. Résumé : Ce travail de recherche porte sur le développement et l'application de méthodes d'apprentissage dites non supervisées. Les applications visées par ces méthodes sont l'analyse de données forensiques et la classification d'images hyperspectrales en télédétection. Dans un premier temps, une méthodologie de classification non supervisée fondée sur l'optimisation symbolique d'une mesure de distance inter-échantillons est proposée. Cette mesure est obtenue en optimisant une fonction de coût reliée à la préservation de la structure de voisinage d'un point entre l'espace des variables initiales et l'espace des composantes principales. Cette méthode est appliquée à l'analyse de données forensiques et comparée à un éventail de méthodes déjà existantes. En second lieu, une méthode fondée sur une optimisation conjointe des tâches de sélection de variables et de classification est implémentée dans un réseau de neurones et appliquée à diverses bases de données, dont deux images hyperspectrales. Le réseau de neurones est entraîné à l'aide d'un algorithme de gradient stochastique, ce qui rend cette technique applicable à des images de très haute résolution. Les résultats de l'application de cette dernière montrent que l'utilisation d'une telle technique permet de classifier de très grandes bases de données sans difficulté et donne des résultats avantageusement comparables aux méthodes existantes.