Introduction: Les particules ultrafines (PUF) constituent la plus petite fraction de matières particulaires (MP) actuellement mesurable avec des diamètres aérodynamiques <0,1 µm. En raison de leur taille plus petite et de leur surface collective plus élevée que d’autres particules plus grossières, les PUF pourraient avoir des effets plus importants sur la santé. Objectifs: L’objectif ultime de ce projet doctoral est de développer des surfaces d’exposition aux PUF les plus valides possible pour la région de Montréal. Afin d’atteindre cet objectif ultime, nous devions répondre à trois objectifs méthodologiques intermédiaires : Premièrement, évaluer la variabilité spatiale et temporelle des PUF dans la région de Montréal. Deuxièmement, construire des modèles LUR (Land Use Regression) spécifiques à deux saisons (été, hiver) avec différentes transformations de variables et stratégies de sélection de variables. Troisièmement, dériver des surfaces d’exposition aux PUF à l’aide de prévisions de modèles LUR. Méthodes: Nous avons utilisé les données recueillies lors de campagnes de surveillance en site fixe dans la région de Montréal, où les PUF ont été mesurées en hiver (mars 2013) et en été (août-septembre 2015) sur 249 sites d’échantillonnage. Lors de chaque campagne, chacun des sites d’échantillonnage a été visité trois fois pendant trois semaines consécutives. Chaque visite d’échantillonnage comprenait une période de mesure de 20 minutes pour les PUF avec une mesure toutes les secondes. Pour le premier sous-objectif, nous avons étudié la variabilité spatiale des PUF par rapport à deux catégories géographiques différentes de sites d’échantillonnage. Nous avons étudié la variabilité temporelle des PUF à différents niveaux de temps. Pour le deuxième sous-objectif, nous avons calculé près d’une centaine de prédicteurs candidats à la construction des modèles LUR. Nous avons normalisé la distribution des prédicteurs les plus biaisés. Nous avons sélectionné les prédicteurs les plus pertinents parmi leurs variations avant de construire des modèles LUR parcimonieux spécifiques à la saison en suivant des algorithmes de sélection progressive avant et arrière. Pour le troisième sous-objectif, nous avons généré les surfaces d’exposition aux PUF correspondants aux modèles LUR basés sur la prédiction qui comprenaient trente prédicteurs (appelés « modèles complets »). Résultats et conclusions: Nous avons constaté qu’il existe une variabilité considérable des niveaux des PUF dans le temps et dans l’espace. Les niveaux de PUF en hiver étaient presque deux fois plus élevés que ceux observés en été. Les prédicteurs candidats dont la distribution est asymétrique doivent être normalisés avant la construction du modèle afin de minimiser les valeurs aberrantes des PUF qui pourraient ne pas être représentatives des PUF réelles dans la zone d’étude. Nous avons pu générer des contrastes plus fins des PUF dans les surfaces d’exposition générées à l’aide de modèles complets, qui pourraient être plus représentatifs de la distribution spatiale réelle des PUF dans la zone d’étude, par rapport aux modèles parcimonieux classiquement utilisés dans la littérature. Les surfaces d’exposition spécifiques aux saisons générées à l’aide de modèles complets pourraient contribuer à réduire les erreurs de classification non-différentielles d’exposition aux PUF dans les études épidémiologiques. / Introduction: Ultrafine Particles (UFPs) are currently the smallest measurable fraction of particulate matter (PM) with aerodynamic diameters <0.1 µm. Due to their smaller size and higher collective surface area than larger PM, UFPs are hypothesized to have stronger health effects than larger PM. Objectives: The ultimate objective of this doctoral project is to develop UFP exposure surfaces as valid as possible for the Montreal area. In order to achieve the objective, we needed to address three intermediate methodological objectives: First, to evaluate spatial and temporal variability of UFPs in the Montreal area. Second, to build season specific land use regression (LUR) models with different variable transformations and variable selection strategies. Third, to derive UFP exposure surfaces using prediction based LUR models. Methods: We used data collected during fixed-site monitoring campaigns in the Montreal area, where UFPs were measured in winter (March 2013) and in summer (August-September 2015) at 249 sampling sites. During each campaign, each of the sampling sites was visited three times during three consecutive weeks. Each sampling visit entailed a 20-minute measurement period for UFPs with a measurement every second. For our first sub-objective, we studied spatial variability of UFPs with respect to two different geographic categorizations of sampling sites. We studied the temporal variability of UFPs at different levels of time and between seasons. For our second sub-objective, we computed close to a hundred candidate predictors. We normalized the distribution of the skewed predictors. We selected the most relevant predictors among their variations before building season-specific parsimonious LUR models following forward and backward stepwise selection algorithms. For our third sub-objective, we derived UFP exposure surfaces using prediction based LUR models that included thirty candidate predictors (referred to as “Full models”). Results and Conclusions: We have identified that there is considerable variability in UFP levels with respect to time and space. UFP levels during winter were almost twice those observed during summer. Candidate predictors with skewed distribution should be normalized before model building in order to minimize UFP outliers that might not be representative of the actual UFP levels in the study area. We were able to generate finer UFP contrasts in the exposure surfaces derived from full models that might be more representative of the actual spatial distribution of UFPs in the study area, compared to parsimonious models, classically used in the literature. Season-specific UFP exposure surfaces derived using full models could help reduce non-differential exposure misclassification among the epidemiological study subjects when used for risk assessment.

Monoklonala antikroppsbaserade läkemedel (mAb) är ett av de snabbast växande segmenten på läkemedelsmarknaden, främst på grund av deras tillämpning inom onkologi, immunologi och hematologi. Traditionellt sker den industriella produktionen av mAb med fed-batch-odling. Detta är en relativt lätthanterlig process med mAb-utbyten på 5-10 g/L, men dess brist på kontroll över kritiska processparametrar (CPP) orsakar höga mAb-förluster på grund av att kvalitetsspecifikationer inte uppfylls. Ökande marknadskrav och regulatoriska förändringar pådriver läkemedelsindustrin iinnovation inom mAb-tillverkningsprocessen, för att nå kontinuerlig tillverkning. För närvarande, som ett övergångssteg till kontinuerlig tillverkning, sker investeringar i intensifierade fed-batch-odlingar. Dessa uppnår högre celldensiteter på cirka 25-30 g/L, men detta är fortfarande mycket lägre än motsvarande mAb-koncentrationer på 130 g/L som kan uppnås med perfusionsprocesser. Andra fördelar med perfusionsprocesser är att de tillåter flexibla produktionsanläggningar och möjliggör en nivå av processkontroll som skulle tillåta realtidstestning av release. För att upprätthålla en perfusionsprocess under de specificerade förhållandena som garanterar den önskade mAb-kvaliteten, måste CPP kontrolleras noggrant. Process Analytical Technologies (PAT) kan mäta CPP i realtid på ett icke-destruktivt sätt. Denna studie undersökte tillämpningen av två PAT, ArgusEye®-sensorerna och Time-gated Raman-spektroskopi, på perfusionsprocesser. Vi visade att ArgusEye®-sensorerna kan användas för att mäta IgG i perfusionsprover med ganska bra korrelation med referensmetoden. Vi har också visat att multivariata Raman-baserade modeller kan konstrueras för att förutsäga flera CPP, baserat på samma spektra. Framförallt belyser denna studie komplexiteten i tillämpningen av dessa PAT för att kontrollera perfusionsprocesser. För ArgusEye® drar vi slutsatsen att för att få exakta mätningar måste vi ta hänsyn till förändringarna i koncentrationen av värdcellsprotein under en perfusionsprocess, eftersom deras ospecifika bindning till sensorerna är den troliga orsaken till variationen i IgG-mätningarna. För de Raman-baserade modellerna, visar denna studie att en stor mängd data krävs för att bygga korrekta prediktionsmodeller, något som rapprterats om i litteraturen. Sammantaget visar denna rapport att dessa PAT har en stor tillämpningspotential, men de måste förbättras ytterligare innan de kan användas som automatiska återkopplingskontrollverktyg. / Monoclonal antibody-based therapeutics (mAb) are one of the fastest-growing segments in the pharmaceutical market, mainly due to their application in oncology, immunology, and hematology. Traditionally, the industrial production of mAb is done with fed-batch cultivation. This is a relatively easy to operate process with mAb yields of 5-10 g/L, but its lack of control over critical process parameters (CPP) causes high mAb losses due to unmet quality specifications. Driven by increasing market demands and regulatory changes, the pharmaceutical industry is innovating in the mAb manufacturing process to reach continuous manufacturing. Currently, as a transition step to continuous manufacturing, the pharmaceutical industry is investing in intensified fed-batch cultivations. They achieve higher cells densities and present yields around 25-30 g/L, but this is still much lower than the equivalent mAb titers of 130 g/L that can be achieved with perfusion processes. Other advantages of perfusion processes are that they allow the existence of flexible production facilities and enable a level of process control that would permit Real-Time Release Testing. To maintain a perfusion process under the specified conditions to guarantee the desired mAb quality, the CPP need to be closely controlled. Process Analytical Technologies (PAT) can measure CPP in real-time and non-destructively. This study evaluated the application of two PAT, the ArgusEye® sensors and Time-gated Raman spectroscopy, on perfusion processes. We showed that the ArgusEye® sensors can be used to measure IgG in perfusion samples with quite good correlation to the reference method. We have also shown that multivariate Raman-based models can be constructed to predict several CPP based on the same spectra. Most importantly, this study highlights the complexity of the application of these PAT to control perfusion processes. For the ArgusEye®, we conclude that to obtain accurate measurements, we need to account for the changes in the concentration of host cell protein during a perfusion process, as their unspecific binding to the sensors is the probable cause for the variation in the IgG measurements. For the Raman-based models, as previously reported in the literature, this study shows that a high volume of data is require to build accurate prediction models. Overall, this report shows that these PAT have a great potential of application, but they need to be further improved prior to their use as automatic feedback control tools.

[ES] Esta tesis doctoral se centra en el estudio, implementación y aplicación de técnicas de aprendizaje automático y estadística multivariante en el emergente campo de la analítica deportiva, concretamente en el fútbol. Se aplican procedimientos comunmente utilizados y métodos nuevos para resolver cuestiones de investigación en diferentes áreas del análisis del fútbol, tanto en el ámbito del rendimiento deportivo como en el económico. Las metodologías empleadas en esta tesis enriquecen las técnicas utilizadas hasta el momento para obtener una visión global del comportamiento de los equipos de fútbol y pretenden ayudar al proceso de toma de decisiones. Además, la metodología se ha implementado utilizando el software estadístico libre R y datos abiertos, lo que permite la replicabilidad de los resultados. Esta tesis doctoral pretende contribuir a la comprensión de los modelos de aprendizaje automático y estadística multivariante para la predicción analítica deportiva, comparando su capacidad predictiva y estudiando las variables que más influyen en los resultados predictivos de estos modelos. Así, siendo el fútbol un juego de azar donde la suerte juega un papel importante, se proponen metodologías que ayuden a estudiar, comprender y modelizar la parte objetiva de este deporte. Esta tesis se estructura en cinco bloques, diferenciando cada uno en función de la base de datos utilizada para alcanzar los objetivos propuestos. El primer bloque describe las áreas de estudio más comunes en la analítica del fútbol y las clasifica en función de los datos utilizados. Esta parte contiene un estudio exhaustivo del estado del arte de la analítica del fútbol. Así, se recopila parte de la literatura existente en función de los objetivos alcanzados, conjuntamente con una revisión de los métodos estadísticos aplicados. Estos modelos son los pilares sobre los que se sustentan los nuevos procedimientos aquí propuestos. El segundo bloque consta de dos capítulos que estudian el comportamiento de los equipos que alcanzan la Liga de Campeones o la Europa League, descienden a segunda división o permanecen en mitad de la tabla. Se proponen varias técnicas de aprendizaje automático y estadística multivariante para predecir la posición de los equipos a final de temporada. Una vez realizada la predicción, se selecciona el modelo con mejor precisión predictiva para estudiar las acciones de juego que más discriminan entre posiciones. Además, se analizan las ventajas de las técnicas propuestas frente a los métodos clásicos utilizados hasta el momento. El tercer bloque consta de un único capítulo en el que se desarrolla un código de web scraping para facilitar la recuperación de una nueva base de datos con información cuantitativa de las acciones de juego realizadas a lo largo del tiempo en los partidos de fútbol. Este bloque se centra en la predicción de los resultados de los partidos (victoria, empate o derrota) y propone la combinación de una técnica de aprendizaje automático, random forest, y la regresión Skellam, un método clásico utilizado habitualmente para predecir la diferencia de goles en el fútbol. Por último, se compara la precisión predictiva de los métodos clásicos utilizados hasta ahora con los métodos multivariantes propuestos. El cuarto bloque también comprende un único capítulo y pertenece al área económica del fútbol. En este capítulo se aplica un novedoso procedimiento para desarrollar indicadores que ayuden a predecir los precios de traspaso. En concreto, se muestra la importancia de la popularidad a la hora de calcular el valor de mercado de los jugadores, por lo que este capítulo propone una nueva metodología para la recogida de información sobre la popularidad de los jugadores. En el quinto bloque se revelan los aspectos más relevantes de esta tesis para la investigación y la analítica en el fútbol, incluyendo futuras líneas de trabajo. / [CA] Aquesta tesi doctoral se centra en l'estudi, implementació i aplicació de tècniques d'aprenentatge automàtic i estadística multivariant en l'emergent camp de l'analítica esportiva, concretament en el futbol. S'apliquen procediments comunament utilitzats i mètodes nous per a resoldre qu¿estions d'investigació en diferents àrees de l'anàlisi del futbol, tant en l'àmbit del rendiment esportiu com en l'econòmic. Les metodologies emprades en aquesta tesi enriqueixen les tècniques utilitzades fins al moment per a obtindre una visió global del comportament dels equips de futbol i pretenen ajudar al procés de presa de decisions. A més, la metodologia s'ha implementat utilitzant el programari estadístic lliure R i dades obertes, la qual cosa permet la replicabilitat dels resultats. Aquesta tesi doctoral pretén contribuir a la comprensió dels models d'aprenentatge automàtic i estadística multivariant per a la predicció analítica esportiva, comparant la seua capacitat predictiva i estudiant les variables que més influeixen en els resultats predictius d'aquests models. Així, sent el futbol un joc d'atzar on la sort juga un paper important, es proposen metodologies que ajuden a estudiar, comprendre i modelitzar la part objectiva d'aquest esport. Aquesta tesi s'estructura en cinc blocs, diferenciant cadascun en funció de la base de dades utilitzada per a aconseguir els objectius proposats. El primer bloc descriu les àrees d'estudi més comuns en l'analítica del futbol i les classifica en funció de les dades utilitzades. Aquesta part conté un estudi exhaustiu de l'estat de l'art de l'analítica del futbol. Així, es recopila part de la literatura existent en funció dels objectius aconseguits, conjuntament amb una revisió dels mètodes estadístics aplicats. Aquests models són els pilars sobre els quals se sustenten els nous procediments ací proposats. El segon bloc consta de dos capítols que estudien el comportament dels equips que aconsegueixen la Lliga de Campions o l'Europa League, descendeixen a segona divisió o romanen a la meitat de la taula. Es proposen diverses tècniques d'aprenentatge automàtic i estadística multivariant per a predir la posició dels equips a final de temporada. Una vegada realitzada la predicció, se selecciona el model amb millor precisió predictiva per a estudiar les accions de joc que més discriminen entre posicions. A més, s'analitzen els avantatges de les tècniques proposades enfront dels mètodes clàssics utilitzats fins al moment. El tercer bloc consta d'un únic capítol en el qual es desenvolupa un codi de web scraping per a facilitar la recuperació d'una nova base de dades amb informació quantitativa de les accions de joc realitzades al llarg del temps en els partits de futbol. Aquest bloc se centra en la predicció dels resultats dels partits (victòria, empat o derrota) i proposa la combinació d'una tècnica d'aprenentatge automàtic, random forest, i la regressió Skellam, un mètode clàssic utilitzat habitualment per a predir la diferència de gols en el futbol. Finalment, es compara la precisió predictiva dels mètodes clàssics utilitzats fins ara amb els mètodes multivariants proposats. El quart bloc també comprén un únic capítol i pertany a l'àrea econòmica del futbol. En aquest capítol s'aplica un nou procediment per a desenvolupar indicadors que ajuden a predir els preus de traspàs. En concret, es mostra la importància de la popularitat a l'hora de calcular el valor de mercat dels jugadors, per la qual cosa aquest capítol proposa una nova metodologia per a la recollida d'informació sobre la popularitat dels jugadors. En el cinqué bloc es revelen els aspectes més rellevants d'aquesta tesi per a la investigació i l'analítica en el futbol, incloent-hi futures línies de treball. / [EN] This doctoral thesis focuses on studying, implementing, and applying machine learning and multivariate statistics techniques in the emerging field of sports analytics, specifically in football. Commonly used procedures and new methods are applied to solve research questions in different areas of football analytics, both in the field of sports performance and in the economic field. The methodologies used in this thesis enrich the techniques used so far to obtain a global vision of the behaviour of football teams and are intended to help the decision-making process. In addition, the methodology was implemented using the free statistical software R and open data, which allows for reproducibility of the results. This doctoral thesis aims to contribute to the understanding of the behaviour of machine learning and multivariate models for analytical sports prediction, comparing their predictive capacity and studying the variables that most influence the predictive results of these models. Thus, since football is a game of chance where luck plays an important role, this document proposes methodologies that help to study, understand, and model the objective part of this sport. This thesis is structured into five blocks, differentiating each according to the database used to achieve the proposed objectives. The first block describes the most common study areas in football analytics and classifies them according to the available data. This part contains an exhaustive study of football analytics state of the art. Thus, part of the existing literature is compiled based on the objectives achieved, with a review of the statistical methods applied. These methods are the pillars on which the new procedures proposed here are based. The second block consists of two chapters that study the behaviour of teams concerning the ranking at the end of the season: top (qualifying for the Champions League or Europa League), middle, or bottom (relegating to a lower division). Several machine learning and multivariate statistical techniques are proposed to predict the teams' position at the season's end. Once the prediction has been made, the model with the best predictive accuracy is selected to study the game actions that most discriminate between positions. In addition, the advantages of our proposed techniques compared to the classical methods used so far are analysed. The third block consists of a single chapter in which a web scraping code is developed to facilitate the retrieval of a new database with quantitative information on the game actions carried out over time in football matches. This block focuses on predicting match outcomes (win, draw, or loss) and proposing the combination of a machine learning technique, random forest, and Skellam regression model, a classical method commonly used to predict goal difference in football. Finally, the predictive accuracy of the classical methods used so far is compared with the proposed multivariate methods. The fourth block also comprises a single chapter and pertains to the economic football area. This chapter applies a novel procedure to develop indicators that help predict transfer fees. Specifically, it is shown the importance of popularity when calculating the players' market value, so this chapter is devoted to propose a new methodology for collecting players' popularity information. The fifth block reveals the most relevant aspects of this thesis for research and football analytics, including future lines of work. / Malagón Selma, MDP. (2023). Machine Learning and Multivariate Statistical Tools for Football Analytics [Tesis doctoral]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/197630

O camarão rosa (Farfantepenaeus subtilis) explotado pela pesca industrial na plataforma continental amazônica brasileira possui um ciclo de vida curto, mas complexo, habitando áreas oceânicas, mais ao norte da área de ocorrência, na fase adulta e larval, e áreas estuarinas e lagunares na fase de pós-larva e juvenil. O período de maior intensidade de reprodução se estende de maio a setembro e logo após a reprodução as larvas eclodem e iniciam sua migração para áreas costeiras, passando por diversas fases, onde se assentam e residem principalmente entre junho e outubro. A partir de setembro até janeiro do ano seguinte é maior a intensidade de recrutamento de juvenis às áreas oceânicas, onde passam a amadurecer e, a partir de dezembro, começam a ser capturados pela pesca industrial. A maior abundância da população adulta em termos de biomassa vai de março a agosto quando também se verificam as maiores capturas. As fêmeas crescem mais que os machos e estão presentes sempre em maior proporção nas capturas (61%). Os comprimentos assintóticos foram estimados em 231 mm ( k = 1,6 \'ano POT.-1\') e 205 mm (k = 0,94 \'ano POT.-1\'), para fêmeas e machos respectivamente. A população apresenta taxa de mortalidade natural relativamente elevada, 2,53 \'ano POT.-1\' para fêmeas e 1,83 \'ano POT.-1\' para machos, sendo observadas acentuadas flutuações de recrutamento e abundância, com evidências de que são fortemente governadas pelas condições ambientais. O estoque vem sendo explotado em níveis moderados nos anos recentes (E = 0,45), embora tenha sofrido elevadas taxas de explotação na década de 80, o que levou a uma redução do tamanho da população. O rendimento máximo sustentável, considerado uma média de longo prazo, foi estimado em 4.032 toneladas de cauda por ano, para um esforço de pesca de 19.370 dias de mar. Nos últimos anos, se observa uma tendência de recuperação da biomassa populacional, mas com as oscilações anuais características da espécie. A vazão do rio Amazonas é o fator ambiental que governa com mais intensidade as condições do ambiente costeiro na região e verificou-se que suas flutuações estão correlacionadas a alterações na abundância da população da espécie. Postula-se que o aporte e sobrevivência das larvas e pós-larvas no ambiente costeiro seja influenciada pela intensidade da vazão do rio. O período em que se assentam nos berçários na zona costeira coincide com a estação de vazante do rio, sendo a sobrevivência favorecida por vazões abaixo da média e vice-versa. Portanto, medidas de ordenamento voltadas para o uso sustentável do recurso devem estar associadas ao conhecimento das condições ambientais nesta fase, bem como a estudos sobre a abundância de pós-larvas e juvenis na faixa costeira. / The brown shrimp (Farfantepenaeus subtilis) exploited by the industrial fishery on the continental shelf of the Brazilian Amazon has a short but complex life cyele, inhabiting oceanic areas, at the north of the area of occurrence, during the adult and larval stages, and estuarine areas and lagoons in post-larval and juvenile. The period of highest intensity of reproduction extends from May to September and soon after the hatch, the larvae start their migration to coastal areas, passing through several stages, where they settle and remain resident between June and October. From September to January of the following year the intensity of recruitment to ocean areas is higher, and once there they start to mature and are caught by the industrial fishery from December on. The highest abundance of the adult population in terms of biomass is observed from March to August when the largest catches also occur. Females grow larger than males and are always present in greater proportion in catches (61%). The asymptotic lengths were estimated at 231 mm (k = 1.6 \'year POT.-1\') and 205 mm (k = 0.94 \'year POT.-1\') for females and males respectively. The population has a natural mortality rate relatively high, 2.53 \'year POT.-1\' for females and 1.83 \'years POT.-1\' for males, and pronounced fluctuations in recruitment and abundance are observed, with evidence of being strongly governed by environmental conditions. The stock has been exploited at moderate levels in recent years (E = 0.45), although it has suffered high rates of exploitation in the 80\'s, which led to a reduction in population size. The maximum sustainable yield, considered a long-term average, was estimated at 4,032 ton of tail per year for a fishing effort of 19,370 days at sea. In recent years, it is observed a tendency of recovering of the population biomass, but annual fluctuations are characteristics of the species. The flow of the Amazon River is the main environmental facto r that governs the conditions of the coastal environment in the region and it was found that it is correlated with the fluctuatícn of the brown shrimp population abundance. It is postulated that the uptake and survival of larvae and post larvae in the coastal environment is lnfluenced by the intensity of river flow, The period during which they settle at the nurseries in the coastal zone coincides with the dry season and their survival is favored when the flow of the river is below the average, and vice versa. Therefore, management measures aimed at sustainable use of the resource must be associated with the knowledge of environmental conditions during this phase, as well as studies on the abundance of post-larvae and juveniles in the coastal zone.

Les deux dernières décennies ont été marquées par une augmentation significative des données collectées dans les systèmes d'informations. Cette masse de données contient des informations riches et peu exploitées. Cette réalité s’applique au secteur de la santé où l'informatisation est un enjeu pour l’amélioration de la qualité des soins. Les méthodes existantes dans les domaines de l'extraction de processus, de l'exploration de données et de la modélisation mathématique ne parviennent pas à gérer des données aussi hétérogènes et volumineuses que celles de la santé. Notre objectif est de développer une méthodologie complète pour transformer des données de santé brutes en modèles de simulation des parcours de soins cliniques. Nous introduisons d'abord un cadre mathématique dédié à la découverte de modèles décrivant les parcours de soin, en combinant optimisation combinatoire et Process Mining. Ensuite, nous enrichissons ce modèle par l’utilisation conjointe d’un algorithme d’alignement de séquences et de techniques classiques de Data Mining. Notre approche est capable de gérer des données bruitées et de grande taille. Enfin, nous proposons une procédure pour la conversion automatique d'un modèle descriptif des parcours de soins en un modèle de simulation dynamique. Après validation, le modèle obtenu est exécuté pour effectuer des analyses de sensibilité et évaluer de nouveaux scénarios. Un cas d’étude sur les maladies cardiovasculaires est présenté, avec l’utilisation de la base nationale des hospitalisations entre 2006 et 2015. La méthodologie présentée dans cette thèse est réutilisable dans d'autres aires thérapeutiques et sur d'autres sources de données de santé. / During the last two decades, the amount of data collected in Information Systems has drastically increased. This large amount of data is highly valuable. This reality applies to health-care where the computerization is still an ongoing process. Existing methods from the fields of process mining, data mining and mathematical modeling cannot handle large-sized and variable event logs. Our goal is to develop an extensive methodology to turn health data from event logs into simulation models of clinical pathways. We first introduce a mathematical framework to discover optimal process models. Our approach shows the benefits of combining combinatorial optimization and process mining techniques. Then, we enrich the discovered model with additional data from the log. An innovative combination of a sequence alignment algorithm and of classical data mining techniques is used to analyse path choices within long-term clinical pathways. The approach is suitable for noisy and large logs. Finally, we propose an automatic procedure to convert static models of clinical pathways into dynamic simulation models. The resulting models perform sensitivity analyses to quantify the impact of determinant factors on several key performance indicators related to care processes. They are also used to evaluate what-if scenarios. The presented methodology was proven to be highly reusable on various medical fields and on any source of event logs. Using the national French database of all the hospital events from 2006 to 2015, an extensive case study on cardiovascular diseases is presented to show the efficiency of the proposed framework.

O camarão rosa (Farfantepenaeus subtilis) explotado pela pesca industrial na plataforma continental amazônica brasileira possui um ciclo de vida curto, mas complexo, habitando áreas oceânicas, mais ao norte da área de ocorrência, na fase adulta e larval, e áreas estuarinas e lagunares na fase de pós-larva e juvenil. O período de maior intensidade de reprodução se estende de maio a setembro e logo após a reprodução as larvas eclodem e iniciam sua migração para áreas costeiras, passando por diversas fases, onde se assentam e residem principalmente entre junho e outubro. A partir de setembro até janeiro do ano seguinte é maior a intensidade de recrutamento de juvenis às áreas oceânicas, onde passam a amadurecer e, a partir de dezembro, começam a ser capturados pela pesca industrial. A maior abundância da população adulta em termos de biomassa vai de março a agosto quando também se verificam as maiores capturas. As fêmeas crescem mais que os machos e estão presentes sempre em maior proporção nas capturas (61%). Os comprimentos assintóticos foram estimados em 231 mm ( k = 1,6 \'ano POT.-1\') e 205 mm (k = 0,94 \'ano POT.-1\'), para fêmeas e machos respectivamente. A população apresenta taxa de mortalidade natural relativamente elevada, 2,53 \'ano POT.-1\' para fêmeas e 1,83 \'ano POT.-1\' para machos, sendo observadas acentuadas flutuações de recrutamento e abundância, com evidências de que são fortemente governadas pelas condições ambientais. O estoque vem sendo explotado em níveis moderados nos anos recentes (E = 0,45), embora tenha sofrido elevadas taxas de explotação na década de 80, o que levou a uma redução do tamanho da população. O rendimento máximo sustentável, considerado uma média de longo prazo, foi estimado em 4.032 toneladas de cauda por ano, para um esforço de pesca de 19.370 dias de mar. Nos últimos anos, se observa uma tendência de recuperação da biomassa populacional, mas com as oscilações anuais características da espécie. A vazão do rio Amazonas é o fator ambiental que governa com mais intensidade as condições do ambiente costeiro na região e verificou-se que suas flutuações estão correlacionadas a alterações na abundância da população da espécie. Postula-se que o aporte e sobrevivência das larvas e pós-larvas no ambiente costeiro seja influenciada pela intensidade da vazão do rio. O período em que se assentam nos berçários na zona costeira coincide com a estação de vazante do rio, sendo a sobrevivência favorecida por vazões abaixo da média e vice-versa. Portanto, medidas de ordenamento voltadas para o uso sustentável do recurso devem estar associadas ao conhecimento das condições ambientais nesta fase, bem como a estudos sobre a abundância de pós-larvas e juvenis na faixa costeira. / The brown shrimp (Farfantepenaeus subtilis) exploited by the industrial fishery on the continental shelf of the Brazilian Amazon has a short but complex life cyele, inhabiting oceanic areas, at the north of the area of occurrence, during the adult and larval stages, and estuarine areas and lagoons in post-larval and juvenile. The period of highest intensity of reproduction extends from May to September and soon after the hatch, the larvae start their migration to coastal areas, passing through several stages, where they settle and remain resident between June and October. From September to January of the following year the intensity of recruitment to ocean areas is higher, and once there they start to mature and are caught by the industrial fishery from December on. The highest abundance of the adult population in terms of biomass is observed from March to August when the largest catches also occur. Females grow larger than males and are always present in greater proportion in catches (61%). The asymptotic lengths were estimated at 231 mm (k = 1.6 \'year POT.-1\') and 205 mm (k = 0.94 \'year POT.-1\') for females and males respectively. The population has a natural mortality rate relatively high, 2.53 \'year POT.-1\' for females and 1.83 \'years POT.-1\' for males, and pronounced fluctuations in recruitment and abundance are observed, with evidence of being strongly governed by environmental conditions. The stock has been exploited at moderate levels in recent years (E = 0.45), although it has suffered high rates of exploitation in the 80\'s, which led to a reduction in population size. The maximum sustainable yield, considered a long-term average, was estimated at 4,032 ton of tail per year for a fishing effort of 19,370 days at sea. In recent years, it is observed a tendency of recovering of the population biomass, but annual fluctuations are characteristics of the species. The flow of the Amazon River is the main environmental facto r that governs the conditions of the coastal environment in the region and it was found that it is correlated with the fluctuatícn of the brown shrimp population abundance. It is postulated that the uptake and survival of larvae and post larvae in the coastal environment is lnfluenced by the intensity of river flow, The period during which they settle at the nurseries in the coastal zone coincides with the dry season and their survival is favored when the flow of the river is below the average, and vice versa. Therefore, management measures aimed at sustainable use of the resource must be associated with the knowledge of environmental conditions during this phase, as well as studies on the abundance of post-larvae and juveniles in the coastal zone.

The master thesis is focused on creating a methodology for quantification of the nerve fiber layer on photographs of the retina. The introductory part of the text presents a medical motivation of the thesis and mentions several studies dealing with this issue. Furthermore, the work describes available textural features and compares their ability to quantify the thickness of the nerve fiber layer. Based on the described knowledge, the methodology to make different regression models enabling prediction of the retinal nerve fiber layer thickness was developed. Then, the methodology was tested on the available image dataset. The results showed, that the outputs of regression models achieve a high correlation between the predicted output and the retinal nerve fiber layer thickness measured by optical coherence tomography. The conclusion discusses an usability of the applied solution.

In the context of both projects Enhancing literacy development in European languages, work package 2 and The early literacy development and its variability in children at risk of specific learning disabilities, we monitored child development of literacy in preschool age and during the first years of school attendance in a four-stage process. The research group (n = 76) compound of typically developing children (BV = 37), children with the family risk of dyslexia (RR = 22) and children with specific language impairment (NVŘ = 17). We evaluated development of phonemic/phonological, lexical/semantic and morphological/syntactic skills, preliteracy skills and early literacy skills. The last fifth test stage included the assessment of literacy development in 3rd graders. First, a group of children with literacy deficits (n = 9) was identified via the latent profile analysis method. Subsequently, four predictive models of literacy deficits for each stage were created by means of lasso or L-1 penalized regression method. Predictive models follows the trend that until literacy skills are fully automatized (preschool age and the 1st grade), phonemic and phonological skills predominate, but later - after the formal learning to read and write proceeds - early literacy skills are becoming more and more...

Indiana University-Purdue University Indianapolis (IUPUI) / With the increasing availability of Electronic Health Records (EHRs) and advances in deep learning techniques, developing deep predictive models that use EHR data to solve healthcare problems has gained momentum in recent years. The majority of clinical predictive models benefit from structured data in EHR (e.g., lab measurements and medications). Still, learning clinical outcomes from all possible information sources is one of the main challenges when building predictive models. This work focuses mainly on two sources of information that have been underused by researchers; unstructured data (e.g., clinical notes) and a patient network. We propose a novel hybrid deep learning model, DeepNote-GNN, that integrates clinical notes information and patient network topological structure to improve 30-day hospital readmission prediction. DeepNote-GNN is a robust deep learning framework consisting of two modules: DeepNote and patient network. DeepNote extracts deep representations of clinical notes using a feature aggregation unit on top of a state-of-the-art Natural Language Processing (NLP) technique - BERT. By exploiting these deep representations, a patient network is built, and Graph Neural Network (GNN) is used to train the network for hospital readmission predictions. Performance evaluation on the MIMIC-III dataset demonstrates that DeepNote-GNN achieves superior results compared to the state-of-the-art baselines on the 30-day hospital readmission task. We extensively analyze the DeepNote-GNN model to illustrate the effectiveness and contribution of each component of it. The model analysis shows that patient network has a significant contribution to the overall performance, and DeepNote-GNN is robust and can consistently perform well on the 30-day readmission prediction task. To evaluate the generalization of DeepNote and patient network modules on new prediction tasks, we create a multimodal model and train it on structured and unstructured data of MIMIC-III dataset to predict patient mortality and Length of Stay (LOS). Our proposed multimodal model consists of four components: DeepNote, patient network, DeepTemporal, and score aggregation. While DeepNote keeps its functionality and extracts representations of clinical notes, we build a DeepTemporal module using a fully connected layer stacked on top of a one-layer Gated Recurrent Unit (GRU) to extract the deep representations of temporal signals. Independent to DeepTemporal, we extract feature vectors of temporal signals and use them to build a patient network. Finally, the DeepNote, DeepTemporal, and patient network scores are linearly aggregated to fit the multimodal model on downstream prediction tasks. Our results are very competitive to the baseline model. The multimodal model analysis reveals that unstructured text data better help to estimate predictions than temporal signals. Moreover, there is no limitation in applying a patient network on structured data. In comparison to other modules, the patient network makes a more significant contribution to prediction tasks. We believe that our efforts in this work have opened up a new study area that can be used to enhance the performance of clinical predictive models.

In this thesis, the problem of predicting the full dose distribution from a partially modeled dose calculation is addressed. Two solutions were studied: a vanilla Hierarchically Densely Connected U-net (HDUnet) and a Conditional Generative Adversarial Network (CGAN) with HDUnet as a generator. The CGAN approach is a 3D version of Pix2Pix [1] for Image to Image translation which we name Dose2Dose. The research question that this project tackled is whether the Dose2Dose can learn more effective dose transformations than the vanilla HDUnet. To answer this, the models were trained using dose calculations of phantom slabs generated for the problem in pairs of inputs (doses without magnetic field) and targets (doses with magnetic field). Once trained, the models were evaluated and compared in various aspects. The evidence gathered suggests that the vanilla HDUnet model can learn to generate better dose predictions than the generative model. However, in terms of the resulting dose distributions, the samples generated from the Dose2Dose are as likely to belong to the target dose calculation distribution as those of the vanilla HDUnet. The results contain errors of considerable magnitude, and do not accomplish clinical suitability tests. / I denna avhandling har problemet med att förutsäga full dosfördelning från en delvis modellerad dosberäkning tagits upp. Två lösningar studerades: ett vanilla HDUnet och ett betingat generativt nätverk (CGAN) med HDUnet som generator. CGAN -metoden var en 3D-version av Pix2Pix [1] för översättning av bild till bild med namnet Dose2Dose. Forskningsfrågan som detta projekt tog upp var om Dose2Dose kan lära sig mer effektiva dostransformationer än vanilla HDUnet. För att svara på detta tränades modellerna med hjälp av parvisa dosberäkningar, i indata (doser utan magnetfält) och mål (doser med magnetfält).. När de var tränade utvärderades modellerna och jämfördes i olika aspekter. De samlade bevisen tyder på att Vanilla HDUnet -modellen kan lära sig att generera bättre dosförutsägelser än den generativa modellen. När det gäller de resulterande dosfördelningarna är emellertid de prover som genererats från Dose2Dose lika sannolikt att tillhöra måldosberäkningsfördelningen som de för vanilla HDUnet. Resultaten innehåller stora storleksfel och uppfyller inte kraven för klinisk tillämpbarhet.