Dissertations / Theses on the topic 'Prognostics prediction model'

Prognostics and healthmanagement (PHM) is an engineering discipline that aims to maintain the systembehaviour and function, and assure the mission success, safety andeffectiveness. Health management using a proper condition-based maintenance (CBM)deployment is a worldwide accepted technique and has grown very popular in manyindustries over the past decades. These techniques are relevant in environmentswhere the prediction of a failure and the prevention and mitigation of itsconsequences increase the profit and safety of the facilities concerned.Prognosis is the most critical part of this process and is nowadays recognizedas a key feature in maintenance strategies, since estimation of the remaininguseful life (RUL) is essential.PHM can provide a stateassessment of the future health of systems or components, e.g. when a degradedstate has been found. Using this technology, one can estimate how long it willtake before the equipment will reach a failure threshold, in future operatingconditions and future environmental conditions. This thesis focuses especiallyon physics-based prognostic approaches, which depend on a fundamentalunderstanding of the physical system in order to develop condition monitoringtechniques and to predict the RUL.The overall research objective of thework performed for this thesis has been to improve the accuracy and precisionof RUL predictions. The research hypothesis is that fusing the output of morethan one method will improve the accuracy and precision of the RUL estimation,by developing a new approach to prognostics that combines different remaininglife estimators and physics-based and data-driven methods. There are two waysof acquiring data for data-driven models, namely measurements of real systemsand syntactic data generation from simulations. The thesis deals with two casestudies, the first of which concerns the generation of synthetic data andindirect measurement of dynamic bearing loads and was performed atBillerudKorsäs paper mill at Karlsborg in Sweden. In this study the behaviourof a roller in a paper machine was analysed using the finite element method(FEM). The FEM model is a step towards the possibility of generating syntheticdata on different failure modes, and the possibility of estimating crucialparameters like dynamic bearing forces by combining real vibration measurementswith the FEM model. The second case study deals with the development ofprognostic methods for battery discharge estimation for Mars-based rovers. Herephysical models and measurement data were used in the prognostic development insuch a way that the degradation behaviour of the battery could be modelled andsimulated in order to predict the life-length. A particle filter turned out tobe the method of choice in performing the state assessment and predicting thefuture degradation. The method was then applied to a case study of batteriesthat provide power to the rover.
Godkänd; 2015; 20151116 (madmis); Nedanstående person kommer att hålla licentiatseminarium för avläggande av teknologie licentiatexamen. Namn: Madhav Mishra Ämne: Drift och underhållsteknik/Operation and Maintenance Engineering Uppsats: Model-based Prognostics for Prediction of Remaining Useful Life Examinator: Professor Uday Kumar Institutionen för samhällsbyggnad och naturresurser Avdelning Drift, underhåll och akustik Luleå tekniska universitet Diskutant: Accos. Professor Jyoti Kumar Sinha University of Manchester, Aerospace and Civil Engineering, Manchester Tid: Torsdag 17 december 2015 kl 10.00 Plats: F1031, Luleå tekniska universitet

La dégradation est un phénomène inévitable qui affecte les composants et les systèmes d'ingénierie, et qui peut entraîner leurs défaillances avec des conséquences potentiellement catastrophiques selon l'application. La motivation de cette Thèse est d'essayer de modéliser, d'analyser et de prédire les défaillances par des méthodes pronostiques qui peuvent permettre une gestion prédictive de la maintenance des actifs. Cela permettrait aux décideurs d'améliorer la planification de la maintenance, augmentant ainsi la disponibilité et la sûreté du système en minimisant les arrêts imprévus. Dans cet objectif, la recherche au cours de la thèse a été consacrée à l'adaptation et à l'utilisation d'approches basées sur des modèles et d'approches pilotées par les données pour traiter les processus de dégradation qui peuvent conduire à différents modes de défaillance dans les composants industriels, en utilisant différentes sources d'informations et de données pour effectuer des prédictions sur l'évolution de la dégradation et estimer la durée de vie utile restante (RUL).Les travaux de thèse ont porté sur deux applications pronostiques spécifiques: les pronostics basés sur des modèles pour la prédiction de la croissance des fissures par fatigue et les pronostics pilotées par les données pour les prédictions à pas multiples des données de séries chronologiques des composants des Centrales Nucléaires.Les pronostics basé sur des modèles compter sur le choix des modèles adoptés de Physics-of-Failure (PoF). Cependant, chaque modèle de dégradation ne convient qu'à certains processus de dégradation dans certaines conditions de fonctionnement, qui souvent ne sont pas connues avec précision. Pour généraliser, des ensembles de multiples modèles de dégradation ont été intégrés dans la méthode pronostique basée sur les modèles afin de tirer profit des différentes précisions des modèles spécifiques aux différentes dégradations et conditions. Les principales contributions des approches pronostiques proposées basées sur l'ensemble des modèles sont l'intégration d'approches de filtrage, y compris le filtrage Bayésien récursif et le Particle Filtering (PF), et de nouvelles stratégies d'ensemble pondérées tenant compte des précisions des modèles individuels dans l'ensemble aux étapes de prédiction précédentes. Les méthodes proposées ont été validées par des études de cas de croissance par fissures de fatigue simulées dans des conditions de fonctionnement variables dans le temps.Quant à la prédictions à pas multiples, elle reste une tâche difficile pour le Prognostics and Health Management (PHM) car l'incertitude de prédiction a tendance à augmenter avec l'horizon temporel de la prédiction. La grande incertitude de prédiction a limité le développement de pronostics à pas multiples dans les applications. Pour résoudre le problème, de nouveaux modèles de prédiction à pas multiples basés sur la Long Short-Term Memory (LSTM), un réseau de neurones profond développé pour traiter les dépendances à long terme dans les données de séries chronologiques, ont été développés dans cette Thèse. Pour des applications pratiques réalistes, les méthodes proposées abordent également les problèmes supplémentaires de détection d'anomalie, d'optimisation automatique des hyper-paramètres et de quantification de l'incertitude de prédiction. Des études de cas pratiques ont été envisagées, concernant les données de séries chronologiques collectées auprès des Générateurs de Vapeur et de Pompes de Refroidissement de Réacteurs de Centrales Nucléaires
Degradation is an unavoidable phenomenon that affects engineering components and systems, and which may lead to their failures with potentially catastrophic consequences depending on the application. The motivation of this Thesis is trying to model, analyze and predict failures with prognostic methods that can enable a predictive management of asset maintenance. This would allow decision makers to improve maintenance planning, thus increasing system availability and safety by minimizing unexpected shutdowns. To this aim, research during the Thesis has been devoted to the tailoring and use of both model-based and data-driven approaches to treat the degradation processes that can lead to different failure modes in industrial components, making use of different information and data sources for performing predictions on the degradation evolution and estimating the Remaining Useful Life (RUL).The Ph.D. work has addressed two specific prognostic applications: model-based prognostics for fatigue crack growth prediction and data-driven prognostics for multi-step ahead predictions of time series data of Nuclear Power Plant (NPP) components.Model-based prognostics relies on the choice of the adopted Physics-of-Failure (PoF) models. However, each degradation model is appropriate only to certain degradation process under certain operating conditions, which are often not precisely known. To generalize this, ensembles of multiple degradation models have been embedded in the model-based prognostic method in order to take advantage of the different accuracies of the models specific to different degradations and conditions. The main contributions of the proposed ensemble of models-based prognostic approaches are the integration of filtering approaches, including recursive Bayesian filtering and Particle Filtering (PF), and novel weighted ensemble strategies considering the accuracies of the individual models in the ensemble at the previous time steps of prediction. The proposed methods have been validated by case studies of fatigue crack growth simulated with time-varying operating conditions.As for multi-step ahead prediction, it remains a difficult task of Prognostics and Health Management (PHM) because prediction uncertainty tends to increase with the time horizon of the prediction. Large prediction uncertainty has limited the development of multi-step ahead prognostics in applications. To address the problem, novel multi-step ahead prediction models based on Long Short- Term Memory (LSTM), a deep neural network developed for dealing with the long-term dependencies in the time series data have been developed in this Thesis. For realistic practical applications, the proposed methods also address the additional issues of anomaly detection, automatic hyperparameter optimization and prediction uncertainty quantification. Practical case studies have been considered, concerning time series data collected from Steam Generators (SGs) and Reactor Coolant Pumps (RCPs) of NPPs

Detecting abnormal operating conditions, which will lead to faults developing later, has important economic implications for industries trying to meet their performance and production goals. It is unacceptable to wait for failures that have potential safety, environmental and financial consequences. Moving from a “reactive” strategy to a “proactive” strategy can improve critical equipment reliability and availability while constraining maintenance costs, reducing production deferrals, decreasing the need for spare parts. Once the fault initiates, predicting its progression and deterioration can enable timely interventions without risk to personnel safety or to equipment integrity. This work presents an online-integrated condition monitoring and prognostics framework that addresses the above issues holistically. The proposed framework aligns fully with ISO 17359:2011 and derives from the I-P and P-F curve. Depending upon the running state of machine with respect to its I-P and P-F curve an algorithm will do one of the following: (1) Predict the ideal behaviour and any departure from the normal operating envelope using a combination of Evolving Clustering Method (ECM), a normalised fuzzy weighted distance and tracking signal method. (2) Identify the cause of the departure through an automated diagnostics system using a modified version of ECM for classification. (3) Predict the short-term progression of fault using a modified version of the Dynamic Evolving Neuro-Fuzzy Inference System (DENFIS), called here MDENFIS and a tracking signal method. (4) Predict the long term progression of fault (Prognostics) using a combination of Autoregressive Integrated Moving Average (ARIMA)- Empirical Mode Decomposition (EMD) for predicting the future input values and MDENFIS for predicting the long term progression of fault (output). The proposed model was tested and compared against other models in the literature using benchmarks and field data. This work demonstrates four noticeable improvements over previous methods: (1) Enhanced testing prediction accuracy, (2) comparable processing time if not better, (3) the ability to detect sudden changes in the process and finally (4) the ability to identify and isolate the problem source with high accuracy.

The ability to estimate the asset reliability and the probability of failure is critical to reducing maintenance costs, operation downtime, and safety hazards. Predicting the survival time and the probability of failure in future time is an indispensable requirement in prognostics and asset health management. In traditional reliability models, the lifetime of an asset is estimated using failure event data, alone; however, statistically sufficient failure event data are often difficult to attain in real-life situations due to poor data management, effective preventive maintenance, and the small population of identical assets in use. Condition indicators and operating environment indicators are two types of covariate data that are normally obtained in addition to failure event and suspended data. These data contain significant information about the state and health of an asset. Condition indicators reflect the level of degradation of assets while operating environment indicators accelerate or decelerate the lifetime of assets. When these data are available, an alternative approach to the traditional reliability analysis is the modelling of condition indicators and operating environment indicators and their failure-generating mechanisms using a covariate-based hazard model. The literature review indicates that a number of covariate-based hazard models have been developed. All of these existing covariate-based hazard models were developed based on the principle theory of the Proportional Hazard Model (PHM). However, most of these models have not attracted much attention in the field of machinery prognostics. Moreover, due to the prominence of PHM, attempts at developing alternative models, to some extent, have been stifled, although a number of alternative models to PHM have been suggested. The existing covariate-based hazard models neglect to fully utilise three types of asset health information (including failure event data (i.e. observed and/or suspended), condition data, and operating environment data) into a model to have more effective hazard and reliability predictions. In addition, current research shows that condition indicators and operating environment indicators have different characteristics and they are non-homogeneous covariate data. Condition indicators act as response variables (or dependent variables) whereas operating environment indicators act as explanatory variables (or independent variables). However, these non-homogenous covariate data were modelled in the same way for hazard prediction in the existing covariate-based hazard models. The related and yet more imperative question is how both of these indicators should be effectively modelled and integrated into the covariate-based hazard model. This work presents a new approach for addressing the aforementioned challenges. The new covariate-based hazard model, which termed as Explicit Hazard Model (EHM), explicitly and effectively incorporates all three available asset health information into the modelling of hazard and reliability predictions and also drives the relationship between actual asset health and condition measurements as well as operating environment measurements. The theoretical development of the model and its parameter estimation method are demonstrated in this work. EHM assumes that the baseline hazard is a function of the both time and condition indicators. Condition indicators provide information about the health condition of an asset; therefore they update and reform the baseline hazard of EHM according to the health state of asset at given time t. Some examples of condition indicators are the vibration of rotating machinery, the level of metal particles in engine oil analysis, and wear in a component, to name but a few. Operating environment indicators in this model are failure accelerators and/or decelerators that are included in the covariate function of EHM and may increase or decrease the value of the hazard from the baseline hazard. These indicators caused by the environment in which an asset operates, and that have not been explicitly identified by the condition indicators (e.g. Loads, environmental stresses, and other dynamically changing environment factors). While the effects of operating environment indicators could be nought in EHM; condition indicators could emerge because these indicators are observed and measured as long as an asset is operational and survived. EHM has several advantages over the existing covariate-based hazard models. One is this model utilises three different sources of asset health data (i.e. population characteristics, condition indicators, and operating environment indicators) to effectively predict hazard and reliability. Another is that EHM explicitly investigates the relationship between condition and operating environment indicators associated with the hazard of an asset. Furthermore, the proportionality assumption, which most of the covariate-based hazard models suffer from it, does not exist in EHM. According to the sample size of failure/suspension times, EHM is extended into two forms: semi-parametric and non-parametric. The semi-parametric EHM assumes a specified lifetime distribution (i.e. Weibull distribution) in the form of the baseline hazard. However, for more industry applications, due to sparse failure event data of assets, the analysis of such data often involves complex distributional shapes about which little is known. Therefore, to avoid the restrictive assumption of the semi-parametric EHM about assuming a specified lifetime distribution for failure event histories, the non-parametric EHM, which is a distribution free model, has been developed. The development of EHM into two forms is another merit of the model. A case study was conducted using laboratory experiment data to validate the practicality of the both semi-parametric and non-parametric EHMs. The performance of the newly-developed models is appraised using the comparison amongst the estimated results of these models and the other existing covariate-based hazard models. The comparison results demonstrated that both the semi-parametric and non-parametric EHMs outperform the existing covariate-based hazard models. Future research directions regarding to the new parameter estimation method in the case of time-dependent effects of covariates and missing data, application of EHM in both repairable and non-repairable systems using field data, and a decision support model in which linked to the estimated reliability results, are also identified.

Le pronostic est le processus de prédiction de la durée de vie résiduelle utile (RUL) des composants, sous-systèmes ou systèmes. Cependant, jusqu'à présent, le pronostic a souvent été abordé au niveau composant sans tenir compte des interactions entre les composants et l'impact de l'environnement, ce qui peut conduire à une mauvaise prédiction du temps de défaillance dans des systèmes complexes. Dans ce travail, une approche de pronostic au niveau du système est proposée. Cette approche est basée sur un nouveau cadre de modélisation : le modèle d'inopérabilité entrée-sortie (IIM), qui permet de prendre en compte les interactions entre les composants et les effets du profil de mission et peut être appliqué pour des systèmes hétérogènes. Ensuite, une nouvelle méthodologie en ligne pour l'estimation des paramètres (basée sur l'algorithme de la descente du gradient) et la prédiction du RUL au niveau système (SRUL) en utilisant les filtres particulaires (PF), a été proposée. En détail, l'état de santé des composants du système est estimé et prédit d'une manière probabiliste en utilisant les PF. En cas de divergence consécutive entre les estimations a priori et a posteriori de l'état de santé du système, la méthode d'estimation proposée est utilisée pour corriger et adapter les paramètres de l'IIM. Finalement, la méthodologie développée, a été appliquée sur un système industriel réaliste : le Tennessee Eastman Process, et a permis une prédiction du SRUL dans un temps de calcul raisonnable
Prognostics is the process of predicting the remaining useful life (RUL) of components, subsystems, or systems. However, until now, the prognostics has often been approached from a component view without considering interactions between components and effects of the environment, leading to a misprediction of the complex systems failure time. In this work, a prognostics approach to system-level is proposed. This approach is based on a new modeling framework: the inoperability input-output model (IIM), which allows tackling the issue related to the interactions between components and the mission profile effects and can be applied for heterogeneous systems. Then, a new methodology for online joint system RUL (SRUL) prediction and model parameter estimation is developed based on particle filtering (PF) and gradient descent (GD). In detail, the state of health of system components is estimated and predicted in a probabilistic manner using PF. In the case of consecutive discrepancy between the prior and posterior estimates of the system health state, the proposed estimation method is used to correct and to adapt the IIM parameters. Finally, the developed methodology is verified on a realistic industrial system: The Tennessee Eastman Process. The obtained results highlighted its effectiveness in predicting the SRUL in reasonable computing time

Wind turbines components are subject to considerable stresses and fatigue due to extreme environmental conditions to which they are exposed, especially those located offshore. Also, the most common faults present in wind turbine components have been investigated for years by the research community and that has led to propose a fault diagnosis and fault tolerant control wind turbine benchmark which include a set of faults that affect the sensors and actuators of several wind turbine components. This thesis presents some contributions to the fields of fault diagnosis, fault-tolerant control, prognostics and its integration with wind turbine control which leads to proposing a control approach called health-aware model predictive control (HAMPC). The contributions are summarized below: - Model-based fault diagnosis: to perform fault detection and isolation interval-based observers together with a set of analytical redundant relations (ARRs) are obtained based on a structural analysis and the fault signature matrix that relates the ARRs with the faults. - Fault tolerant control: it is proposed a fault tolerant control scheme that integrates fault detection and an algorithm for fault accommodation. The scheme has the objective to avoid the increment of blades and tower loads when a fault in the rotor azimuth angle sensor occurs using the individual pitch control technique (IPC). - Wind turbine blades fatigue prognostics and degradation: fatigue is assessed using the rainflow counting algorithm which is used to estimate the accumulated damage and for degradation, it is used a stiffness degradation model of blades material which is used to make predictions of remaining useful life (RUL). - Wind turbines health control: the module for the health of the system based on fatigue damage estimation and RUL predictions is integrated with model predictive control (MPC) leading to the proposed control approach (HAMPC). The contributions presented in this thesis have been validated on a wind turbine study case that uses a 5MW wind turbine reference model implemented in a high fidelity wind turbine simulator (FAST).
Els components dels aerogeneradors estan sotmesos a considerable estrès i fatiga, degut a les condicions ambientals extremes a les quals estan exposats, especialment els localitzats en alta mar. Per aquest motiu, al comunitat científica durant els últims anys ha investigat les averies més comunes presents en els aerogeneradors, fet que ha portat a proposar un cas d'estudi de diagnosi i control tolerant de fallades que inclou un conjunt de fallades que afecten a diversos components dels aerogeneradors. Aquesta tesi presenta algunes contribucions en els camps de la diagnosi de fallades, el control tolerant de fallades i la prognosi, així com la seva integració amb el control d'aerogeneradors, fet que ha portat a proposar una tècnica de control anomenada control predictiu basada en models conscients de la salut del sistema (HAMPC). Concretament les aportacions es poden resumir en: - Diagnosi de fallades basada en models: per a la detecció s'utilitzen observadors intervalars i l'aïllament de la fallada es fa en base el conjunt d'ARRs obtinguts de l'anàlisi estructural i de la matriu de signatures de fallades que relaciona les ARRs amb les fallades. - Control tolerant de fallades: es proposa un esquema de control tolerant a fallades que integra la detecció de fallades i algoritme d'acomodació de fallades, i té per objectiu evitar l'augment de càrregues en la pala i la torre quan es produeix una fallada en el sensor azimuth quan es fa un control individual de la inclinació de les pales (IPC). - Prognosi de la fatiga i la degradació de les pales: la fatiga s'avalua amb un algorisme denominat "rainflow counting" amb el qual es fa estimació del dany acumulat i per a la degradació es fa servir un model de degradació de la rigidesa del material amb el qual es fan prediccions de la vida útil restant (RUL). - Control de la salut d'aerogeneradors: s'ha integrat la gestió de la salut del sistema basat en danys per fatiga o prediccions de RUL amb control predictiu basat en models (MPC) donant lloc al control que anomenem HAMPC. Les contribucions presentades en aquesta tesi han sigut validades en un cas d'estudi d'aerogeneradors basat en un aerogenerador de referència de 5MW de potència implementat en el simulador d'aerogeneradors d'alta fidelitat conegut amb el nom de FAST.

Dissertação de Mestrado Integrado em Medicina Veterinária
A cólica é uma patologia de importância preeminente em equinos, cuja identificação da causa nem sempre é fácil, fazendo com que a determinação precoce de um prognóstico seja essencial. Assim, foi realizado um estudo retrospetivo em 92 casos de cólica recebidos pelo Serviço de Cirurgia e Urgência em Equinos da FMV-ULisboa. Os objetivos do presente estudo foram: 1) caracterizar os casos de cólica referenciados para o SCUE FMV-ULisboa, avaliando o tipo de intervenção clínica, a causa de cólica e a taxa de alta hospitalar; 2) avaliar o valor prognóstico de cada um dos indicadores recolhidos na admissão; 3) comparar o valor destes indicadores entre os dois tipos de intervenção clínica, médica e cirúrgica; e 4) elaborar um modelo multivariado de predição de prognóstico. Estimou-se que 82% dos animais submetidos a intervenção cirúrgica e 75% dos animais tratados medicamente tiveram alta hospitalar, e que 25% dos animais submetidos a laparotomia sofreram íleo pós-cirúrgico. Foram recolhidos na admissão os seguintes dados: idade, tempo entre sinalização e admissão hospitalar, refluxo gastrointestinal, frequência cardíaca, hematócrito, proteínas totais séricas, proteínas totais do líquido peritoneal, lactato peritoneal e lactato sanguíneo. Nas cólicas médicas, os indicadores hematócrito, frequência cardíaca e lactato peritoneal foram considerados estatisticamente significativos (p<0,05), o lactato sanguíneo marginalmente significativo (p=0,053) e as proteínas do líquido peritoneal tendencialmente significativas (p<0,10). Foram elaborados dois modelos de predição multivariável. O modelo de 3 preditores (lactato sanguíneo, frequência cardíaca e hematócrito) com especificidade de 42,9% e sensibilidade de 96,0%. O modelo de 5 preditores (lactato sanguíneo, frequência cardíaca, hematócrito, idade e proteínas totais séricas) com especificidade de 71,4% e sensibilidade de 95,7%. Nas cólicas cirúrgicas, não foi possível determinar preditores significativos nem elaborar modelos de predição. Foi, ainda, criada uma aplicação informática de cálculo de probabilidade de alta hospitalar com base nos modelos descritos. Finalmente, conclui-se que a recolha de líquido peritoneal deverá ser feita com mais frequência pois os seus indicadores parecem transmitir informação valiosa. O modelo de 3 preditores, apesar de ter uma especificidade menor para a amostra em estudo, será provavelmente mais fiável do ponto de vista clínico, para utilização futura. Para além disso, é espectado que com o aumento da dimensão da amostra, estes modelos se tornem mais robustos.
ABSTRACT - A RETROSPECTIVE REVIEW OF 92 EQUINE COLIC CASES REFERRED FOR HOSPITAL TREATMENT - Colic is a really important syndrome in the equine species. To identify a diagnosis can be a true challenge, so the early determination of a prognosis is essential. Therefore, a retrospective study was performed in 92 colic cases admitted at the “Equine Surgery and Emergency Services” (Lisbon University). The objectives of this study were: 1) describe the colic cases and evaluate the clinical approach (medical or surgical), the origin of the problem and rate of survival; 2) estimate the prognostic value of each one of the collected predictors at the admission process; 3) compare the predictors according to the clinical approach; and 4) elaborate a multivariable prognostic prediction model. In this study, the survival rate was 82% for the horses submitted to surgical intervention and 75% for the horses treated medically; and, 25% of the horses in which laparotomy was performed developed post-operative ileus. The following data were collected during admission at the hospital: age, time between the onset of clinical signs and referral, gastrointestinal reflux, cardiac frequency, haematocrit, blood total protein, peritoneal fluid total protein, peritoneal fluid lactate, blood lactate. In medical colics, haematocrit, cardiac frequency and peritoneal fluid lactate were statistically significant (p<0,05), blood lactate was marginally significant (p=0,053) and peritoneal fluid total protein was tendentially significant (p<0,10). Two multivariable prognostic prediction models were elaborated. The three predictors model (blood lactate, cardiac frequency and haematocrit) had a specificity of 42,9% and a sensibility of 96,0%. The five predictors model (blood lactate, cardiac frequency, haematocrit, blood total protein and age) had a specificity of 71,4% and a sensibility of 95,7%. In surgical colics, it wasn’t possible to determine statistically significant predictors neither to elaborate prediction models. Based on the previously descript models, a computerized application to calculate the survival probability was created. It was concluded that peritoneal fluid should be collected more often, since peritoneal lactate and peritoneal fluid total protein seem to be providers of valuable information. Even though, the three predictors model has a reduced specificity for the study sample, it will be probably more reliable from the clinical point of view for further applications. Furthermore, it’s expected that with the increasing of the sample size, these models will get more robust.
N/A

Dissertação apresentada para obtenção do Grau de Doutor em Engenharia Electrotécnica e de Computadores – Sistemas Digitais e Percepcionais pela Universidade Nova de Lisboa, Faculdade de Ciências e Tecnologia
There are several clinical prognostic models in the medical field. Prior to clinical use, the outcome models of longitudinal cohort data need to undergo a multi-centre evaluation of their predictive accuracy. This thesis evaluates the possible gain in predictive accuracy in multicentre evaluation of a flexible model with Bayesian regularisation, the (PLANN-ARD), using a reference data set for breast cancer, which comprises 4016 records from patients diagnosed during 1989-93 and reported by the BCCA, Canada, with follow-up of 10 years. The method is compared with the widely used Cox regression model. Both methods were fitted to routinely acquired data from 743 patients diagnosed during 1990-94 at the Christie Hospital, UK, with follow-up of 5 years following surgery. Methodological advances developed to support the external validation of this neural network with clinical data include: imputation of missing data in both the training and validation data sets; and a prognostic index for stratification of patients into risk groups that can be extended to non-linear models. Predictive accuracy was measured empirically with a standard discrimination index, Ctd, and with a calibration measure, using the Hosmer-Lemeshow test statistic. Both Cox regression and the PLANN-ARD model are found to have similar discrimination but the neural network showed marginally better predictive accuracy over the 5-year followup period. In addition, the regularised neural network has the substantial advantage of being suited for making predictions of hazard rates and survival for individual patients. Four different approaches to stratify patients into risk groups are also proposed, each with a different foundation. While it was found that the four methodologies broadly agree, there are important differences between them. Rules sets were extracted and compared for the two stratification methods, the log-rank bootstrap and by direct application of regression trees, and with two rule extraction methodologies, OSRE and CART, respectively. In addition, widely used clinical breast cancer prognostic indexes such as the NPI, TNM and St. Gallen consensus rules, were compared with the proposed prognostic models expressed as regression trees, concluding that the suggested approaches may enhance current practice. Finally, a Web clinical decision support system is proposed for clinical oncologists and for breast cancer patients making prognostic assessments, which is tailored to the particular characteristics of the individual patient. This system comprises three different prognostic modelling methodologies: the NPI, Cox regression modelling and PLANN-ARD. For a given patient, all three models yield a generally consistent but not identical set of prognostic indices that can be analysed together in order to obtain a consensus and so achieve a more robust prognostic assessment of the expected patient outcome.

OBJECTIVE. To synthesize the evidence regarding prognostic models to predict mortality in patients diagnosed with pulmonary tuberculosis. METHODOLOGY. The current study followed the guidelines of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (The PRISMA Group, 2020) Statement. A literature search on prognostic models aimed to predict mortality in patients diagnosed with pulmonary tuberculosis was conducted by three revisers. We included prospective and retrospective studies where prognostic models predicting mortality were either developed or validated in patients diagnosed with pulmonary tuberculosis. Three reviewers independently assessed the quality of the included studies using the PROBAST tool. (¨Prediction model study Risk Of Bias Assessment Tool¨), which assesses both the risk of bias (RoB) and the applicability of each model. A descriptive analysis of each of the prediction models developed, their performance and the population characteristics of each article was conducted. RESULTS. Only 6 articles met the selection criteria. There was a total of 6 prognostic rules, one in each article. Most studies (5 out of 6) were retrospective cohorts, only 1 study was a prospective case-control study. When adding the population of all the studies, there were a total of 3,553 participants, with samples ranging from 103 participants to 1070 participants. All the studies had a high risk of bias according to the PROBAST tool in the overall assessment. The overall assessment showed that 3 studies had a low concern of applicability, 2 high concern and 1 unclear concern. Only 5 studies developed new prediction rules. In general, the presented models had a good discriminatory ability, with areas under the curve fluctuating between 0.65 up to 0.91. The predictive model with the highest discriminatory power was the one reported by Horita, et - al. with an AUC of 0.910 in the development cohort and 0.893 in the validation cohort. CONCLUSION. Considering that pulmonary tuberculosis is a highly prevalent disease in low-income countries, it would be very useful to have quality tools that allow healthcare personnel to be able to catalog patients with a higher risk of death so that they can receive priority medical attention.
OBJETIVO Sintetizar la evidencia acerca de modelos pronósticos que predicen mortalidad en pacientes con tuberculosis pulmonar. METODOLOGÍA. El siguiente estudio sigue las guías PRISMA del año 2020 (Preferred Reporting Items for Systematic Reviews and Meta-Analyses). Se realizó una búsqueda literaria, por tres revisores, de modelos pronósticos que se enfocaban en predecir mortalidad en pacientes diagnosticados con tuberculosis pulmonar. Se incluyeron estudios prospectivos y retrospectivos, donde los modelos pronósticos que predecían mortalidad habían sido desarrollados o validados en pacientes con tuberculosis pulmonar. De manera independiente, tres revisores evaluaron la calidad de los estudios incluidos usando la herramienta PROBAST (¨Prediction model study Risk Of Bias Assessment Tool¨), la cual evalúa el riesgo de sesgo y la aplicabilidad de cada modelo. Se realizó un análisis descriptivo de cada modelo de predicción, su performance, y las características de la población. RESULTADOS. Solo 6 artículos cumplieron los criterios de selección. Hubo un total de 6 modelos pronósticos, uno en cada artículo. La mayoría de los estudios (5 de 6) fueron cohortes retrospectivas, y solo uno fue un estudio de casos y controles prospectivo. Al sumar la población total de los estudios, hubo un total de 3,553 participantes, con muestras desde 103 hasta 1070 participantes. Todos los estudios obtuvieron un alto riesgo de sesgo, de acuerdo a la herramienta PROBAST, en la evaluación global. Además, la evaluación global mostró que 3 estudios obtuvieron una baja preocupación de aplicabilidad, 2 alta preocupación y un estudio preocupación indeterminada. Solo 5 estudios desarrollaron nuevas reglas de predicción, mientras que uno válido una ya existente. En general los modelos de predicción mostraron una buena habilidad discriminatoria, con valores de área bajo la curva que fluctuaban entre 0.65 hasta 0.91. El modelo de predicción con mayor poder discriminatorio fue el reportado por Horita, et – al con un valor de área bajo la curva de 0.910 en la cohorte de desarrollo y 0.893 en la cohorte de validación. CONCLUSIÓN. Tomando en cuenta que la tuberculosis pulmonar es una enfermedad prevalente en países de desarrollo, sería útil contar con herramientas que ayuden a los profesionales de la salud a catalogar a los pacientes con mayor riesgo de mortalidad, para que así ellos puedan recibir atención médica prioritaria.
Tesis

The development and implementation challenges associated with a proposed load allocation paradigm for fault risk assessment and system health management based on uncertain fault diagnostic and failure prognostic information are investigated. Health management actions are formulated in terms of a value associated with improving system reliability, and a cost associated with inducing deviations from a system's nominal performance. Three simulated case study systems are considered to highlight some of the fundamental challenges of formulating and solving an optimization on the space of available supervisory control actions in the described health management architecture. Repeated simulation studies on the three case-study systems are used to illustrate an empirical approach for tuning the conservatism of health management policies by way of adjusting risk assessment metrics in the proposed health management paradigm. The implementation and testing of a real-world prognostic system is presented to illustrate model development challenges not directly addressed in the analysis of the simulated case study systems. Real-time battery charge depletion prediction for a small unmanned aerial vehicle is considered in the real-world case study. An architecture for offline testing of prognostics and decision making algorithms is explained to facilitate empirical tuning of risk assessment metrics and health management policies, as was demonstrated for the three simulated case study systems.

Submitted by Pedro Barros (pedro.silvabarros@ufpe.br) on 2018-06-26T22:26:10Z No. of bitstreams: 2 license_rdf: 811 bytes, checksum: e39d27027a6cc9cb039ad269a5db8e34 (MD5) DISSERTAÇÃO Caio Bezerra Souto Maior.pdf: 3924685 bytes, checksum: 6968386bf75059f45ee80306322d2a56 (MD5)
Made available in DSpace on 2018-06-26T22:26:10Z (GMT). No. of bitstreams: 2 license_rdf: 811 bytes, checksum: e39d27027a6cc9cb039ad269a5db8e34 (MD5) DISSERTAÇÃO Caio Bezerra Souto Maior.pdf: 3924685 bytes, checksum: 6968386bf75059f45ee80306322d2a56 (MD5) Previous issue date: 2017-02-21
CAPES
The useful life time of equipment is an important variable related to reliability and maintenance. The knowledge about the useful remaining life of operation system by means of a prognostic and health monitoring could lead to competitive advantage to the corporations. There are numbers of models trying to predict the reliability’s variable behavior, such as the remaining useful life, from different types of signal (e.g. vibration signal), however several could not be realistic due to the imposed simplifications. An alternative to those models are the learning methods, used when exist many observations about the variable. A well-known method is Support Vector Machine (SVM), with the advantage that is not necessary previous knowledge about neither the function’s behavior nor the relation between input and output. In order to achieve the best SVM’s parameters, a Particle Swarm Optimization (PSO) algorithm is coupled to enhance the solution. Empirical Mode Decomposition (EMD) and Wavelets rise as two preprocessing methods seeking to improve the input data analysis. In this paper, EMD and wavelets are used coupled with PSO+SVM to predict the rolling bearing Remaining Useful Life (RUL) from a vibration signal and compare with the prediction without any preprocessing technique. As conclusion, EMD models presented accurate predictions and outperformed the other models tested.
O tempo de vida útil de um equipamento é uma importante variável relacionada à confiabilidade e à manutenção, e o conhecimento sobre o tempo útil remanescente de um sistema em operação, por meio de um monitoramento do prognóstico de saúde, pode gerar vantagens competitivas para as corporações. Existem diversos modelos utilizados na tentativa de prever o comportamento de variáveis de confiabilidade, tal como a vida útil remanescente, a partir de diferentes tipos de sinais (e.g. sinal de vibração), porém alguns podem não ser realistas, devido às simplificações impostas. Uma alternativa a esses modelos são os métodos de aprendizado, utilizados quando se dispõe de diversas observações da variável. Um conhecido método de aprendizado supervisionado é o Support Vector Machine (SVM), que gera um mapeamento de funções de entrada-saída a partir de um conjunto de treinamento. Para encontrar os melhores parâmetros do SVM, o algoritmo de Particle Swarm Optimization (PSO) é acoplado para melhorar a solução. Empirical Mode Decomposition (EMD) e Wavelets são usados como métodos pré-processamento que buscam melhorar a qualidade dos dados de entrada para PSO+SVM. Neste trabalho, EMD e Wavelets foram usadas juntamente com PSO+SVM para estimar o tempo de vida útil remanescente de rolamentos a partir de sinais de vibração. Os resultados obtidos com e sem as técnicas de pré-processamento foram comparados. Ao final, é mostrado que modelos baseados em EMD apresentaram boa acurácia e superaram o desempenho dos outros modelos testados.

In this work a data-driven prognostic approach based on AutoRegressive (AR) estimation and hidden Markov models (HMMs) is addressed. In particular, the approach is capable of achieving Prognostic and Health Management (PHM) tasks such as real time detection and Remaining Useful Life (RUL) estimation. The approach can be seen as composed of a training part (offline) and an exploitation part (online). The offline part relies upon the use of a scalar health indicator coming from the system identification field: the Itakura Saito (IS) spectral distance. In particular, raw acceleration data, gathered in an unsupervised framework from the machine, are modeled by AR processes and then transformed into IS. Then, HMMs are used to map such IS signals into a finite number of parameters. Moreover, in the training procedure of HMMs, a left-to-right clustering of unsupervised data, based on Mixture of Gaussians (MOG) distribution is proposed. During the online exploitation a simulation of a running signal is tested against trained ones in order to carry out PHM tasks in real time. Simulations have been performed using a public benchmark available in ”NASA prognostic data repository”. It contains run-to-failure tests on bearings, on which acceleration signals are gathered. In particular the gathering experiment simulates an industry application, under constant operating conditions. Results of simulations, performed on real time data, validate the proposed prognostic approach and make the combined use of IS an HMMs a reliable way in achieving PHM goals.

Despite professional requirements, educational research across disciplines, provides limited evidence indicating undergraduate health students, are confident with or intend to use evidence in their clinical practice after graduation. Using Bandura's self-efficacy theory, this research investigated factors influencing undergraduate nursing and paramedicine students' intention to use and their current use of evidence-based practice (EBP). Through development and validation of two multivariate prediction models, the study identified EBP self-efficacy as one important factor necessary for supporting students' intentions to translate EBP into clinical contexts. The research results provide theoretically-based components for curriculum developers when designing strategies to support students' advancement in EBP.

Modelos preditivos estimam o risco de eventos ou agravos relacionados à saúde e podem ser utilizados como ferramenta auxiliar em tomadas de decisão por gestores e profissionais de saúde. Algoritmos de machine learning (ML), por sua vez, apresentam potencial para identificar relações complexas e não-lineares presentes nos dados, com consequências positivas na performance preditiva desses modelos. A presente pesquisa objetivou aplicar técnicas supervisionadas de ML e comparar sua performance em problemas de classificação e de regressão para predizer respostas de interesse para a saúde pública e a medicina. Os resultados e discussão estão organizados em três artigos científicos. O primeiro apresenta um tutorial para o uso de ML em pesquisas de saúde, utilizando como exemplo a predição do risco de óbito em até 5 anos (frequência do desfecho 15%; n=395) para idosos do estudo \"Saúde, Bem-estar e Envelhecimento\" (n=2.677), segundo variáveis relacionadas ao seu perfil demográfico, socioeconômico e de saúde. Na etapa de aprendizado, cinco algoritmos foram aplicados: regressão logística com e sem penalização, redes neurais, gradient boosted trees e random forest, cujos hiperparâmetros foram otimizados por validação cruzada (VC) 10-fold. Todos os modelos apresentaram área abaixo da curva (AUC) ROC (Receiver Operating Characteristic) maior que 0,70. Para aqueles com maior AUC ROC (redes neurais e regressão logística com e sem penalização) medidas de qualidade da probabilidade predita foram avaliadas e evidenciaram baixa calibração. O segundo artigo objetivou predizer o risco de tempo de vida ajustado pela qualidade de vida de até 30 dias (frequência do desfecho 44,7%; n=347) em pacientes com câncer admitidos em Unidade de Terapia Intensiva (UTI) (n=777), mediante características obtidas na admissão do paciente à UTI. Seis algoritmos (regressão logística com e sem penalização, redes neurais, árvore simples, gradient boosted trees e random forest) foram utilizados em conjunto com VC aninhada para estimar hiperparâmetros e avaliar performance preditiva. Todos os algoritmos, exceto a árvore simples, apresentaram discriminação (AUC ROC > 0,80) e calibração satisfatórias. Para o terceiro artigo, características socioeconômicas e demográficas foram utilizadas para predizer a expectativa de vida ao nascer de municípios brasileiros com mais de 10.000 habitantes (n=3.052). Para o ajuste do modelo preditivo, empregou-se VC aninhada e o algoritmo Super Learner (SL), e para a avaliação de performance, o erro quadrático médio (EQM). O SL apresentou desempenho satisfatório (EQM=0,17) e seu vetor de valores preditos foi utilizado para a identificação de overachievers (municípios com expectativa de vida superior à predita) e underachievers (município com expectativa de vida inferior à predita), para os quais características de saúde foram comparadas, revelando melhor desempenho em indicadores de atenção primária para os overachievers e em indicadores de atenção secundária para os underachievers. Técnicas para a construção e avaliação de modelos preditivos estão em constante evolução e há poucas justificativas teóricas para se preferir um algoritmo em lugar de outro. Na presente tese, não foram observadas diferenças substanciais no desempenho preditivo dos algoritmos aplicados aos problemas de classificação e de regressão analisados. Espera-se que a maior disponibilidade de dados estimule a utilização de algoritmos de ML mais flexíveis em pesquisas de saúde futuras.
Predictive models estimate the risk of health-related events or injuries and can be used as an auxiliary tool in decision-making by public health officials and health care professionals. Machine learning (ML) algorithms have the potential to identify complex and non-linear relationships, with positive implications in the predictive performance of these models. The present research aimed to apply various ML supervised techniques and compare their performance in classification and regression problems to predict outcomes of interest to public health and medicine. Results and discussion are organized into three articles. The first, presents a tutorial for the use of ML in health research, using as an example the prediction of death up to 5 years (outcome frequency=15%; n=395) in elderly participants of the study \"Saúde, Bemestar e Envelhecimento\" (n=2,677), using variables related to demographic, socioeconomic and health characteristics. In the learning step, five algorithms were applied: logistic regression with and without regularization, neural networks, gradient boosted trees and random forest, whose hyperparameters were optimized by 10-fold cross-validation (CV). The area under receiver operating characteristic (AUROC) curve was greater than 0.70 for all models. For those with higher AUROC (neural networks and logistic regression with and without regularization), the quality of the predicted probability was evaluated and it showed low calibration. The second article aimed to predict the risk of quality-adjusted life up to 30 days (outcome frequency=44.7%; n=347) in oncologic patients admitted to the Intensive Care Unit (ICU) (n=777), using patients\' characteristics obtained at ICU admission. Six algorithms (logistic regression with and without regularization, neural networks, basic decision trees, gradient boosted trees and random forest) were used with nested CV to estimate hyperparameters values and to evaluate predictive performance. All algorithms, with exception of basic decision trees, presented acceptable discrimination (AUROC > 0.80) and calibration. For the third article, socioeconomic and demographic characteristics were used to predict the life expectancy at birth of Brazilian municipalities with more than 10,000 inhabitants (n=3,052). Nested CV and the Super Learner (SL) algorithm were used to adjust the predictive model, and for evaluating performance, the mean squared error (MSE). The SL showed good performance (MSE=0.17) and its vector of predicted values was used for the identification of underachievers and overachievers (i.e. municipalities showing worse and better outcome than predicted, respectively). Health characteristics were analyzed revealing that overachievers performed better on primary health care indicators, while underachievers fared better on secondary health care indicators. Techniques for constructing and evaluating predictive models are constantly evolving and there is scarce theoretical justification for preferring one algorithm over another. In this thesis no substantial differences were observed in the predictive performance of the algorithms applied to the classification and regression problems analyzed herein. It is expected that increase in data availability will encourage the use of more flexible ML algorithms in future health research.

This thesis presents sorne contributions to the field of Health-Aware Control (HAC) of dynamic systems. In the first part of this thesis, a review of the concepts and methodologies related to reliability versus degradation and fault tolerant control versus health-aware control is presented. Firstly, in an attempt to unify concepts, an overview of HAC, degradation, and reliability modeling including some of the most relevant theoretical and applied contributions is given. Moreover, reliability modeling is formalized and exemplified using the structure function, Bayesian networks (BNs) and Dynamic Bayesian networks (DBNs) as modeling tools in reliability analysis. In addition, some Reliability lmportance Measures (RIMs) are presented. In particular, this thesis develops BNs models for overall system reliability analysis through the use of Bayesian inference techniques. Bayesian networks are powerful tools in system reliability assessment due to their flexibility in modeling the reliability structure of complex systems. For the HAC scheme implementation, this thesis presents and discusses the integration of actuators health information by means of RIMs and degradation in Model Predictive Control (MPC) and Linear Quadratic Regulator algorithms. In the proposed strategies, the cost function parameters are tuned using RIMs. The methodology is able to avoid the occurrence of catastrophic and incipient faults by monitoring the overall system reliability. The proposed HAC strategies are applied to a Drinking Water Network (DWN) and a multirotor UAV system. Moreover, a third approach, which uses MPC and restricts the degradation of the system components is applied to a twin rotor system. Finally, this thesis presents and discusses two reliability interpretations. These interpretations, namely instantaneous and expected, differ in the manner how reliability is evaluated and how its evolution along time is considered. This comparison is made within a HAC framework and studies the system reliability under both approaches.
Aquesta tesi presenta algunes contribucions al camp del control basat en la salut dels components "Health-Aware Control" (HAC) de sistemes dinàmics. A la primera part d'aquesta tesi, es presenta una revisió dels conceptes i metodologies relacionats amb la fiabilitat versus degradació, el control tolerant a fallades versus el HAC. En primer lloc, i per unificar els conceptes, s'introdueixen els conceptes de degradació i fiabilitat, models de fiabilitat i de HAC incloent algunes de les contribucions teòriques i aplicades més rellevants. La tesi, a més, el modelatge de la fiabilitat es formalitza i exemplifica utilitzant la funció d'estructura del sistema, xarxes bayesianes (BN) i xarxes bayesianes dinamiques (DBN) com a eines de modelat i anàlisi de la fiabilitat com també presenta algunes mesures d'importància de la fiabilitat (RIMs). En particular, aquesta tesi desenvolupa models de BNs per a l'anàlisi de la fiabilitat del sistema a través de l'ús de tècniques d'inferència bayesiana. Les xarxes bayesianes són eines poderoses en l'avaluació de la fiabilitat del sistema gràcies a la seva flexibilitat en el modelat de la fiabilitat de sistemes complexos. Per a la implementació de l?esquema de HAC, aquesta tesi presenta i discuteix la integració de la informació sobre la salut i degradació dels actuadors mitjançant les RIMs en algoritmes de control predictiu basat en models (MPC) i control lineal quadràtic (LQR). En les estratègies proposades, els paràmetres de la funció de cost s'ajusten utilitzant els RIMs. Aquestes tècniques de control fiable permetran millorar la disponibilitat i la seguretat dels sistemes evitant l'aparició de fallades a través de la incorporació d'aquesta informació de la salut dels components en l'algoritme de control. Les estratègies de HAC proposades s'apliquen a una xarxa d'aigua potable (DWN) i a un sistema UAV multirrotor. A més, un tercer enfocament fent servir la degradació dels actuadors com a restricció dins l'algoritme de control MPC s'aplica a un sistema aeri a dos graus de llibertat (TRMS). Finalment, aquesta tesi també presenta i discuteix dues interpretacions de la fiabilitat. Aquestes interpretacions, nomenades instantània i esperada, difereixen en la forma en què s'avalua la fiabilitat i com es considera la seva evolució al llarg del temps. Aquesta comparació es realitza en el marc del control HAC i estudia la fiabilitat del sistema en tots dos enfocaments.
Esta tesis presenta algunas contribuciones en el campo del control basado en la salud de los componentes “Health-Aware Control” (HAC) de sistemas dinámicos. En la primera parte de esta tesis, se presenta una revisión de los conceptos y metodologíasrelacionados con la fiabilidad versus degradación, el control tolerante a fallos versus el HAC. En primer lugar, y para unificar los conceptos, se introducen los conceptos de degradación y fiabilidad, modelos de fiabilidad y de HAC incluyendo algunas de las contribuciones teóricas y aplicadas más relevantes. La tesis, demás formaliza y ejemplifica el modelado de fiabilidad utilizando la función de estructura del sistema, redes bayesianas (BN) y redes bayesianas diná-micas (DBN) como herramientas de modelado y análisis de fiabilidad como también presenta algunas medidas de importancia de la fiabilidad (RIMs). En particular, esta tesis desarrolla modelos de BNs para el análisis de la fiabilidad del sistema a través del uso de técnicas de inferencia bayesiana. Las redes bayesianas son herramientas poderosas en la evaluación de la fiabilidad del sistema gracias a su flexibilidad en el modelado de la fiabilidad de sistemas complejos. Para la implementación del esquema de HAC, esta tesis presenta y discute la integración de la información sobre la salud y degradación de los actuadores mediante las RIMs en algoritmos de control predictivo basado en modelos (MPC) y del control cuadrático lineal (LQR). En las estrategias propuestas, los parámetros de la función de coste se ajustan utilizando las RIMs. Estas técnicas de control fiable permitirán mejorar la disponibilidad y la seguridad de los sistemas evitando la aparición de fallos a través de la incorporación de la información de la salud de los componentes en el algoritmo de control. Las estrategias de HAC propuestas se aplican a una red de agua potable (DWN) y a un sistema UAV multirotor. Además, un tercer enfoque que usa la degradación de los actuadores como restricción en el algoritmo de control MPC se aplica a un sistema aéreo con dos grados de libertad (TRMS). Finalmente, esta tesis también presenta y discute dos interpretaciones de la fiabilidad. Estas interpretaciones, llamadas instantánea y esperada, difieren en la forma en que se evalúa la fiabilidad y cómo se considera su evolución a lo largo del tiempo. Esta comparación se realiza en el marco del control HAC y estudia la fiabilidad del sistema en ambos enfoques.

La maintenance prédictive joue un rôle important dans le maintien des systèmes de production continue car elle peut aider à réduire les interventions inutiles ainsi qu'à éviter des pannes imprévues. En effet, par rapport à la maintenance conditionnelle, la maintenance prédictive met en œuvre une étape supplémentaire, appelée le pronostic. Les opérations de maintenance sont planifiées sur la base de la prédiction des états de détérioration futurs et sur l'estimation de la vie résiduelle du système. Dans le cadre du projet européen FP7 SUPREME (Sustainable PREdictive Maintenance for manufacturing Equipment en Anglais), cette thèse se concentre sur le développement des modèles de détérioration stochastiques et sur des méthodes d'estimation de la vie résiduelle (Remaining Useful Life – RUL en anglais) associées pour les adapter aux cas d'application du projet. Plus précisément, les travaux présentés dans ce manuscrit sont divisés en deux parties principales. La première donne une étude détaillée des modèles de détérioration et des méthodes d'estimation de la RUL existant dans la littérature. En analysant leurs avantages et leurs inconvénients, une adaptation d’une approche de l'état de l'art est mise en œuvre sur des cas d'études issus du projet SUPREME et avec les données acquises à partir d’un banc d'essai développé pour le projet. Certains aspects pratiques de l’implémentation, à savoir la question de l'échange d'informations entre les partenaires du projet, sont également détaillées dans cette première partie. La deuxième partie est consacrée au développement de nouveaux modèles de détérioration et les méthodes d'estimation de la RUL qui permettent d'apporter des éléments de solutions aux problèmes de modélisation de détérioration et de prédiction de RUL soulevés dans le projet SUPREME. Plus précisément, pour surmonter le problème de la coexistence de plusieurs modes de détérioration, le concept des modèles « multi-branche » est proposé. Dans le cadre de cette thèse, deux catégories des modèles de type multi-branche sont présentées correspondant aux deux grands types de modélisation de l'état de santé des système, discret ou continu. Dans le cas discret, en se basant sur des modèles markoviens, deux modèles nommés Mb-HMM and Mb-HsMM (Multi-branch Hidden (semi-)Markov Model en anglais) sont présentés. Alors que dans le cas des états continus, les systèmes linéaires à sauts markoviens (JMLS) sont mis en œuvre. Pour chaque modèle, un cadre à deux phases est implémenté pour accomplir à la fois les tâches de diagnostic et de pronostic. A travers des simulations numériques, nous montrons que les modèles de type multi-branche peuvent donner des meilleures performances pour l'estimation de la RUL par rapport à celles obtenues par des modèles standards mais « mono-branche »
Predictive maintenance plays a crucial role in maintaining continuous production systems since it can help to reduce unnecessary intervention actions and avoid unplanned breakdowns. Indeed, compared to the widely used condition-based maintenance (CBM), the predictive maintenance implements an additional prognostics stage. The maintenance actions are then planned based on the prediction of future deterioration states and residual life of the system. In the framework of the European FP7 project SUPREME (Sustainable PREdictive Maintenance for manufacturing Equipment), this thesis concentrates on the development of stochastic deterioration models and the associated remaining useful life (RUL) estimation methods in order to be adapted in the project application cases. Specifically, the thesis research work is divided in two main parts. The first one gives a comprehensive review of the deterioration models and RUL estimation methods existing in the literature. By analyzing their advantages and disadvantages, an adaption of the state of the art approaches is then implemented for the problem considered in the SUPREME project and for the data acquired from a project's test bench. Some practical implementation aspects, such as the issue of delivering the proper RUL information to the maintenance decision module are also detailed in this part. The second part is dedicated to the development of innovative contributions beyond the state-of-the-are in order to develop enhanced deterioration models and RUL estimation methods to solve original prognostics issues raised in the SUPREME project. Specifically, to overcome the co-existence problem of several deterioration modes, the concept of the "multi-branch" models is introduced. It refers to the deterioration models consisting of different branches in which each one represent a deterioration mode. In the framework of this thesis, two multi-branch model types are presented corresponding to the discrete and continuous cases of the systems' health state. In the discrete case, the so-called Multi-branch Hidden Markov Model (Mb-HMM) and the Multi-branch Hidden semi-Markov model (Mb-HsMM) are constructed based on the Markov and semi-Markov models. Concerning the continuous health state case, the Jump Markov Linear System (JMLS) is implemented. For each model, a two-phase framework is carried out for both the diagnostics and prognostics purposes. Through numerical simulations and a case study, we show that the multi-branch models can help to take into account the co-existence problem of multiple deterioration modes, and hence give better performances in RUL estimation compared to the ones obtained by standard "single branch" models

La maintenance aéronautique est fortement régulée, notamment à travers l’établissement d’un planning de maintenance obligatoire, permettant de garantir la sureté structurale. La fréquence des arrêts en maintenance est déterminée de manière très conservative en vue d’assurer les exigences de fiabilité. Développer des stratégies de maintenance moins conservatives et plus efficaces peut alors représenter une voie pour une nouvelle croissance des compagnies aériennes. Les systèmes de monitoring embarqué de structures, sont progressivement introduits dans l’industrie aéronautique. Ces développements pourraient alors permettre de nouvelles stratégies de maintenance structurale basées sur la prévision de l’état de santé de chaque élément structural, plutôt que basée sur une maintenance programmée, tel qu’implémentée actuellement. Dans ce cadre général, ce travail se concentre sur le suivi par un système embarqué de la propagation de fissures de fatigue dans les panneaux de fuselage. Une nouvelle méthode de prévision des fissures basée sur des modèles de propagation est développée, qui permet de filtrer le bruit des mesures du système embarqué, identifier la taille actuelle de la fissure et prédire son évolution future et par conséquent la fiabilité des panneaux. Cette approche prédictive est intégrée dans le processus de maintenance structurale aéronautique et deux types de maintenances prédictives sont proposés. L’étude numérique montre que ces stratégies de maintenance prédictive peuvent réduire de manière significative les coûts de maintenance en réduisant le nombre d’arrêts en maintenance et le nombre de réparations inutiles
Aircraft maintenance represents a major economic cost for the aviation industry. Traditionally, the aircraft maintenance is highly regulated based on fixed schedules (thus called scheduled maintenance) in order to ensure safety. The frequency of scheduled maintenance is designed to be very conservative to maintain a desirable level of reliability. Developing efficient maintenance can be an important way for airlines to allow a new profit growth. With the development of sensor technology, structural health monitoring (SHM) system, which employ a sensor network sealing inside aircraft structures to monitor the damage state, are gradually being introduced in the aviation industry. Once it is possible to monitor the structure damage state automatically and continuously by SHM systems, it enables to plan the maintenance activities according to the actual or predicted health state of the aircraft rather than a fixed schedule. This work focus on the fatigue crack propagation in the fuselage panels. The SHM system is assumed to be employed. A model-based prognostics method is developed, which enables to filter the noise of SHM data to estimate the crack size, and to predict the future health state of the panels. This predictive information is integrated into the maintenance decision-making and two types of predictive maintenance are developed. The numerical study shows that the predictive maintenance significantly reduces the maintenance cost by reducing the number of maintenance stop and the repaired panels

Background: Venous thromboembolism (VTE), comprising the conditions of deep vein thrombosis (DVT) and pulmonary embolism (PE), is a common acute cardiovascular event associated with increased long-term morbidity, functional disability, all-cause mortality, and high rates of recurrence. Major advances in identification, prophylaxis, and treatment over the past 3-decades have likely changed its clinical epidemiology. However, there are little published data describing contemporary, population-based, trends in VTE prevention and management. Objectives: To examine recent trends in the epidemiology of clinically recognized VTE and assess the risk of recurrence after a first acute episode of VTE. Methods: We used population-based surveillance to monitor trends in acute VTE among residents of the Worcester, Massachusetts, metropolitan statistical area (WMSA) from 1985 through 2009, including in-hospital and ambulatory settings. Results: Among 5,025 WMSA residents diagnosed with acute PE and/or lower-extremity DVT between 1985 and 2009 (mean age = 65 years), 46% were men and 95% were white. Age- and sex-adjusted annual event rates (per 100, 000) of clinically recognized acute first-time and recurrent VTE was 142 overall, increasing from 112 in 1985/86 to 168 in 2009, due primarily to increases in PE occurrence. During this period, non-invasive diagnostic VTE testing increased, vi while treatment shifted from the in-hospital (chiefly with warfarin and unfractionated heparin) to out-patient setting (chiefly with low-molecular-weight heparins and newer anticoagulants). Among those with community-presenting first-time VTE, subsequent 3-year cumulative event rates of key outcomes decreased from 1999 to 2009, including all-cause mortality (41% to 26%), major bleeding episodes (12% to 6%), and recurrent VTE (17% to 9%). Active-cancer (with or without chemotherapy), a hypercoagulable state, varicose vein stripping, and Inferior vena cava filter placement were independent predictors of recurrence during short- (3-month) and long-term (3-year) follow-up after a first acute episode of VTE. We developed risk score calculators for VTE recurrence based on a 3-month prognostic model for all patients and separately for patients without active cancer. Conclusions: Despite advances in identification, prophylaxis, and treatment between 1985 and 2009, the disease burden from VTE in residents of central Massachusetts remains high, with increasing annual events. Declines in the frequency of major adverse outcomes between 1999 and 2009 were reassuring. Still, mortality, major bleeding, and recurrence rates remained high, suggesting opportunities for improved prevention and treatment. Clinicians may be able to use the identified predictors of recurrence and risk score calculators to estimate the risk of VTE recurrence and tailor outpatient treatments to individual patients.

Background: Venous thromboembolism (VTE), comprising the conditions of deep vein thrombosis (DVT) and pulmonary embolism (PE), is a common acute cardiovascular event associated with increased long-term morbidity, functional disability, all-cause mortality, and high rates of recurrence. Major advances in identification, prophylaxis, and treatment over the past 3-decades have likely changed its clinical epidemiology. However, there are little published data describing contemporary, population-based, trends in VTE prevention and management. Objectives: To examine recent trends in the epidemiology of clinically recognized VTE and assess the risk of recurrence after a first acute episode of VTE. Methods: We used population-based surveillance to monitor trends in acute VTE among residents of the Worcester, Massachusetts, metropolitan statistical area (WMSA) from 1985 through 2009, including in-hospital and ambulatory settings. Results: Among 5,025 WMSA residents diagnosed with acute PE and/or lower-extremity DVT between 1985 and 2009 (mean age = 65 years), 46% were men and 95% were white. Age- and sex-adjusted annual event rates (per 100, 000) of clinically recognized acute first-time and recurrent VTE was 142 overall, increasing from 112 in 1985/86 to 168 in 2009, due primarily to increases in PE occurrence. During this period, non-invasive diagnostic VTE testing increased, vi while treatment shifted from the in-hospital (chiefly with warfarin and unfractionated heparin) to out-patient setting (chiefly with low-molecular-weight heparins and newer anticoagulants). Among those with community-presenting first-time VTE, subsequent 3-year cumulative event rates of key outcomes decreased from 1999 to 2009, including all-cause mortality (41% to 26%), major bleeding episodes (12% to 6%), and recurrent VTE (17% to 9%). Active-cancer (with or without chemotherapy), a hypercoagulable state, varicose vein stripping, and Inferior vena cava filter placement were independent predictors of recurrence during short- (3-month) and long-term (3-year) follow-up after a first acute episode of VTE. We developed risk score calculators for VTE recurrence based on a 3-month prognostic model for all patients and separately for patients without active cancer. Conclusions: Despite advances in identification, prophylaxis, and treatment between 1985 and 2009, the disease burden from VTE in residents of central Massachusetts remains high, with increasing annual events. Declines in the frequency of major adverse outcomes between 1999 and 2009 were reassuring. Still, mortality, major bleeding, and recurrence rates remained high, suggesting opportunities for improved prevention and treatment. Clinicians may be able to use the identified predictors of recurrence and risk score calculators to estimate the risk of VTE recurrence and tailor outpatient treatments to individual patients.

Tese de mestrado em Ciência de Dados, 2021, Universidade de Lisboa, Faculdade de Ciências
A Esclerose Lateral Amiotrófica (ELA) é uma doença neurodegenerativa devastadora que causa degeneração rápida dos neurónios motores e geralmente leva à morte por falência respiratória. Não existe cura, pelo que o principal objetivo dos tratamentos consiste em melhorar os sintomas e prolongar a sobrevivência. A Ventilação não invasiva (VNI) é um tratamento eficaz que estende a expectativa de vida e melhora a sua qualidade. Neste contexto, é imperativo prever a necessidade de VNI de modo a administrá-¬la preventiva e adequadamente. Assim, propomos utilizar métodos de extração de padrões transacionais e sequenciais para descobrir padrões de apresentação e progressão da doença, respetivamente. Isto é feito analisando dados estáticos recolhidos aquando do diagnóstico, e dados longitudinais recolhidos no seguimento dos pacientes. O objetivo é utilizar estes padrões como variáveis em modelos de prognóstico, assim permitindo a utilização da progressão da doença na predição e melhorando a interpretação dos modelos. Inicialmente previmos a necessidade de VNI a 90, 180 e 365 dias da última consulta (predições a curto, médio e longo prazo) para os pacientes da coorte portuguesa, através do Portuguese ALS dataset. Os modelos de prognósticos obtiveram resultados promissores, especialmente quando introduzidas medidas de similaridade que permitem a verificação parcial de padrões. A avaliação de padrões através de taxas de crescimento de suporte entre classes sugere que a função bulbar e amplitude de resposta do nervo frénico, para além da função respiratória, são variáveis significantes para determinar a evolução dos pacientes. Isto confirma o conhecimento clínico em relação a biomarcadores de progressão de doença para a insuficiência respiratória. Devido a heterogeneidade existente entre doentes de ELA, dividimos também os pacientes em três grupos de progressão: Rápida, Neutra e Lenta, de acordo com as taxas de declínio da escala ALS¬FRS¬R, aplicando a abordagem a cada um dos grupos individualmente. Os resultados melhoraram significativamente relativamente àqueles obtidos inicialmente. Também foram exploradas janelas de predição específicas a cada grupo, que obtiveram resultados igualmente bons. A avaliação de padrões pode indicar que a função bulbar é mais relevante para a necessidade de VNI em pacientes com progressão da doença mais lenta, sendo as variáveis respiratórias os principais indicadores em pacientes com progressão rápida.
Amyotrophic Lateral Sclerosis (ALS) is a devastating neurodegenerative disease causing rapid degeneration of motor neurons and usually leading to death by respiratory failure. Since there is no cure, treatment’s goal is to improve symptoms and prolong survival. Non¬invasive Ventilation (NIV) is an effective treatment, leading to extended life expectancy and improved quality of life. In this scenario, it is paramount to predict its need in order to allow preventive or timely administration. Thus, we propose to use itemset mining together with sequential pattern mining to unravel disease presentation patterns together with disease progression patterns by analysing, respectively, static data collected at diagnosis and longitudinal data from patient follow¬up. The goal is to use these static and temporal patterns as features in prognostic models, enabling to take disease progression into account in predictions and promoting model interpretability. We initially predict the need for NIV within 90, 180 and 365 days of the last appointment (short, mid and long¬term predictions) for the portuguese ALS cohort, through the Portuguese ALS dataset. The learnt prognostic models are promising, especially when using similarity measures to assess partial pattern verification. Pattern evaluation through support growth rates suggests bulbar function and phrenic nerve response amplitude, additionally to respiratory function, are significant features towards determining patient evolution. This confirms clinical knowledge regarding relevant biomarkers of disease progression towards respiratory insufficiency. Due to known heterogeneity among ALS patients, we have also split the population into three progression groups: Fast, Neutral and Slow progressors, according to decline rates of the ALS¬FRS¬R score, and applied the approach separately to each group. Results have greatly improved from the initial approach. Group ¬specific prediction windows have also been explored, obtaining equally good results. Pattern evaluation through growth rates may indicate that bulbar function is more relevant towards needing NIV in patients with slower disease progression, with respiratory tests being the main indicators in Fast progressors.

Aims: This thesis explored and examined the clinical factors associated with the outcomes of chemotherapy-induced febrile neutropenia for adult cancer patients and confirms the independent predictive value of these factors. Established as predictors, the factors were used to formulate a multivariable prognostic model to stratify patients according to their risk groupings (high- or low-risk) for adverse outcomes for febrile neutropenia. Newly developed models underwent preliminary validation for their performance as prognostic models for febrile neutropenia outcomes. Background: Accuracy in risk stratification for cancer patients presenting with chemotherapy-induced febrile neutropenia is of critical importance. Serious morbidity may result when treatment is tailored according to misclassified levels of risk. New predictors and prediction tools used for risk stratification have been reported in the recent years. A systematic review was conducted on this topic as part of the thesis and the findings showed a lack of conclusive information on predictive values for some factors identified as predictors, and limitations in prognostic research studies’ methodologies which affect the internal and external validity of the risk prediction tools. Methods: Clinical factors identified through the systematic review contributed to the candidate factors investigated. Additional factors were also included based on other primary studies not included in the systematic review. A retrospective review of patients’ medical records was conducted. Tests of association using univariate analysis were conducted on these variables. Significant variables were tested and adjusted for confounders in a multivariate logistic regression analysis to formulate a multivariable tool for risk stratification of patients presenting with febrile neutropenia. Results: Predictive values for some variables were re-established while some variables failed to demonstrate their predictive values in a univariate analysis. After statistically adjusting to the current factors used in existing prognostic models, a new risk prediction tool was developed predict the risk of adverse outcomes. This tool has been subjected to preliminary validation that confirmed its potential utility. Limitations of the study included single-centre data and the small sample size. Conclusions: Application of a risk prediction tool has its benefits and limitations. However, enhancement of the methodological rigor and comprehensiveness of reporting of results in prognosis research needs to be emphasised for clarity in interpretation and implementation of the studies’ findings. Despite the promising initial validation of the tool developed in this thesis, further extensive validation and evaluation of the tool’s performance are needed to show the true impact of the tool on clinical practice.
Thesis (Ph.D.) -- University of Adelaide, School of Translational Health Science, 2014

博士
國立彰化師範大學
機電工程學系
107
In this work, a compositional prognostic-based assessment using the gamma process and Monte Carlo simulation was implemented to monitor the likelihood values of test Lithium-ion batteries on the failure threshold associated with capacity loss whose evaluation used a novel dual dynamic stress accelerated degradation test, called D2SADT for the test LiFePO4 batteries. D2SADT is an enable technique developed by us to simulate a situation when driving an electric vehicle in the city. The Norris and Landzberg reliability model was applied to estimate activation energy of the test batteries. The test results show that the battery capacity always decreased at each measurement time-step during D2SADT to enable the novel test method. The variation of the activation energies for the test batteries indicate that the capacity loss of the test battery operated under certain power and temperature cycling conditions, which can be accelerated when the charge–discharge cycles increase. The modeling results show that the gamma process combined with Monte Carlo simulations provides superior accuracy for predicting the lifetimes of the test batteries compared with the baseline lifetime data (i.e., real degradation route and lifetimes). The results presented high prediction quality for the proposed model as the error rates were close-or-within ±5% and were obtained for all test batteries after a certain quantity of capacity loss. In conclusion, the proposed model and test method could help engineers not only understand the degradation behavior according to the indicator of activation energy, but also enable monitoring of the health states of Li-ion batteries more precisely in certain real conditions.

Trabalho de Projeto do Mestrado Integrado em Engenharia Biomédica apresentado à Faculdade de Ciências e Tecnologia
A previsão do tempo de internamento dos pacientes é de grande importância para os hospitais, uma vez que pode determinar a utilização de recursos, melhorar o agendamento de futuros internamentos e cirurgias, e auxiliar no planeamento dos cuidados de saúde dos pacientes, desde a admissão até à alta. Consequentemente, uma melhor qualidade dos cuidados de saúde prestados pode ser proporcionada aos pacientes, sendo este o principal objetivo dos hospitais. Neste projecto, quatro abordagens diferentes foram implementadas para desenvolver modelos de previsão de tempo de internamento: i) exploração de modelos de risco existentes (SCORE), ii) aplicação de modelos típicos de inteligência computacional (Random Forest, Support Vector Machine e Multilayer Perceptron), iii) desenvolvimento de um modelo interpretável e personalizável ao paciente com base em regras e iv) integração de dados dinâmicos (sinais vitais) nos modelos anteriores. Os dados clínicos usados neste trabalho foram fornecidos pelo CHUC (Centro Hospitalar e Universitário de Coimbra) e pela Philips Electronics Nederland B.V., compreendendo 1544 pacientes admitidos na unidade de cuidados intensivos de cardiologia do Hospital dos Covões (Coimbra) e 189 pacientes bariátricos admitidos para cirurgia no Catharina Hospital (Eindhoven), respetivamente.O conjunto inicial de variáveis dos pacientes cardíacos foi obtido através de uma revisão da literatura e do conhecimento clínico de um cardiologista da unidade de cuidados intensivos de cardiologia do CHUC. Para os pacientes bariátricos, este conjunto resultou de uma revisão da literatura para a determinação das variáveis relevantes. Posteriormente, as variáveis de entrada dos modelos de previsão de tempo de internamento foram selecionadas desse conjunto inicial usando o coeficiente de correlação tau de Kendall. Adicionalmente, as variáveis de entrada selecionadas para os pacientes cardíacos foram também validadas pelo cardiologista. O desempenho dos modelos referidos, medido através da média geométrica (GE) e do F1 score, foi determinado aplicando este conjunto final de variáveis de entrada a cada um deles.Finalmente, através da aplicação do teste Friedman e do correspondente teste post-hoc Nemenyi, foi possível ordenar os modelos em função do seu desempenho.A performance do modelo baseado no SCORE foi significativamente baixa, obtendo uma GE de 0.50. Assim, apesar deste modelo de risco ser de grande importância na prática cardiológica europeia, não é adequado para estimar o tempo de internamento hospitalar. A segunda abordagem (modelo Black-box) superou o modelo anterior. Os melhores resultados foram obtidos pelo Multilayer perceptron com uma GE de 0.62 ± 0.03 para os pacientes cardíacos e 0.64 ± 0.08 para os bariátricos, respetivamente. Verificou-se ainda que o desempenho do modelo interpretável e personalizável foi superior ao modelo Black-box, para os dois tipos de pacientes, com uma GE de 0.66 ± 0.02 para os pacientes cardíacos e 0.83 ± 0.05 para os pacientes bariátricos. Adicionalmente, a inclusão de sinais vitais aos modelos de previsão mostrou-se vantajosa por levar a um aumento da performance em todos os classificadores. Estes resultados sugerem que a incorporação de dados dinâmicos em modelos de previsão de tempo de internamento deve ser explorada de forma aprofundada em estudos posteriores.A análise dos resultados permitiu-nos concluir que, apesar de aceitável, a performance dos modelos desenvolvidos não parece ser adequada para o seu uso na prática clínica (GE máxima de 0.66 e 0.83 para os pacientes cardíacos e bariátricos, respetivamente). Este facto pode-se justificar pela dificuldade e complexidade que o problema apresenta. O estudo de outras variáveis, não só determinadas aquando a admissão, mas durante as primeiras horas ou no primeiro dia de internamento do doente, poderia ser uma estratégia a explorar no futuro.
Predicting the patients' length of stay (LOS) is of major importance for hospitals, since it can determine the resource utilization, improve the scheduling of admissions and surgeries and helping in the development of effective clinical pathways. Consequently, a better quality of care can be provided to the patients, which is the main goal of the hospitals.In this project, four different approaches were implemented to develop LOS prediction models: i) exploration of available risk tools (SCORE), ii) application of typical computational intelligence models (Random Forest, Support Vector Machine and Multilayer Perceptron), iii) development of an interpretable and patient customized model based on rules and iv) integration of dynamic data (vital signs) in the previous models. The clinical data used in this work was provided by the CHUC (Coimbra Hospital and University Center) and by Philips Electronics Nederland B.V., comprising 1544 patients admitted in the cardiac intensive care unit of Hospital dos Covões (Coimbra) and 189 bariatric patients admitted to surgery in Catharina Hospital (Eindhoven), respectively.The initial set of features of the cardiac patients was obtained through a literature review and the clinical knowledge of an ICU cardiologist of CHUC. For the bariatric patients, this set resulted from a literature review for the determination of the relevant features. Then, the input features of the LOS prediction models were selected from this initial set using the Kendall's tau coefficient correlation. Moreover, the selected input features for the cardiac patients were also validated by the cardiologist. The performance of the referred models, measured in terms of the geometric mean (GE) and F1 score, was determined by employing this final set of input variables to each one of them. Finally, through the application of the Friedman test and the corresponding post-hoc Nemenyi test, it was possible to order the models according to their performance.The SCORE model performance was significantly low, achieving a geometric mean (GE) of 0.50. Thus, although this risk tool is of high importance in the European cardiology practice, it is not sufficiently accurate to estimate the actual LOS. The second approach (Black-box model) outperformed the previous model. The best results were achieved by the multilayer perceptron with a GE of 0.62 ± 0.03 for the cardiac patients and 0.64 ± 0.08 for the bariatric ones. Furthermore, we verified that the performance of the interpretable and customized model was higher than the Black-box model, for both types of patients, obtaining a GE of 0.66 ± 0.02 for the cardiac patients and 0.83 ± 0.05 for the bariatric patients. Moreover, the addition of the vital signs to the prediction models was proved to be advantageous since it leaded to an increase of performance in all the classifiers. These results suggest that the incorporation of dynamic data in LOS prediction models is worthy of further exploratory studies.The analysis of the results allowed us to conclude that, although acceptable, the performance of the developed models does not seem to be adequate for their use in clinical practice (maximum GE of 0.66 and 0.83 for the cardiac and bariatric patients, respectively). This fact may be justified by the difficulty and complexity that the problem presents. The study of other variables, not only determined at admission time, but during the first hours or on the first day of the patient's stay, could be a strategy to explore in the future.

Personalized medicine offers the most effective treatment protocols to the individual Chronic Myeloid Leukemia (CML) patients. Understanding the molecular biology that causes CML assists in providing efficient treatment. After the identification of an activated tyrosine kinase BCR-ABL1 as the causative lesion in CML, the first-generation Tyrosine Kinase inhibitors (TKI) imatinib (Glivec®), were developed to inhibit BCR-ABL1 activity and approved as a treatment for CML. Despite the remarkable increase in the survival rate of CML patients treated with imatinib, some patients discontinued imatinib therapy due to intolerance, resistance or progression. These patients may benefit from the use of secondgeneration TKIs, such as nilotinib (Tasigna®) and dasatinib (Sprycel®). All three of these TKIs are currently approved for use as frontline treatments. Prognostic scores and molecularbased predictive assays are used to personalize the care of CML patients by allocating risk groups and predicting responses to therapy. Although prognostic scores remain in use today, they are often inadequate for three main reasons. Firstly, since each prognostic score may generate conflicting prognoses for the risk index and it can be difficult to know how to treat patients with conflicting prognoses. Secondly, since prognostic score systems are developed over time, patients can benefit from newly developed systems and information. Finally, the earlier scores use mostly clinically oriented factors instead of those directly related to genetic or molecular indicators. As the current CML treatment guidelines recommend the use of TKI therapy, a new tool that combines the well-known, molecular-based predictive assays to predict molecular response to TKI has not been considered in previous research. Therefore, the main goal of this research is to improve the ability to manage CML disease in individual CML patients and support CML physicians in TKI therapy treatment selection by correctly allocating patients to risk groups and predicting their molecular response to the selected treatment. To achieve this objective, the research detailed here focuses on developing a prognostic model and a predictive model for use as a personalized medicine support system. The system will be considered a knowledge-based clinical decision support system that includes two models embedded in a decision tree. The main idea is to classify patients into risk groups using the prognostic model, while the patients identified as part of the high-risk group should be considered for more aggressive imatinib therapy or switched to secondgeneration TKI with close monitoring. For patients assigned to the low-risk group to imatinib should be predicted using the predictive model. The outcomes should be evaluated by comparing the results of these models with the actual responses to imatinib in patients from a previous medical trial and from patients admitted to hospitals. Validating such a predictive system could greatly assist clinicians in clinical decision-making geared toward individualized medicine. Our findings suggest that the system provides treatment recommendations that could help improve overall healthcare for CML patients. Study limitations included the impact of diversity on human expertise, changing predictive factors, population and prediction endpoints, the impact of time and patient personal issues. Further intensive research activities based on the development of a new predictive model and the method for selecting predictive factors and validation can be expanded to other health organizations and the development of models to predict responses to other TKIs.
Thesis (Ph.D.) -- University of Adelaide, School of Computer Science, 2018