RCAAP Repository

Visually-Impaired People (VIP) that have the desire to partake in recreational activities, such as playing video games or exploring virtual environments will often find barriers with the lack of accessibility features available. This is mostly due to the fact that such environments are heavily designed with the visual video component in mind. Still, among the number of people with visual impairments, blind people have the most difficulty in adapting to these environments. Although there are options that suit their impairments, either digital audio or text-based games, these experiences will often offer a lesser recreational value compared to the more complex and detailed options available in the market of video games. In this work, we delved into how blind people experience Virtual Reality and Virtual Environments. We analysed video content from blind content producers, identifying their biggest barriers while playing video games that weren’t completely accessible. At the same time, we registered and discussed about the many ways, either via elements outside or from the games they play, that they managed to overcome such barriers with their own hands and ingenuity. A contribution on its own, these observed experiences are supported by demonstrative videos and may serve to guide developers on how to better create or modify Virtual Environments to be accessible to blind people. We went further in our study, using the indications we gathered previously and created a Virtual Environment of our own. The implementation of possible solutions for some of the most common barriers observed was evaluated through testing with 7 participants, who also shared their opinions and suggestions. We expect that this can serve as a practical example on how to initiate the creation or adaptation of Virtual Environments to be accessible for blind people

This work describes the designing, prototyping and real-life testing of an Internet of Things (IoT) wearable device, the SALUS device. The solution consists of a purpose-built device, carefully designed for hospital use, that measures Heart Rate (HR), Blood Oxygen Saturation (SpO2), several Heart Rate Variability (HRV) parameters, Breathing Rate (BR) and Peripheral Vascular Resistance (PVR), and, therefore, acquiring most of the data needed for a cardiovascular patient evaluation. One of the critical challenges was the calculation of a novel way of assessing PVR using PPG. To do so, a regression analysis was performed between the Full Width at Half Maximum (FWHM) of the PPG peak and the value of PVR measured by the Task Force Monitor (TFM). Healthcare professionals, however, need more information to evaluate a cardiovascular patient than PVR alone, namely, HRV indexes. The SALUS device can provide these values by using the HeartPy library to analyse the data gathered. SALUS was tested in Hospital de Santa Marta following the 2008 Front-loaded Glyceryl Trinitrate (GTN) head-up tilt protocol used for the tilt test implemented in the Syncope Unit of the Hospital de Santa Marta and the results compared with the gold standard devices TFM and uMEC10, a cardiovascular monitor. There were limitations to this study since the number of patients considered was small and the sample was clinically heterogeneous. However, the main goal of this pilot study was to show that the SALUS device could be used in a clinical context. Despite the limitations, the results obtained were positive for the values of BPM (R 2=0.99), SpO2 (R 2=0.83), and LF/HF (R 2=0.77). The wearable device here developed holds promise to be a valuable and dependable diagnostic tool, as shown by the values when compared to the gold standard. Validation of this instrument is an ongoing process. Further testing and confirmation of other parameters are needed to assess the improvement of this acquisition for rapid screening of cardiac dysautonomia and Cardiovascular Disease (CVD).

Energy consumption in the EU has been increasing at an astonishing rate in the recent years. The European Commission has decided counteract on energy consumed by buildings, which is one of the areas that has contributed more to this increase. More than 40% of energy consumed in the EU is currently spent in buildings, whereas more than 75% of buildings in this region are not energy efficient. The SATO project addresses these challenges by providing self-assessments towards optimization (SATO) of energy and comfort management in residential and service buildings from multiple partners. This thesis focuses on specific contributions in the design, development, deployment, and evaluation of a first prototype of the cloud-based platform being proposed by this project. More specifically, this thesis contributed to the design of a middleware component that connects the buildings to the platform and streams events from devices to where data is processed. The activities included exploring proprietary and open-source specialized platforms for Internet of Things (IoT), general-purpose streaming platforms, among others. The acquired knowledge on these topics enabled us to design a platform architecture that integrated our components with the ones from other frameworks developed in the SATO project. Additionally, the systematic deployment of the SATO platform is another contribution from this thesis that helped the project to deal with the cumbersome complexity of this type of platforms. Finally, this thesis identified some appropriate component placement strategies for the SATO platform, optimized and scaled up many of its components, and evaluated its first prototype to reach a throughput of up to 60 thousand messages (of 500 bytes each) per second.

Most Functional Magnetic Resonance Imaging (fMRI) studies with the goal of understanding the Functional Connectivity (FC) of the brain have applied Resting-State fMRI to study connectivity between different brain regions. With the objective of applying a new method, we developed a novel approach using a passive task-based fMRI acquisition created to induce the shifting of brain networks, and thus increase between network connectivity. This approach consists of an array of images and stimuli created to excite specific networks and thus to enhance Functional Connectivity. We applied this novel paradigm in both typically developing youth and youth with Anorexia Nervosa (AN). AN is a severe psychological disorder characterized by an obsessive desire to lose weight by refusing to eat. We hypothesised that, using only healthy participants, viewing the Passive-Task will create a stronger shifts between different networks, translated in a higher FC between regions of distinct networks, when compared to the shift of networks at Resting-State. Also we hypothesised that no significant differences within the connectivity of regions of the same networks would be found between Resting-State and Passive-Task. Results show a greater connectivity during the Passive-Task in areas related to Executive Control, such as the insula and the Lateral Prefrontal Cortex, with areas from the Secondary Visual and Lateral Parietal Occipital networks. Higher connectivity at rest was found in the majority of the pairs of connected regions, especially related to the Cerebellar-Occipital networks. Although our results were not as hypothesised, this work proposes a new view on the use of passive viewing tasks, in the study of FC and network shifts. In the future, the goal of the task is to be used to compare FC in cases of AN with controls.

Genetic interactions, both between genetic material and between this and the surrounding environment of a given individual are important factors for understanding the process of evolution of natural populations. The study of this interactions as well as their integration in evolutionary terms may have several applications such as, for example, understanding the observed prevalence of antibiotic resistance in natural populations. In this thesis the patterns of genetic interactions between multiple-resistances to antibiotics were explored. In particular, epistasis and genotype-by-environment interactions operating among antibiotic resistances were studied. To measure levels of epistasis occurring between multiple-antibiotic-resistances in the complete absence of antibiotics, mutants resistant to three antibiotics commonly used in clinic were first generated: nalidixic acid, rifampicin and streptomycin. With these mutants double resistant mutants were created. By measuring the costs of each mutation individually and jointly the type of epistasis operating between different sets of resistance mutations was determined. The cost of double resistance was majorly lower than expected, revealing the presence of positive epistasis between mutations conferring resistance to the studied antibiotics. The same patterns were observed for interactions between a set of resistance mutations and conjugative plasmids with multiple-resistance factors. However, in this case, extreme cases of positive epistasis coined sign epistasis were observed for a large fraction of combinations. This type of interactions was also observed in the previous study yet less frequently. Naturally occurring bacteria are often faced with a multitude of environments. The action of environmental changes in the effects of antibiotic resistance mutations was studied for a group of mutations resistant to four antibiotics in three environments differing in the CHAPTER I viii number of comprised environmental stresses. This study revealed, for most cases, a strong pattern of interactions between the different genotypes and the environment.

There is an urgent need for the development of new anti-HIV drugs that can complement existing medicines to be used against resistant strains. Here, we report the anti-HIV-1 peptide pepRF1, a human serum-resistant peptide derived from the Dengue virus capsid protein. In vitro, pepRF1 shows a 50% inhibitory concentration of 1.5 nM with a potential therapeutic window higher than 53 000. This peptide is specific for CXCR4-tropic strains, preventing viral entry into target cells by binding to the viral coreceptor CXCR4, acting as an antagonist of this receptor. pepRF1 is more effective than T20, the only peptide-based HIV-1 entry inhibitor approved, and excels in inhibiting a HIV-1 strain resistant to T20. Potentially, pepRF1 can be used alone or in combination with other anti-HIV drugs. Furthermore, one can also envisage its use as a novel therapeutic strategy for other CXCR4-related diseases.

We report the synthesis of water soluble cyclodextrin (CD) polymers prepared by crosslinking pyromellitic dianhydride (PMDA) with two CD derivatives (methyl-β-CD - MβCD and (2-hydroxy)propyl-β-CD - HPβCD) and their evaluation as functional sub-micron sized carriers in the development of antiretroviral drug delivery systems. Using the protease inhibitor lopinavir (LPV) as model drug, LPV loaded CD polymers (pHPβCD and pMβCD) were prepared and fully characterized. The physicochemical characterization and in vitro drug release confirmed the successful synthesis of pHPβCD and pMβCD, the formation of sub-micron sized particles and a 12–14 fold increase in LPV solubility. Cytotoxicity assays indicated that both pHPβCD and pMβCD were able to improve the safety profile of LPV while the viral infectivity assay revealed a concentration independent anti-HIV-1 effect for both pHPβCD and pMβCD with a maximum percentage inhibition (MPI) of 79 and 91% respectively. After LPV loading, the antiviral profile of pHPβCD was reversed to the sigmoidal dose–response profile of LPV, while pMβCD maintained its dose-independent profile followed by a LPV mediated increase in viral inhibition. Overall, both pHPβCD and pMβCD demonstrated anti-HIV-1 activity, while drug loaded pMβCD indicated its potential as functional sub-micron sized drug delivery polymers for achieving synergistic anti-HIV activity.

Introduction: Structural modulation of previously identified lead spiro-β-lactams with antimicrobial activity was carried out. Objective: The main objective of this work was to synthesize and evaluate the biological activity of novel spiro-lactams based on previously identified lead compounds with antimicrobial activity. Methods: The target chiral spiro-γ-lactams were synthesized through 1,3-dipolar cycloaddition reaction of a diazo-γ-lactam with electron-deficient dipolarophiles. In vitro activity against HIV and Plasmodium of a wide range of spiro-β-lactams and spiro-γ-lactams was evaluated. Among these compounds, one derivative with good anti-HIV activity and two with promising antiplasmodial activity (IC50 < 3.5 µM) were identified. Results: A novel synthetic route to chiral spiro-γ-lactams has been established. The studied β- and γ- lactams were not cytotoxic, and three compounds with promising antimicrobial activity were identified, whose structural modulation may lead to new and more potent drugs. Conclusion: The designed structural modulation of biologically active spiro-β-lactams involved the replacement of the four-membered β-lactam ring by a five-membered γ-lactam ring. Although conformational and superimposition computational studies revealed no significant differences between β- and γ- lactam pharmacophoric features, the studied structural modulation did not lead to compounds with a similar biological profile. The observed results suggest that the β-lactamic core is a requirement for the activity against both HIV and Plasmodium.

Currently, there are several route recommendation systems, such as Google Maps or Waze, that suggest to their users routes between two given locations while considering different means of travel (on foot, by car, or by public transport). These systems generally recommend the shortest, the fastest, and/or the cheapest route (where factors such as fuel consumption, toll payment, and/or ticket purchasing are considered), indicating a greater focus on public and private modes of transportation rather than on pedestrians or cyclists. That is, these services are strictly limited to economic criteria and do not allow for the integration of other factors that are relevant to each user. In urban environments, it is typical that a person’s choice of walking a certain path is not based on the shortest or the fastest route, but on other subjective criteria valued by pedestrians, such as safety or overall exposure to air pollutants, for instance. The main objective of this work is to design and develop a mobile system (named Mobile HealthyTrack) that recommends pedestrian routes for the city of Lisbon by considering, simultaneously, multiple criteria tailored to the preferences of each registered user, with the additional integration of a voting functionality in order to improve the overall quality of the recommendations. Mobile HealthyTrack is a mobile system (exclusive to the Android operating system) in which users input the points of origin and destination of the routes they want to travel, based on the preferences they have established for each of the three quality criteria considered for the system: distance, air quality, and proximity to green areas. This system combines various modules that work together to provide the most customizable and tailored experience possible for the user. The experiments carried out in a controlled offline environment, as well as the user studies conducted with volunteers, allowed not only to observe that the routes recommended by the system correspond to the best possible routes considering the established parameters, but also to attest to the application’s usability and to obtain essential feedback to further optimize the system.

Ultrasound neuromodulation is a promising ultrasound modality since it combines high spatial resolution and high coverage of the brain while remaining minimally invasive. It is performed by exciting a transducer with an alternating pulse signal at its resonance frequency. These pulse signals are typically sent in bursts. The piezoelectric transducer converts this signal into an acoustic wave, resulting in multiple acoustic waves within each burst. Bulk ceramic piezoelectric transducers have a high-quality factor, meaning that for a single pulse excitation, it produces several pulses with decaying amplitude. The main goal of this dissertation was to explore potential power savings in the ultrasound transmitter by removing pulses from a long burst of pulses and bursts from the neuromodulation cycle while minimizing the decay of the acoustic wave amplitude. This goal was accomplished by developing an experimental setup for ultrasound stimulation using two different transducers: High-Intensity Focused Ultrasound and Air-Backing transducers. In this study, it was also simulated two different approaches: removing pulses by short and open circuits. The results showed that using the Air-Backing transducer, it’s possible to save more energy than with the other transducer. The simulations from the Air-Backing transducer, in which only a percentage of the pulse amplitude was removed, showed better results than the total removal of the pulses, and among the results presented, the 30% amplitude removal is the case with the highest efficiency. Another comparison made in this project suggests that performing pulse removal through a short circuit generates very positive results while performing pulse removal through an open circuit was not beneficial to the hypothesis under study. Future work to be done should include experiments and simulations with neurons since the behavior of ultrasound in neurons is still not fully understood, and the obtained results may prove to be quite different from the simulations.

Ultrasound neuromodulation is a promising ultrasound modality since it combines high spatial resolution and high coverage of the brain while remaining minimally invasive. It is performed by exciting a transducer with an alternating pulse signal at its resonance frequency. These pulse signals are typically sent in bursts. The piezoelectric transducer converts this signal into an acoustic wave, resulting in multiple acoustic waves within each burst. Bulk ceramic piezoelectric transducers have a high-quality factor, meaning that for a single pulse excitation, it produces several pulses with decaying amplitude. The main goal of this dissertation was to explore potential power savings in the ultrasound transmitter by removing pulses from a long burst of pulses and bursts from the neuromodulation cycle while minimizing the decay of the acoustic wave amplitude. This goal was accomplished by developing an experimental setup for ultrasound stimulation using two different transducers: High-Intensity Focused Ultrasound and Air-Backing transducers. In this study, it was also simulated two different approaches: removing pulses by short and open circuits. The results showed that using the Air-Backing transducer, it’s possible to save more energy than with the other transducer. The simulations from the Air-Backing transducer, in which only a percentage of the pulse amplitude was removed, showed better results than the total removal of the pulses, and among the results presented, the 30% amplitude removal is the case with the highest efficiency. Another comparison made in this project suggests that performing pulse removal through a short circuit generates very positive results while performing pulse removal through an open circuit was not beneficial to the hypothesis under study. Future work to be done should include experiments and simulations with neurons since the behavior of ultrasound in neurons is still not fully understood, and the obtained results may prove to be quite different from the simulations.

The COVID-19 pandemic caused a rise in preprinting, triggered by the need for open and rapid dissemination of research outputs. We surveyed authors of COVID-19 preprints to learn about their experiences with preprinting their work and also with publishing their work in a peer-reviewed journal. Our research had the following objectives: 1. to learn about authors’ experiences with preprinting, their motivations, and future intentions; 2. to consider preprints in terms of their effectiveness in enabling authors to receive feedback on their work; 3. to compare the impact of feedback on preprints with the impact of comments of editors and reviewers on papers submitted to journals. In our survey, 78% of the new adopters of preprinting reported the intention to also preprint their future work. The boost in preprinting may therefore have a structural effect that will last after the pandemic, although future developments will also depend on other factors, including the broader growth in the adoption of open science practices. A total of 53% of the respondents reported that they had received feedback on their preprints. However, more than half of the feedback was received through “closed” channels–privately to the authors. This means that preprinting was a useful way to receive feedback on research, but the value of feedback could be increased further by facilitating and promoting “open” channels for preprint feedback. Almost a quarter of the feedback received by respondents consisted of detailed comments, showing the potential of preprint feedback to provide valuable comments on research. Respondents also reported that, compared to preprint feedback, journal peer review was more likely to lead to major changes to their work, suggesting that journal peer review provides significant added value compared to feedback received on preprints.

Metagenomic next-generation sequencing (mNGS), the shotgun sequencing of RNA and DNA from clinical samples, has proved useful for broad-spectrum pathogen detection and the genomic surveillance of viral outbreaks. An additional target enrichment step is generally needed for high-sensitivity pathogen identification in low-titre infections, yet available methods using PCR or capture probes can be limited by high cost, narrow scope of detection, lengthy protocols and/or cross-contamination. Here, we developed metagenomic sequencing with spiked primer enrichment (MSSPE), a method for enriching targeted RNA viral sequences while simultaneously retaining metagenomic sensitivity for other pathogens. We evaluated MSSPE for 14 different viruses, yielding a median tenfold enrichment and mean 47% (±16%) increase in the breadth of genome coverage over mNGS alone. Virus detection using MSSPE arboviral or haemorrhagic fever viral panels was comparable in sensitivity to specific PCR, demonstrating 95% accuracy for the detection of Zika, Ebola, dengue, chikungunya and yellow fever viruses in plasma samples from infected patients. Notably, sequences from re-emerging and/or co-infecting viruses that have not been specifically targeted a priori, including Powassan and Usutu, were successfully enriched using MSSPE. MSSPE is simple, low cost, fast and deployable on either benchtop or portable nanopore sequencers, making this method directly applicable for diagnostic laboratory and field use.

Com o desenvolvimento dos mercados financeiros e de capitais, os bancos viram os seus negócios de crédito e depósitos começar a diminuir. Para se tornarem novamente competitivos e obterem uma maior satisfação por parte do cliente, viram-se na necessidade de iniciar e desenvolver novos negócios, nomeadamente o Negócio de Bancassurance. Esta é uma área que tem ganho interesse ao longo dos anos pelos diversos bancos e que continua a crescer. Assim, o presente relatório resulta de um estágio realizado no Banco Caixa Geral de Depósitos, no departamento de Negócio e Bancassurance, para a obtenção do grau de mestre em Matemática Aplicada à Economia e Gestão. Este trabalho consiste em estudar o perfil do Cliente que subscreve um seguro de saúde através do canal bancário, isto é, encontrar as principais características definidoras do perfil desse mesmo cliente, de modo a direcionar eventuais campanhas publicitárias e aumentar a probabilidade de subscrição do seguro de saúde. Para tal será aplicado um Modelo de Regressão Logística onde a variável resposta é a subscrição, ou não, de um seguro de saúde no contexto da atividade bancária, sendo assim possível aferir quais as variáveis que caracterizam um cliente deste tipo.

No summary/description provided

No summary/description provided

As Redes Neuronais Convolucionais (CNNs) são frequentemente utilizadas em medicina para executar diferentes tarefas. Um exemplo desta utilização é a classificação de imagens médicas. Assim, essa classificação pode ser designada de multi-classe ou binária. No primeiro caso, as CNNs conseguem prever, com elevada precisão, se um indivíduo apresenta ou não alguma das doenças pertencente ao conjunto para as quais o modelo foi treinado para reconhecer. No segundo caso, estes modelos permitem verificar e analisar se uma determinada imagem pertence a um indivíduo saudável ou se, por seu lado, são reveladoras de alguma patologia. Outra tarefa para a qual CNNs são frequentemente utilizadas é a segmentação de órgãos em imagens médicas. Estes modelos têm a capacidade de retirar informações complexas de dados pertencentes a milhões de pacientes, e utilizar esse conhecimento para tornar a medicina mais personalizada a cada indivíduo. No entanto, apesar do seu sucesso, estes modelos são caixas negras, extremamente complexos e difíceis de entender. Além disso, também não fornecem explicações ou justificações para as suas decisões. Isto é considerado um problema, especialmente em áreas de elevado risco como é o caso da medicina, principalmente se os modelos tiverem algum defeito inesperado, como por exemplo bias derivadas dos datasets utilizados para treinar os modelos. Numa tentativa de minimizar estes problemas, surgiu a área de estudo da inteligência artificial ‘explicável’ (XAI). Existem vários métodos de XAI os quais permitem obter explicações sobre as previsões realizadas por modelos de inteligência artificial. As ‘explicações’ fornecidas pelos métodos de XAI baseiam-se em características conhecidas dos modelos de inteligência artificial, tais como a propagação pelas diversas camadas do erro entre o esperado e o previsto. Existem vários estudos nesta área, no entanto, estes nem sempre se focam na análise quantitativa das explicações obtidas por métodos de XAI. A falta de investigação neste aspeto deve-se sobretudo à inexistência de uma ground truth que possa ser utilizada para comparar o obtido nas explicações com o que se espera obter por uma explicação ‘correta’. O objetivo deste trabalho passa, assim, por criar uma framework para quantificar a qualidade das explicações obtidas por métodos de XAI e utilizar esta framework para investigar como a transferência de conhecimento (TL) influencia as explicações. A TL consiste em treinar, inicialmente, um modelo numa tarefa para a qual existe uma grande quantidade de dados de treino etiquetados e após o treino utilizar os pesos aprendidos neste treino para classificar um novo problema (para o qual tipicamente se tem menos dados com etiquetas). A vantagem desta técnica é que, após o treino inicial, apenas é necessário ajustar os pesos do modelo para a nova tarefa (fine-tuning do modelo), sendo necessário uma menor quantidade de dados e potencialmente treinar apenas algumas das suas camadas. Por este motivo, a TL é muito utilizada em medicina, área onde as bases de dados existentes têm tendência a ser de reduzida dimensão e a obtenção de novos dados etiquetados difícil, demorada e dispendiosa. O treino inicial destes modelos é tipicamente feito recorrendo a imagens naturais para as quais existem enormes bases de dados, como a ImageNet. No entanto, existe na literatura a discussão sobre se o treino inicial deve ser feito com imagens naturais ou com imagens mais próximas do problema para o qual este modelo será utilizado, ou seja, imagens médicas. O principal argumento desta discussão sugere que o segundo caso é mais benéfico, uma vez que as formas existentes em imagens médicas são muito diferentes das formas que existem nas imagens naturais. Devido a esta diferença, a informação aprendida na primeira tarefa, pode não ser muito relevante para a segunda. A framework criada neste estudo permite gerar uma base de dados de imagens médicas sintéticas para as quais se conhece a ground truth, possibilitando, desta forma, a avaliação quantitativa das explicações que se obtêm com os métodos de XAI. Esta base de dados é composta por diversas imagens axiais de ressonância magnética (MRI) do cérebro às quais se adicionam lesões sintéticas hiperintensas. A quantidade de imagens criadas é variável e, a cada imagem, um número aleatório de lesões (entre 3 e 5) é adicionado em locais aleatórios. As lesões adicionadas à MRI são também criadas com formas aleatórias. O contraste entre as lesões e o cérebro pode também ser ajustada. As imagens do cérebro são obtidas do Human Connectome Program (HCP) e pertencem a indivíduos saudáveis entre os 22 e os 37 anos. A ground truth corresponde às lesões que são adicionadas à imagem de MRI do cérebro. Estas lesões são de dois tipos (regulares ou irregulares) e a tarefa para a qual os modelos em estudo são treinados consiste em distinguir entre estes dois tipos de lesão. As imagens do cérebro da base de dados criadas foram utilizadas para ajustar camada a camada CNNs previamente treinadas para outros problemas. Um destes modelos iniciais foi treinado com várias imagens naturais de 1000 categorias diferentes, pertencentes à base de dados ImageNet. O outro modelo foi treinado com MRIs axiais de cérebros sem lesões com o objetivo de as distinguir entre género feminino e masculino. Ambos os modelos utilizados neste trabalho têm a arquitetura VGG-16, sendo compostos por 16 camadas treináveis, 13 das quais convolucionais e as restantes fully connected. Destas CNNs pré-treinadas foram obtidos diversos modelos com vários graus de fine-tuning (de apenas 1 conjunto de camadas até à totalidade da arquitetura). Cada grau de fine-tuning corresponde às camadas existentes entre duas camadas de max-pooling consecutivas. As explicações foram obtidas por oito métodos de XAI. Após a obtenção das explicações, a ground truth foi utilizada para se calcular a qualidade da explicação, ou seja, qual a percentagem de pixeis mais relevantes da explicação que se encontram na ground truth. Os resultados experimentais deste trabalho, demonstram que existe uma grande dispersão de valores relativos à qualidade das explicações, havendo algumas com uma qualidade muito elevada e outras com uma qualidade mínima. Este comportamento observa-se para todos os métodos de XAI avaliados. Isto acontece mesmo quando apenas se consideram os exemplos corretamente classificados por todos os modelos de inteligência artificial em estudo. Este é o melhor cenário, uma vez que no mundo real, nunca se sabe se o modelo utilizado classificou o exemplo corretamente ou não, pelo que esta dispersão apenas pode aumentar. Destes resultados, observa-se também que a qualidade do modelo e a qualidade da explicação estão correlacionadas. Isto faz sentido uma vez que quanto mais exemplos o modelo conseguir classificar corretamente, mais significativas se espera que sejam as informações aprendidas por este. No entanto, quando apenas se avaliam modelos com precisão semelhante, verifica-se que as explicações dos modelos treinados inicialmente com imagens no mesmo domínio do problema (MRI do cérebro) têm uma melhor qualidade. Verifica-se, ainda assim que, em geral, a qualidade das explicações tende a aumentar com o aumento do grau de fine-tuning do modelo. Este comportamento é mais facilmente verificado para certos métodos de XAI como é o caso de Deconvolution e Guided Backpropagation. Também é possível verificar nos resultados obtidos que modelos inicialmente treinados com MRI de cérebros têm uma qualidade de explicação mais constante dos que os treinados inicialmente com imagens da base de dados ImageNet. No segundo caso, a qualidade das explicações diminui rapidamente com pequenas reduções da qualidade do modelo, o que não é tão visível no primeiro caso. Os resultados obtidos neste trabalho vão de encontro à linha de pensamento de que é mais benéfico utilizar modelos pré-treinados em problemas semelhantes àquele a que o modelo será aplicado. Este é o caso pelo menos para imagens médicas, em particular MRIs do cérebro, que foram o foco deste trabalho. Isto parece ser verdade mesmo quando os modelos obtidos apresentam menor precisão, já que as explicações parecem fornecer informação mais significativa do que quando as imagens utilizadas para treinar inicialmente o modelo pertencem a áreas mais distantes como, neste caso, imagens naturais. Além deste aspeto, os resultados sugerem que é importante continuar a investigar objetivamente e avaliar quantitativamente os métodos de XAI existentes e futuramente criados. Novos estudos que considerem esta avaliação recorrendo a ground truths parecem assim ser cruciais para que seja possível criar métodos melhores e mais robustos. Além disso, estes estudos permitirão ainda avaliar as implicações que a utilização dos métodos de XAI atualmente existentes tem em áreas de alto risco, como é o caso da medicina.

Antibiotics are a very important class of drugs in modern Medicine; its discovery and introduction constituted a major revolution in Medicine. As such, the decrease of their effectiveness due to the rising levels of antibiotic resistance is a great concern to our society. Hospital-Acquired Infections (HAI), also known as nosocomial infections, are infections that a patient acquires in the context of medical treatments. Nosocomial infections are closely related to the problem of resistance to antibiotics, because a large part of these infections are caused by antibiotic resistant bacteria. This aspect motivates the tackling of these two problems jointly. This work was performed in the frame of the RESISTIR project, that aims to develop a Decision Support System for intelligent control of infection and personalized antibiotherapy. The data was collected by a portuguese software company, Maxdata Software, S.A.; then, in the scope of this project, it was pre-processed and inserted into a database. The database contains information generated by clinical episodes with origin in three portuguese hospitals, dated from 2013 to 2016. The available data allowed us to develop predictive models of risk of antibiotic resistance (AMR models) and risk of nosocomial infection (HAI model). For the AMR models, we aggregated the antibiotics into the level 4 of the ATC classification system, and focused on four classes of antibiotics: J01MA (fluoroquinolones), J01CA (penicillins with extended spectrum), J01DC (second-generation cephalosporins) and J01DH (carbapenems). In all predictive models developed, the features generated related to the clinical history of the patients and the health units involved in the episodes were found to be the most significant. We concluded that the main goals of this project were achieved, with the development of prototype predictive models of risk of antibiotic resistance and risk of nosocomial infection. The predictive power of the models, as measured by the ROC-AUC, ranged from 0.720 to 0.857 for the AMR models, and was 0.915 for the HAI model. These prototype predictive models that we were able to develop show that it is possible to deploy in healthcare units a Decision Support System that can help to monitor and reduce the problem of resistance to antibiotics and the nosocomial infections. For the HAI predictive model developed, we made an estimate of its benefits if it would be deployed in production in the three hospitals involved in this project. We drew the conclusion that the following savings would be obtained, in 2007 euros and considering only the hospitalizations: about 20 million euros and a reduction of approximately 60 000 days in the hospital stays, per year.

Particles moving in a fluid interact though flows leading to rich non-linear behaviours. A problem of interest is the synchronisation in biological flows at the microscale, such as the synchronisation of bacteria flagella and the transport of organelles in living organisms by cilia. In this work, we used the lattice Boltzmann method to simulate moving solid particles in a fluid and simulate the fluid-particles interaction. With these simulations, we studied the effect on synchronization of different parameters, such as the fluid viscosity, the distance between oscillators, their natural oscillation frequency, and the initial phase shift. This study is divided in to two parts: systems with particles with the same natural oscillation frequency and systems with particles of different natural oscillation frequencies. This was done with further propose of relating these quantities with the coupling factor of the Kuramoto model. As the simplest Kuramoto model used presented some limitations, we propose a modification to it, where either the natural frequency of the coupling factor is non-constant. With these modified models, we concluded that the necessary coupling factor to synchronize is proportional to the natural frequency and the fluid viscosity, while inversely proportional to the distance between centers of oscillation.

As hemoglobinopatias referem-se a patologias associadas a alterações moleculares nos genes das globinas e podem ser classificadas como variantes da hemoglobina ou talassémias. Os objetivos deste estudo compreenderam a caracterização dos fenótipos hematológicos e bioquímicos típicos dos portadores de β-talassémia e, em casos atípicos, pretendeu-se investigar as suas bases moleculares e a contribuição de fatores genéticos modificadores. O estudo inicial incluiu 133 indivíduos onde a presença de alterações β-talassémicas comuns, em heterozigotia, foi confirmada molecularmente e a coexistência de anemia ferropénica foi excluída. Os seus parâmetros hematológicos e bioquímicos foram tratados estatisticamente e o perfil típico do portador dessa hemoglobinopatia foi estabelecido como possuindo microcitose, hipocromia e HbA2 >3,5%, podendo ser acompanhado de anemia e eritrocitose. Esse conhecimento permitiu-nos selecionar 18 casos suspeitos de hemoglobinopatias relacionadas com as cadeias da β-globina mas apresentando fenótipos atípicos. O estudo desses indivíduos incluiu a análise dos genes HBG2, HBG1, HBD e HBB, e suas regiões regulatórias, e foi realizado através de metodologias de ARMS, ARMS-Multiplex, PCR, Sequenciação de Sanger, MLPA e Gap-PCR. Como resultado foram identificadas as alterações β-talassémicas Cd8 (-AA), IVS-I-1 (G>A), IVS-I-6 (T>C) e IVS-II-1 (G>A), e as variantes de hemoglobina de cadeia β: Hb S, Hb Malmö e Hb O-Arab. Além disso, as variantes de hemoglobina de cadeia δ (HbA2' e Hb A2-Yialousa), a deleção Corfu e uma deleção que remove todo o agrupamento génico da β-globina foram identificadas como justificativas para valores anormalmente baixos de HbA2. Níveis atipicamente elevados de HbF foram explicados pela presença dos polimorfismos -158 (C>T) em HBG2, -195 (C>G) e -179 (T>C) em HBG1, e pelas deleções HPFH-1 e HPFH-2. Concluímos que o conhecimento da base molecular complexa subjacente a casos de hemoglobinopatias de apresentação atípica contribui para o melhor conhecimento da fisiopatologia da doença e de fatores modificadores, e poderá revelar novos alvos terapêuticos.