APRÈN. Portal de la producció docent del professorat de la UPC

The aim of this Final Degree Project (TFG) is to develop a website for the diagnosis of rare diseases. This tool is intended for non-specialized healthcare professionals, such as general practitioners and pediatricians. The goal is that by simply inputting the patient's symptoms, a quick and intuitive diagnosis can be obtained, thus facilitating the referral of the patient to the appropriate specialist for further diagnosis and treatment. The project primarily focuses on creating a user interface that offers a good user experience (UX). It should not only be visually appealing but also easy to use and fast. The tool is meant to serve as a complement to the medical professional, ensuring it does not take more time than the clinician would normally spend making a prediction without the web assistant. The symptoms grouping and their connection to rare diseases has been carried out using HPO (Human Phenotype Ontology). The HPO is a system for organizing and classifying medical information related to human phenotypes, helping link symptoms to the genetic conditions that cause them. Finally, to perform the prediction, an API developed by b2slab-CREB member Rafael León has been used. This API, currently optimized for diagnosing rare kidney diseases, has the potential to be extrapolated to other rare diseases. This tool, called the Minority Disease Diagnostic Assistant (Midia), has the potential to sig- nificantly improve the early diagnosis of rare diseases, which are often difficult to detect by non-specialist professionals. It enables quicker and more accurate identification, optimizing the referral process to the appropriate specialists. This reduces diagnostic delays, leading to a faster start of specific treatments, which is key to avoiding complications and increasing treat- ment effectiveness. In this way, it contributes to more efficient and personalized medical care.

Posterior Agreement (PA) has been proposed as a theoretically-grounded alternative for model robustness assessment in covariate shift settings. In this work, we provide further evidence in favor of these findings, and we explore the use of PA as a model selection criterion for deep learning models in supervised image classification tasks. Starting from the theoretical principles leading to the formulation of PA, we derive an operative version for discrete hypothesis set problems and use it as a robustness metric in the presence of adversarially-perturbed and domain-shifted data. We show that PA outperforms standard accuracy-based metrics in both settings and possesses superior discriminative power and consistency across increasing levels of covariate shift.

Aquest treball estudia l'evolució d'una sèrie temporal dels anys 2019, 2020 i 2021 que recull múltiples dades relacionades amb el camp de la salut d'aproximadament 1.5 milions de persones majors d'edat a Catalunya. L'objectiu és crear un un model d'Intel·ligència Artificial capaç de predir els esdeveniments que han succeït l'any 2021 per a cada pacient, a partir de l'entrenament realitzat amb les dades dels anys 2019 i 2020. Tot i que la predicció cobreix una àmplia gamma de variables relacionades amb la diabetis, com ara hipertensió o malalties cardiovasculars, el focus principal és: predir l'ús de medicaments relacionats amb la regulació de la glucosa en sang, com la insulina, predir si el resultat de l'anàlisi de l'hemoglobina glicada cau dins dels paràmetres normals (menys del 7%) o és més alt, i predir la probabilitat que un pacient sigui diagnosticat amb diabetis tipus 2. El model s'ha desenvolupat a partir d'un model utilitzat en la generació de dades sintètiques i després s'ha adaptat a les necessitats específiques del conjunt de dades utilitzat. Donada la intrínseca dificultat del projecte, s'ha prioritzat l'obtenció de resultats metodològics sobre la consecució de resultats conclusius ja que el model no ha assolit la capacitat predictiva esperada.

The Covid-19 pandemic has posed unprecedented challenges to healthcare systems worldwide, particularly impacting patients with Type 2 Diabetes Mellitus (DM2), Obesity and Renal Chronic Microvascular Disease (RCMD). This study investigates the bidirectional relationship between Covid-19 and metabolic control in patients with DM2, Obesity or RCMD, aiming to understand the evolution of metabolic parameters during the pandemic. Leveraging advanced statistical techniques and supervised learning models, the study identifies distinct patient clusters based on metabolic profiles and predicts post-Covid-19 metabolic evolution. Data from patients' metabolic parameters before and after Covid-19 episodes are collected and analyzed, revealing three clusters: the renal group, lipid group, and HbA1c group, each with unique metabolic characteristics and clinical implications. Supervised learning models demonstrate moderate success in predicting cluster membership, highlighting the complexity of post-Covid-19 metabolic evolution prediction. While significant predictors include age, sex, initial metabolic values and comorbidities, discrepancies between medical interpretations and model results underscore the need for further research and validation. This study lays the groundwork for personalized care strategies for post-Covid-19 patients with DM2, Obesity or RCMD, contributing to the advancement of predictive medicine and targeted interventions in the face of complex metabolic conditions exacerbated by the pandemic.

The RANO criteria are the current standard for monitoring patients with brain tumours as they allow assessment of treatment progress by obtaining the maximum perpendicular diameters in the axial plane. However, these criteria present some inter- and intra-operator ambiguity when drawing conclusions, which can lead to errors. Making use of new tools for automatic tumour segmentation using ML and AI, the aim is to develop a tool capable of obtaining RANO metrics in an automated way in order to obtain accurate results that provide objectivity, standardisation and reproducibility. Based on data extracted from the public repository of The University of California San Francisco Preoperative Diffuse Glioma MRI and the HD-GLIO automatic segmentation tool, an algorithm has been developed capable of obtaining the maximum diameter in the axial plane, as well as its value in units of the decimal metric system and the visualisation of the results in MR images.

When it comes to motorsports, the racing line is a crucial factor in lap time. The goal of this project is to determine the trajectory to be followed along a circuit and the velocity at which to travel it for an autonomous Formula Student vehicle. To do so, two algorithms with different characteristics based on the available computing time are presented, which improve the lap time that would be obtained by tracing the center of the track. When designing racing lines, it is common to use geometric constructs such as curvature or length, although the definition of the optimal path comes down to only one other magnitude: time. This project proposes directly optimizing this metric and demonstrates that it is indeed possible to do so using evolutionary algorithms, obtaining significant improvements over purely geometric approaches. In addition, a tool for random circuit generation is used to statistically evaluate the algorithms presented as well as other similar articles. This comparison is not only carried out theoretically but also uses a vehicle simulation to verify the feasibility of the obtained trajectories. The results show that the proposed technique has advantages in achieving the desired behavior as well as a better lap time, that is, a better racing line.

Metabolomics, which is defined as the comprehensive analysis of metabolites in a biological system, has lots of applications that are leading to an important step forward in medicine. Since it is still an emerging field, researchers are trying to create computational tools that improve the processing, analysis and interpretation of metabolomics data. In this project, we have developed a new library in R based on MAIT (Metabolite Automatic Identification Toolkit) R package. Our algorithm provides a set of state-of-the-art tools to perform the whole metabolomics workflow, from raw liquid-chromatography coupled to mass spectrometry (LC-MS) data to the biological interpretation.

This thesis is focused on the analysis of the results of the paper KÖHLER, Sebastian, et al. Clinical diagnostics in human genetics with semantic similarity searches in ontologies(1) and the improvement of them. The method studied uses symptoms from the Human Phenotype Ontology 2022 (HPO) (2) and tries to predict the disease using another database from the same website (3). This comparison is performed using semantic similarity methods that will be explained in the next chapters. While trying to reproduce the results, many difficulties were found to reach that goal. After many auditions of the code that run the simulation, the performance was not close to that shown in the paper. However, an implementation of this method is already available online by the own authors of the article. This is called the Phenomizer. After generating hundreds of simulated patients following the rules specified in the paper, and using this data in the Phenomizer, one conclusion can be extracted from it, the results are not consistent with the data specified in the article. When imprecision is added to the symptoms (randomly replacing a symptom by one of their ancestors), performance is much worse than stated. On the other hand, adding noise (symptoms unrelated to the disease) has little effect on the output and the performance is coherent with the paper. To try to improve the performance, some other semantic similarity methods were tested using a specialized library called Semantic Measures Library and Toolkit (3). No significant improvements were detected in this investigation, the method used in the article was the best in terms of results quality. The last important point that is clashing with the article is the fact that they use a p-value calculated from the semantic similarity score to obtain the rankings of more probable diseases for the given set of symptoms. According to the paper, the p-value method is better than using simply the similarity scores. However, in all the tests done in this thesis similarity scores performed better than p-values. This means that no big calculations, expensive in time, and a big hard drive capacity are required. In conclusion, the results of the real Phenomizer are worse than stated; no better semantic similarity method than the existent one was found, and disease rankings by similarity scores performed better than by p-value, contrary to the paper assertions.

Over recent years hate speech has rapidly spread over the internet, concretely this work centered on Hateful Memes dataset, a problem that requires both visual and linguistic understanding of the meme to classify it correctly. In order to detect hate speech easily, AI can help. This work has presented a method that can help with it. The approach consists of an architecture that starts in-painting the memes from text, then extracting relevant features (including race), and finally training two models based on transformers (BERT and ERNIE-VIl) in order to ensemble it and get the hatefulness of the meme. At the end of the work, the results are analyzed in order to help the community that deals with similar multimodal problems

This master thesis has two aims. The first main objective is to study and obtain robust voxel-wise (3D) radiomic features for lung cancer and with them compute the habitats (unsupervised clusters) to unravel tumor heterogeneity and facilitate a personalized treatment for each patient and to see the progression of the disease over time.The goal is to obtain robust radiomic features that can be generalized to every database with lung cancer data. The second main objective is to develop a CT-radiomic signature (i.e., imaging phenotype) that will predict the tumor immune profile obtained from the RNA analysis of biopsied sample and allow to identify patients with lung cancer who will respond to immunotherapy.

Fer una recerca de com es pot utilitzar el Big Data per tal de millorar el rendiment d'un equip en diferents esports.

This work presents a dual approach to object recognition: 2D computer vision on images on the one hand, and 3D point cloud processing on the other. The recognition of objects has been applied to a synthetic “supermarket” shelf (called testbench) with household items, carefully chosen so that they have simple geometry and texture information for the algorithms to use. Specifically, the objects to detect are Pepsi and Coca-Cola cans, Milk bricks and a detergent bottle. After tackling both approaches a combined 2D and 3D approach is performed. Regarding the 2D object detection part this work takes a very pragmatic approach. First it tackles the issue of the tooling behind creating and curating image datasets, which is an overlooked aspect in many works. It does so by creating a Graphical User Interface with advanced functionality. Then, it starts off the with the simplest detection methods, like Template Matching, all the way to Convolutional Neural Networks, assessing the results of each and their tradeoffs. Regarding the 3D object recognition many vital aspects of point cloud processing are treated first, mainly filtering and compositions of clouds. Then, object recognition itself is addressed, and its limitations are explained. Finally, the 2D and 3D approaches are combined in a modular manner that is quite uncommon in the literature, so as to leverage the strengths and weaknesses of each method

El Centro de Investigación en Ingeniería Biomédica (CREB) de la Universidad Politécnica de Cataluña (UPC) ha creado una red de I+D+i en tecnologías de la salud denominado XarTEC Salut. Entre sus iniciativas, destaca la idea de crear una incubadora común desde Barcelona de tecnologías innovadoras del sector de la salud creadas en países mediterráneos. En este trabajo, se seleccionan 5 países del arco mediterráneo que no pertenecen a Europa con el objetivo de estudiar una posible vía expansión de la incubadora. Estos países seleccionados pertenecen a la región MENA (Middle East and North Africa), y son los siguientes: Argelia, Líbano, Túnez, Egipto y Marruecos El Centro de Investigación en Ingeniería Biomédica (CREB) de la Universidad Politécnica de Cataluña (UPC) ha creado una red de I+D+i en tecnologías de la salud denominado XarTEC Salut. Entre sus iniciativas, destaca la idea de crear una incubadora común desde Barcelona de tecnologías innovadoras del sector de la salud creadas en países mediterráneos. En este trabajo, se seleccionan 5 países del arco mediterráneo que no pertenecen a Europa con el objetivo de estudiar una posible vía expansión de la incubadora. Estos países seleccionados pertenecen a la región MENA (Middle East and North Africa), y son los siguientes: Argelia, Líbano, Túnez, Egipto y Marruecos Primero de todo, se realiza un estudio general económico y social de los países seleccionados. Para ello, se presentan sus principales indicadores macroeconómicos, su clasificación según el Índice de Democracia Global, su clasificación según el Índice de Desarrollo Humano y su clasificación según el Índice de Competitividad Global. Además, se muestran algunos rankings de interés científico, económico y social, que pueden ayudar al lector a acabar de entender el panorama actual en dichos países. Posteriormente, se realiza un mapeo exhaustivo de las principales universidades, hospitales y centros de investigación de cada país. Unas largas listas de instituciones y organizaciones públicas y privadas que ayudan a situar al talento dentro de un ecosistema de ciencia, tecnología e innovación en plena fase de crecimiento y desarrollo. Después se realiza un estudio del arte de la transferencia de tecnología para cada país de forma individual. Por ese motivo, se muestra el papel que juegan los gobiernos, las universidades, los centros de transferencia de tecnología, los centros de investigación, las entidades financieras y una larga lista de entidades que intentan que las ideas más innovadoras creadas en cada país puedan llegar a ser comercializadas por el bien de todos. Por último, se aportan recomendaciones para la expansión de la incubadora común del sector Health Tech por los 5 países, describiendo los principales atractivos diferenciales de la incubadora y aportando ideas para captar startups, así como requisitos que estas deben tener para entrar en el programa de incubación. También se realiza un estudio de las etapas que la incubadora debe tener para aprovechar al máximo sus puntos fuertes, entre los que destaca la disponibilidad de financiación a través de los fondos europeos Next Generation y su especialización en un nicho de mercado como es el sector de las tecnologías de la salud, todo ello desde Barcelona

Coronary Artery Bypass Grafting is the most common surgery for treating Coronary Artery Disease, and one of the most common surgeries worldwide. Despite having a low mortality rate, CABG has a relatively high complication rate. Medical protocols used by physicians to drive the recovery of these patients are highly standardized, yet the mechanisms that lead to different outcomes are still not well understood. By gaining deeperinsight into the physiologic relationships that drive a patient's recovery from surgery,healthcare providers could better adjust their treatment, potentially improving recoveryand reducing complications. The main goal of this master's thesis is to contribute to finding these relations by looking at trends in the evolution of vital signs and medication doses. This master thesis has been carried on at the Cardiovascular Research Center of the Massachusetts General Hospital and data used for the analysis come from this hospital. The methodology developed in the project consisted of three main steps. First, a tool was created to extract, curate and merge data from two databases of the hospital. The tool, named Tensorizer, can also be used to obtain data for other analyses. Next, a web visualization application was designed and implemented to allow visual examination of signals. This tool was also used to assist in the detection of corrupted data and its use is also extensible to other studies beyond this project. Finally, different clustering methods were applied on a set of 10 vital signs (heart rate; diastolic, systolic and mean pulmonary artery pressure; diastolic, systolic and mean arterial pressure; blood temperature, oxygen saturation and central venous pressure) and 5 medication dose trajectories (epinephrine, norepinephrine, vasopressins, milrinone and phenylephrine) recorded during the first after CABG. These methods include multivariate time-series clustering on the entire stream of vital signs and univariate time-series clustering on individual vital signs and medications.Results from these methods were combined into a table (Tabular Representation) and analyzed together.Clustering results suggest that vital sign and medication trajectories contain predictive power on their trajectories and reveal repeated patterns within and between signals. These patterns can be used to predict bad recovery from surgery, identify patients at risk or assist physicians to provide more personalized treatments. Results motivate and open the doors for future research in the field

The study of the health effects of sport has been studied over the years from many perspectives. Microarray technology allows studies to be carried out from the perspective of gene expression. However, existing studies show a large number of genes and consequently altered biological pathways after intense exercise. This large number of genes makes it difficult to interpret the results. At the same time, it has been observed that after an intense exercise session significant variations in the levels of most blood cells occur. Taking these two factors into account, the following question arise. Is there a relationship between variation in gene expression and variation in blood cell counts? The main objective of this study is to explain the variation in total blood differential expression in response to running intervention by while controlling the variation of the values of the blood cell count. For this purpose, genetic expression data (HuGene2.0st microarrays), complete blood count data and other biological variables (sex, age, performance) are collected from runners before and immediately after the competition “Volta a la Cerdanya 2013”. Runners are stratified by three distance categories; A) 14 km (N=30, Age= 35.07 ± 9.76), B) 35 km (N=40, Age=35.91 ± 7.65) and C) 55 km (N=24, Age=36.58 ± 4.64). A linear regression model was fitted to each TC variation in expression data. The level count of erythrocytes, neutrophils, lymphocytes, monocytes, eosinophils and platelets have been used as explanatory variables. Finally, gene enrichment analysis has been computed over the Gene Ontology (GO) and the Kyoto Encyclopedia of Genes and Genomes (KEGG) Pathway. 91 genes are reported as statistically significant (independent of the cell counts), 774 related to neutrophils, 11 related to lymphocytes, 69 related to monocytes, and 28 related to eosinophils. 138 GO therms and 16 KEGG pathways are reported as statistically significant (independent of the cell counts), 257 GO therms and 16 KEGG pathways related to neutrophils, 22 GO therms related to lymphocytes, 154 GO therms and 12 KEGG pathways related to monocytes, and 1 GO therm related to eosinophils.

The aim of this project is to analyse and parametrize the evolution of the gait in muscular dystrophies such as Duchenne, Becker or Myotonic, and to discriminate them from the control group. The data has been collected with the WalkTest+ System during a 6-MinuteWalk-Test (6MWT). The studied data belongs to 69 tests of different subjects and the variables recorded can be classified into the following categories: spatiotemporal, cardiorespiratory, electromyographic, plantar pressure and kinematics. The groups existent in the data as well as the groups in the predictors have been analysed. It has been determined that the data structure is coincident to the separation between the control group and the group with Duchene muscular dystrophy. Certain parameters collected through the system are able to differentiate significantly the patients with pathologies from the control groups. It has been concluded possible a reduction in the performing test time while maintaining the necessary information to distinguish the groups.

L'objectiu del treball és la implementació d'un aplicatiu web capaç de facilitar l'accés al públic general d'un mètode de traducció capaç de relacionar descripcions simptomàtiques amb termes corrents amb una ontologia clínica de referència. S'expliquen els fonaments teòrics que fan possible la traducció automàtica i en concret, es descriuen els mètodes emprats S'estudien les eines existents per a integrar aplicatius webs capaços d'efectuar prediccions de models d'aprenentatge profund ja entrenats amb llibreries com Keras, Per mitjà del desenvolupament de diversos models de processat de llenguatge natural i a fi de poder emetre prediccions en el navegador web amb la llibreria TensorFlow.js. Es desenvolupa un estudi sobre les solucions actuals per a la representació gràfica d'informació complexa. En un primer terme s'opta per llibreries desenvolupades amb Python. A continuació i com a conseqüència de l'estudi de llibreries especialitzades en la representació d'ontologies, es desenvolupen solucions que resulten especialment útils per a l' entorn web gràcies a la llibreria D3.js. S'arriba a la conclusió que una proposta d'aplicatiu desenvolupada amb Node.js i Express permetrà una millor solució a la problemàtica. Assolint-se un aplicatiu basada tant en la banda del servidor com en la del client capaç de desenvolupar la tasca esmentada així com de transmetre informació complexa sobre estructures ontològiques de manera gràfica, intuïtiva i segons els interessos de l'usuari.