Search

Traditional psychiatric hospitalization models have been criticized for being restrictive and for violating fundamental rights, which can be counterproductive for long-term patient recovery. In response to this situation, new hospitalization models have been developed that aim to create a safe therapeutic environment that promotes patient recovery, while minimizing the need for coercive measures. The psychiatry unit from Germans Trias i Pujol Hospital (HGTiP) has designed a dual model that combines a care method, called Safewards, with a risk prediction tool in emergencies and a risk categorization and workload tool on the ward through a global score. The objective of the model is to safely open the unit doors, taking into account the risks and the level of work of the care team. The tools currently used are not scientifically validated, thus hindering the scalability of the model. Therefore, the purpose of this project is to develop a proposal for intervention to develop two Artificial Intelligence (AI) algorithms, using validated variables, that predict the unit environment and patient risk at the time of admission. First, a qualitative study was conducted with expert personnel from the psychiatric unit at HGTiP to objectify the criteria by which risk level and workload on the ward are defined. A total of 35 items were extracted, and a digital platform was developed to store them. These will be used, along with other dynamic variables, as input into a regression algorithm. In addition, classification algorithms will be developed to predict the level of risk in the first 24 hours of admission. The input data will have variables related to symptoms and conflictive events in the last month. For both cases, the performance of different algorithms will be evaluated, and the one with the best predictive character will be chosen. This dual model, which uses Artificial Intelligence as a tool to support clinical decision-making, sets a precedent in risk control in open door models. Once the model is fully validated, it will be scalable to other units and will allow for an improvement in the quality of care, while maintaining the safety of all individuals involved in the recovery process Los modelos de hospitalización psiquiátrica tradicionales han sido criticados por ser restrictivos y vulnerar derechos fundamentales, lo que puede ser contraproducente para la recuperación de los pacientes a largo plazo. En respuesta a esta situación, se han desarrollado nuevos modelos de hospitalización que pretenden crear un ambiente terapéutico seguro y que promueva la recuperación del paciente, minimizando la necesidad de medidas coercitivas. El Hospital Germans Trias i Pujol, ha diseñado un modelo dual, que combina un método asistencial, llamado Safewards, con una herramienta de predicción de riesgos en urgencias y una categorización de riesgos y de carga asistencial en planta a través de una puntuación global. El objetivo del modelo es abrir las puertas de la unidad de manera segura, teniendo en cuenta los riesgos y el nivel de trabajo del equipo asistencial. Las herramientas que se usan actualmente no están validadas científicamente, dificultando así la escalabilidad del modelo. Por lo que la finalidad de este proyecto es elaborar una propuesta de intervención para desarrollar dos algoritmos de Inteligencia Artificial, con variables validadas, que predigan el ambiente de la unidad y el riesgo de un paciente en el momento de ingreso. En primer lugar, se ha realizado un estudio cualitativo con personal experto de la unidad de psiquiatría del HGTiP para objetivar los criterios por los que se define el nivel de riesgo y la carga asistencial en planta. Se han extraído un total de 35 ítems y se ha desarrollado una plataforma digital para almacenarlos. Estos serán utilizados, junto a otras variables dinámicas, como entrada en un algoritmo de regresión. Por otra parte, se desarrollarán algoritmos de clasificación para predecir el nivel de riesgo en las primeras 24 horas de ingreso. Los datos de entrada serán variables relacionadas con sintomatología y eventos conflictivos en el último mes. Para ambos casos, se evaluará el rendimiento de diferentes algoritmos y se elegirá el que mejor carácter predictivo tenga. Este modelo dual, que utiliza la Inteligencia Artificial como herramienta de apoyo a la decisión clínica, supone un precedente en el control de riesgos en modelos de puertas abiertas. Una vez validado el modelo en su totalidad, sería escalable a otras unidades y permitiría mejorar la calidad asistencial, manteniendo la seguridad de todas las personas implicadas en el proceso de recuperación Els models d'hospitalització psiquiàtrica tradicionals han estat criticats per ser restrictius i vulnerar drets fonamentals, cosa que pot ser contraproduent per a la recuperació dels pacients a llarg termini. En resposta a aquesta situació, s'han desenvolupat nous models d'hospitalització que pretenen crear un ambient terapèutic segur i que promogui la recuperació del pacient, minimitzant la necessitat de mesures coercitives. L'Hospital Germans Trias i Pujol, ha dissenyat un model dual, que combina un mètode assistencial, anomenat Safewards, amb una eina de predicció de riscos a urgències i una categorització de riscos i de càrrega assistencial a planta a través d'una puntuació global. L'objectiu del model és obrir les portes de la unitat de manera segura, tenint en compte els riscos i el nivell de treball de l'equip assistencial. Les eines que es fan servir actualment no estan validades científicament, dificultant així l'escalabilitat del model. La finalitat d'aquest projecte és elaborar una proposta d'intervenció per desenvolupar dos algorismes d'Intel·ligència Artificial, amb variables validades, que prediguin l'ambient de la unitat i el risc d'un pacient a l'ingrés. En primer lloc, s'ha fet un estudi qualitatiu amb personal expert de la unitat de psiquiatria de l'HGTiP per objectivar els criteris pels quals es defineix el nivell de risc i la càrrega assistencial a planta. Se n'han extret un total de 35 ítems i s'ha desenvolupat una plataforma digital per emmagatzemar-los. Aquests seran utilitzats, juntament amb altres variables dinàmiques, com a entrada en un algorisme de regressió. D'altra banda, es desenvoluparan algorismes de classificació per predir el nivell de risc en les primeres 24 hores d’ingrés. Les dades d'entrada seran variables relacionades amb simptomatologia i esdeveniments conflictius a l'últim mes. Per a tots dos casos, s'avaluarà el rendiment de diferents algorismes i es triarà el que millor caràcter predictiu tingui. Aquest model dual, que utilitza la Intel·ligència Artificial com a eina de suport a la decisió clínica, suposa un precedent en el control de riscos en models de portes obertes. Una vegada validat el model en la seva totalitat, seria escalable a altres unitats i permetria millorar la qualitat assistencial, mantenint la seguretat de totes les persones implicades en el procés de recuperació

Technological and medical advances in recent decades have promoted a lengthening of life expectancy, leading to an increase in age-associated cognitive diseases. As a result, there is a growing interest in efficient and cost-effective technological alternatives for the early detection of cognitive problems. Recent studies show growing evidence of a relationship between electrocardiogram (ECG) parameters and cognitive decline. For this reason, this study aims to assess the validity of ECG markers to detect mild cognitive impairment in elderly subjects. The 'Lemon' database, that contains the cognitive and ECG results of 227 healthy adults, has been analysed. Initially, the electrocardiographic signals have been filtered and adapted to extract multiple relevant ECG markers. To assess the reliability of these markers, several tests were conducted. A statistical analysis between cognitive categories, a machine learning classification problem, and multiple regressions with the additional tests included in the database. The analysis focused on elderly patients, who are classified in three cognitive groups according to their impairment (normal, moderate, and severe). Given the statistical significance between cognitive categories and the favourable results obtained in the regression analysis, two markers should be emphasised: the traditional temporal ratios of heart rate variability (HRV), such as the average beat, a measure of the average between QRS peaks, and the fragmentation ratios (HRF), an estimate of cardiac neuroautonomic dysfunction measured by variations in the QRS interval. Among the different classifiers, the artificial neural network (ANN) stands out, providing the best and most stable classification ratios (55.00 % accuracy, 47.60 % sensitivity, 50.00 % precision, in a three class, five-cross-fold problem). The reported tendency to correctly differentiate normal subjects must be emphasized. This is observed both in the statistical tests, where no significance is found between the moderate and severe groups, and in the classification problems, since there is a bias towards error among subjects with impaired condition. In conclusion, a slight relationship between cardiac markers and cognitive decline has been reported. Further research as well as the replication of the conducted study with different databases is recommended. The idea of combining ECG markers with other literary validated technologies, such as electrocardiography (EEG), suggests an interesting approach that may lead to improved detection of cognitive disorders in early stages

The quality of life of people with epilepsy is considerably altered by the risk involved in the accidents that epileptic seizures can cause while carrying out their daily live activities. It is for this reason that several attempts have been made to try to predict epileptic seizures before they occur. This project proposes to make use of the Hierarchical Temporal Memory model, an anomaly detector inspired by the neocortex, for the detection of abnormalities in the EEG signal, specifically, the preictal activity prior to the seizure. The main benefits of this innovative neural network are that it allows online unsupervised learning with streaming data, as well as adaptation to behavioural changes of the input, since it is continuously learning and updating as it sees new records. After understanding the model basis and replicating some available examples, an application programming interface has been created with the HTM model, which returns the anomaly likelihood as output value, a metric that describes the probability of a record of being anomalous. The aim is to obtain an output that allows us stablishing a threshold to differentiate between anomalous and non-anomalous data. The optimal model parameters and their influence on the output have been studied, being resolution and anomaly period the ones with a more direct impact on the model performance. Additionally, it has been discussed how resampling is a crucial step during the pre-processing phase in order to ease the sequential pattern learning of the algorithm. For the study, firstly, the performance of the model has been evaluated with a synthetic dataset constructed from real patient data, and it has been possible to see how the model is capable of detecting epileptic seizures as abnormal, but does not successfully detect anomalies in the data records that precede the seizure. It has also been seen how the model learned seizures as non-anomalous after having seen one before, so tests have been carried out training with an interictal dataset, to later disable continuous learning and pass them datasets containing seizures. Despite the satisfactory results, high anomaly score values make it doubtful that the algorithm is learning the interictal patterns and has a robust idea of what is normal. Trials performed with real patient data have ratified the previous conclusions. Finally, a comparison has been made between the HTM model and Isolation Forests (IF). HTM model has not been found advantageous compared to IF, as IF seem to detect data records before seizures but also suppose an increased risk of false positives. None of them has demonstrated to be a perfect anomaly detector with the EEG dataset used La calidad de vida de las personas con epilepsia se ve considerablemente alterada por el riesgo de sufrir crisis epilépticas y los accidentes que pueden ocasionar durante el desarrollo de sus actividades vitales cotidianas. Es por ello que se han realizado diversos intentos para tratar de predecir las crisis epilépticas antes de que se produzcan. Este proyecto propone hacer uso del modelo Hierarchical Temporal Memory, una red neuronal inspirada en el neocórtex, para la detección de anomalías en la señal EEG, concretamente, la actividad preictal previa a la crisis. Las principales ventajas de esta innovadora red neuronal son que permite el aprendizaje online no supervisado de datos a tiempo real, así como la adaptación a los cambios de comportamiento de los datos de entrada. Tras comprender las bases del modelo y replicar algunos ejemplos disponibles, se ha creado un programa con el modelo HTM, que devuelve como valor de salida la probabilidad de anomalía, una métrica que describe la probabilidad de que un registro sea anómalo. El objetivo es obtener una respuesta que nos permita establecer un umbral para diferenciar entre datos anómalos y no anómalos. Se han estudiado los parámetros óptimos del modelo y su influencia en la salida, siendo la resolución y el periodo de anomalía los que tienen un impacto más directo en el rendimiento del modelo. Además, se ha discutido cómo el remuestreo es un paso crucial durante la fase de preprocesamiento para facilitar el aprendizaje secuencial de patrones del algoritmo. Para el estudio, en primer lugar, se ha evaluado el rendimiento del modelo con señales EEG sintéticas construidas a partir de datos reales de pacientes, y se ha podido comprobar cómo el modelo es capaz de detectar ataques epilépticos como anormales, pero no detecta con éxito anomalías en los registros de datos que preceden a la crisis. También se ha visto como el modelo aprendía las crisis como no anómalas después de haber visto una anteriormente, por lo que se han realizado pruebas entrenando con datos interictales, para posteriormente desactivar el aprendizaje continuo y pasarles datos que contienen crisis. A pesar de los resultados satisfactorios, los valores de anomalía instantánea son muy elevados y hacen dudar sobre si el algoritmo está aprendiendo los patrones interictales y si tiene una idea robusta de lo que es normal. Los ensayos realizados con datos de pacientes reales han ratificado las conclusiones anteriores. Por último, se ha realizado una comparación entre el modelo HTM y los Isolation Forests (IF). El modelo HTM no ha resultado ventajoso en comparación con el IF, ya que el IF parece detectar como anómalos registros de datos antes de las crisis, pero también supone un mayor riesgo de falsos positivos. Ninguno de ellos ha demostrado ser un detector de anomalías perfecto con el conjunto de datos de EEG utilizado La qualitat de vida de les persones amb epilèpsia es veu considerablement alterada pel risc de patir crisis epilèptiques i els accidents que poden ocasionar durant el desenvolupament de les activitats quotidianes. És per això que s'han realitzat diversos intents per predir les crisis epilèptiques abans que es produeixin. Aquest projecte proposa fer ús del model Hierarchical Temporal Memory, una xarxa neuronal inspirada en el neocòrtex, per a la detecció d'anomalies al senyal EEG, concretament, l'activitat preictal prèvia a la crisi. Els principals avantatges d'aquesta innovadora xarxa neuronal són que permet l'aprenentatge en línia no supervisat de dades a temps real, així com l'adaptació als canvis de comportament de les dades d'entrada. Després de comprendre les bases del model i replicar-ne alguns exemples disponibles, s'ha creat un programa amb el model HTM, que torna com a valor de sortida la probabilitat d'anomalia, una mètrica que descriu la probabilitat que un registre sigui anòmal. L'objectiu és obtenir una resposta que ens permeti establir un llindar per diferenciar entre dades anòmales i no anòmales. S'han estudiat els paràmetres òptims del model i la seva influència a la sortida, sent la resolució i el període d'anomalia els que tenen un impacte més directe en el rendiment del model. A més, s'ha discutit com el remostreig és un pas crucial durant la fase de preprocessament per facilitar l'aprenentatge seqüencial de patrons de l'algorisme. Per a l'estudi, en primer lloc, s'ha avaluat el rendiment del model amb senyals EEG sintètics construïts a partir de dades reals de pacients, i s'ha pogut comprovar com el model és capaç de detectar la senyal durant la crisis epilèptica com a anormal, però no detecta amb èxit anomalies als registres de dades que precedeixen la crisi. També s'ha vist com el model aprenia les crisis com no anòmales després d'haver-ne vist una anteriorment, per la qual cosa s'han realitzat proves entrenant-se amb dades interictals, per posteriorment desactivar l'aprenentatge continu i passar dades que contenen crisi. Tot i els resultats satisfactoris, els valors d'anomalia instantània són molt elevats i fan dubtar si l'algorisme està aprenent els patrons interictals i si té una idea robusta del que és normal. Els assaigs realitzats amb dades de pacients reals han ratificat les conclusions anteriors. Per acabar, s'ha realitzat una comparació entre el model HTM i els Isolation Forests (IF). El model HTM no ha resultat avantatjós en comparació amb els IF, ja que aquest últim sembla detectar com a anòmals registres de dades abans de les crisis, però també suposa un risc més gran de falsos positius. Cap no dels dos ha demostrat ser un detector d'anomalies perfecte amb el conjunt de dades d'EEG utilitzat

L’objectiu principal d’aquest treball és explorar els camps de la Intel∙ligència Artificial explicable (XAI per les seves sigles en anglès) i de la Intel∙ligència Artificial multimodal, arribant a desenvolupar un mètode d’explicabilitat per a Intel∙ligència Artificial multimodal en l’àmbit de la salut. S’ha concretat una metodologia per obtenir explicacions dels models que combinen la informació extreta de més d’una modalitat de dades. L’estratègia es basa en extreure característiques de les imatges que es puguin presentar en format tabular per posteriorment ser annexades a les dades clíniques tabulars i aplicant mètodes d’explicabilitat a un model entrenat amb totes les dades tabulars. El mètode desenvolupat està acotat a casos d’ús que combinin imatge mèdica i dades clíniques tabulars, és agnòstic al model, proporciona explicabilitat global i local i és de possible aplicació en múltiples casos d’ús. Per provar i millorar el mètode desenvolupat, s’ha utilitzat un cas d’ús de diagnòstic de la Malaltia d’Alzheimer (MA), incorporant imatges de ressonància magnètica (MRI, per les seves sigles en anglès) i dades clíniques tabulars per a cada pacient. El resultat d’aquest projecte el conformen aquesta Memòria i Annexos, així com tres documents en format notebook amb tot el codi desenvolupat en llenguatge Python. Es conclou que el treball té moltes possibilitats d’ampliació i optimització futures. El principal objetivo de este trabajo es explorar los campos de la Inteligencia Artificial explicable (XAI por sus siglas en inglés) y de la Inteligencia Artificial multimodal, llegando a desarrollar un método de explicabilidad para Inteligencia Artificial multimodal en el ámbito de la salud. Se ha concretado una metodología para obtener explicaciones de los modelos que combinan la información extraída de más de una modalidad de datos. La estrategia se basa en extraer características de las imágenes que se puedan presentar en formato tabular para posteriormente ser anexados a los datos clínicos tabulares y aplicando métodos de explicabilidad a un modelo entrenado con todos los datos tabulares. El método desarrollado está acotado a casos de uso que combinen imagen médica y datos clínicos tabulares, es agnóstico al modelo, proporciona explicabilidad global y local y es de posible aplicación en múltiples casos de uso. Para probar y mejorar el método desarrollado, se ha utilizado un caso de uso de diagnóstico de la Enfermedad de Alzheimer, incorporando imágenes de resonancia magnética (MRI, por sus siglas en inglés) y datos clínicos tabulares para cada paciente. El resultado de este proyecto lo conforman esta Memoria y Anexos, así como tres documentos en formato notebook con todo el código desarrollado en lenguaje Python. Se concluye que el trabajo tiene muchas posibilidades de ampliación y optimización futuros. The main goal of this work is to explore the fields of eXplainable Artificial Intelligence (XAI) and multimodal Artificial Intelligence, developing an explainability method for multimodal Artificial Intelligence in the field of health. This work specifies a methodology to obtain explanations of the models that combine the information extracted from more than one data modality. The strategy is based on extracting characteristics from the images that can be presented in tabular format, attaching them afterwards to the tabular clinical data and applying explainability methods to a model trained with all the tabular data. The developed method is limited to the use cases that combine medical imaging and tabular clinical data, it is agnostic to the model, provides global and local explainability and it can be applied in multiple use cases. To test and improve the developed method, a diagnostic use case of Alzheimer’s Disease (AD) has been used, incorporating magnetic resonance images (MRI) and tabular clinical data for each patient. The result of this project consists on this Report and Annexes, as well as three documents in notebook format with all the code developed in Python. It is concluded that this work has many opportunities of improvement and optimization.

La pandèmia viscuda arreu del món a inicis del món a causa del SARS-CoV-2, agent causant de la malaltia coneguda com a coronavirus, ha vist reflectida la necessitat de models matemàtics que poguessin ajudar a explicar a la població com es trobava la pandèmia usant indicadors, paràmetres i gràfics que, de forma senzilla, donessin un estat el qual s’anava actualitzant diàriament i així poder prendre decisions pel que fa a mesures sanitàries i de control i aïllament social. En aquest treball es vol realitzar una comparativa a partir de les dades reals de la pandèmia al territori català, el Baix Llobregat i a l’Alt Penedès treballades mitjançant els model de Gompertz i ARIMA. Per a tots dos models, amb el programari Matlab, s’han validat les dades de partida, s’han realitzat modelitzacions i s’ha comprovat que, efectivament, són capaces de realitzar pronòstics en quant a casos de coronavirus i en quant a la mortalitat de la malaltia. Pel model de Gompertz, model ja conegut i utilitzat en dinàmiques de poblacions en ser un model amb paràmetres físics, s’obtenen resultats prou satisfactoris a partir d’aquests paràmetres, ajustats a partir de l’optimització no lineal per mitjà de mínims quadrats. Pel que fa al model ARIMA, en ser aquest un model autoregressiu, necessita de l’estudi de la sèrie temporal, el tractament d’aquesta i el tenir un algoritme propi per a poder modelitzar els seus paràmetres, els quals no són físics si no matemàtics. A més, es corre el risc de cometre overfitting, o sobreentrenar el model, si s’utilitza tota la sèrie temporal. Així, s’ha decidit usar mètodes de train/test amb les dades inicials per a dur a terme el model, obtenint uns resultats molt satisfactoris, tot i ser més complicats de tractar. Així doncs, durant l’estudi es trobaran explicacions de caire teòric de les sèries temporals i els models estudiats, com s’han validat i a quins territoris, avaluant els errors relatius comesos respecte als territoris que serveixen de referència, i per últim, ja amb els models validats, es realitzen els pronòstics a tres territoris ben diferenciats per la seva naturalesa: Catalunya, amb molta població, el Baix Llobregat amb una alta densitat de població i l’Alt Penedès, comarca amb força menys població i densitat de població que el Baix Llobregat, i amb els pronòstics s’avaluaran dos dels indicadors principals de la pandèmia com són la incidència acumulada i la velocitat de propagació

Sentinel lymph node status is a crucial prognosis factor for melanomas; nonetheless, the invasive surgery required to obtain it always puts the patient at risk. In this study, we develop a Deep Learning-based approach to predict lymph node metastasis from Whole Slide Images of primary tumours. Albeit very informative, these images come with complexities that hamper their use in machine learning applications, namely their large size and limited datasets. We propose a pre-training strategy based on self-supervised contrastive learning to extract better image feature representations and an attention-based Multiple Instance Learning approach to enhance the model’s performance. With this work, we quantitatively demonstrate that combining both methods improves various classification metrics and qualitatively show that contrastive learning encourages the network to output higher attention scores to tumour tissue and lower scores to image artifacts. Work supported by the Spanish Research Agency (AEI) under project PID2020-116907RB-I00 of the call MCIN/AEI/10.13039/501100011033 and the project 718/C/ 2019 funded by Fundació la Marato de TV3.

On the one hand, clinical trials are of great importance in discovering new treatments for diseases. They teach research things that cannot be learned in the laboratory, that is, what does and does not work in humans. What’s more, they are a really helpful tool for deciding whether the side effects of a new drug or treatment are acceptable compared to the potential benefits. Since results are not known at the beginning, this process is quite uncertain and effective management of doses is required. On the other hand, Reinforcement Learning is a Machine Learning paradigm different from Supervised Learning and Unsupervised Learning. Unlike these two methods, Reinforcement Learning is used when we want a system to accomplish a task by understanding the task by itself and the system must learn based on the “trial and error” rule. In recent years, several Reinforcement Learning applications have brought great advances in different fields. However, this new paradigm has not been used for optimizing the inventory management of clinical trials, so that’s why this master thesis addresses this line of investigation. The contributions of this master thesis are: (1) The definition of the Markov Decision Process formulation for the current problem, (2) The application of Tabular Methods to solve the problem, and (3) The improvement of the used Tabular Methods by modifying their respective Q-Tables. Results show a considerable improvement after the Q-Table’s modification. However, the system does not achieve a great performance mainly due to the high dimensionality of the problem Objectius de Desenvolupament Sostenible::3 - Salut i Benestar

The monitoring of biomechanical data during the crutch gait is essential when assessing the progression of the patient during rehabilitation. However, the complexity of this data may hidden embedded patterns on the user gait that could provide additional information about their health condition, and may affect the evolution of the patient. During the last years, research focusing on the implementation of ML techniques for biomechanical analysis have been gaining relevance, due to their capacity for identifying gait disorders, detecting specific gait events or classifying user activity. The main objective of this project, therefore, is to classify 4 different gait patterns through the extraction of relevant spatio-temporal parameters applying ML techniques. To achieve this objective, gait data were collected from a set of instrumented crutches, which allowed to record the trials of 27 healthy participants, along a 10 meter corridor. The collected gait signals were processed and segmented in cycles, by the identification of swing and stance phases using the pitch angle signal of the crutch. The final dataset obtained after this segmentation was composed by a total of 74 extracted features and 6014 samples. The acquisition of this dataset allowed to build a ML classifier to accurately identify the different gait patterns. In order to reduce the data dimensionality, a feature selection process is performed, comparing the most relevant methods applied in literature, and demonstrating the importance of forward acceleration, pitch and force signals for crutch gait identification. The reduced dataset was evaluated against different classifiers through a hyperparameter optimization process, resulting in a SVM classifier able to identify the crutch patterns with a 93.85% of accuracy. Additionally, a solution to reject samples not belonging to one of the expected gait patterns is presented Objectius de Desenvolupament Sostenible::3 - Salut i Benestar Objectius de Desenvolupament Sostenible::9 - Indústria, Innovació i Infraestructura::9.b - Donar suport al desenvolupament de tecnologies, investigació i innovació nacionals als països en desenvolu­pament, garantint també un entorn normatiu propici a la diversificació industrial i l’addició de valor als productes bàsics, entre d’altres

Traditional machine learning methods amongst others segment an image into different regions on basis of pixel attributes. One of the most well known methods are clustering and thresholding algorithms. Convolutional neural networks are designed to imitate the human visual cortex by applying convolutional filters in form of layers to the input images. That way, deep features are extracted. Since this process is often referred to a “black box” behaviour because it is unknown what is going on inside the model, deep learning interpretability methods are introduced that highlight parts of an image that are important for the algorithm’s classification process. Experiments show that it is possible to establish links between segmented regions extracted by traditional methods and deep features extracted by CNNs. These results rely heavily on the DL interpretability method used and the type of dataset. This extended TFM is carried out in cooperation with a research group from the AMRITA University – school of medicine in India, focusing on biomedical image processing and computer vision. The collaboration started in November 2021 Objectius de Desenvolupament Sostenible::3 - Salut i Benestar::3.b - Donar suport a les activitats d’investigació i desenvolupament de vacunes i medicaments per a les malalties transmissibles i no transmissibles que afecten primordialment països en desenvolupament, i facilitar l’accés a medicaments i a vacunes essencials assequibles de conformitat amb la Declaració de Doha relativa a l’Acord sobre els aspectes dels drets de propietat intel·lectual relacionats amb el comerç (ADPIC) i la salut pública, en la qual s’afirma el dret dels països en desenvolupament a utilitzar al màxim les disposicions de l’ADPIC pel que fa a la flexibilitat per a protegir la salut pública, i en particular, proporcionar accés a medicaments per a totes les persones

The management of sick leaves is a critical task that public and private health systems carry out. This enables the good care of sick workers and guarantees a safe return to their jobs. Most health systems enforce regulations that establish the duration of sick leave according to general rules for groups of diagnoses. However, these regulations do not account for the particularities of workers. On the one hand, an early return to work is sometimes possible, but this does not happen unless the worker pro-actively requests it. On the other hand, the worker’s health condition could demand for one or more leave extensions, but the system requires mandatory and sometimes unnecessary follow-ups, adding nuisance to patients and overhead to health systems. In both cases, the lack and excess of action by the health system represents extra costs for society. This paper proposes the analysis of a voluminous historical dataset of sick leaves (including medical and personal data) to predict the duration of future sick leaves for patients. The data mining process is performed for a large number of diagnoses to assess the possibility of using data driven models for broad decision-making. The nature and characteristics of the data makes it difficult to obtain models using classical methods, which is why the analysis focuses on non-linear machine learning-based survival analysis methods. In sight of the models performance, we move forward to its practical implementation, proposing a tool to manage the decision of what patients should be contacted at a given date using the predictions of the trained models. This tool will manage the whole cycle, continuously training on new data, performing daily predictions, and presenting the results to the health-care decision-maker for their assessment. This work has been funded by a collaboration between Asepeyo and BSC. We want to thank the following people from Asepeyo for their help and support that made this project possible. Xavier Calatrava Petisme, Alex Nogué Martinez, Enric Lleal Serra, Francisco Manuel Ventura Nofuentes, Oscar González Cherta, Francisco Sánchez Algarra, Eulàlia Borén and Vanessa Vegazo. We also want to thank to Nadia Tonello (BSC Data Manager) and Ulises Cortés (UPC/BSC) for their insightful comments.

Mathematical modeling of the glucose–insulin system forms the core of simulators in the field of glucose metabolism. The complexity of human biological systems makes it a challenging task for the physiological models to encompass the entirety of such systems. Even though modern diabetes simulators perform a respectable task of simulating the glucose–insulin action, they are unable to estimate various phenomena affecting the glycemic profile of an individual such as glycemic disturbances and patient behavior. This research work presents a potential solution to this problem by proposing a method for the generation of blood glucose values conditioned on plasma insulin approximation of type 1 diabetes patients using a pixel-to-pixel generative adversarial network. Two type-1 diabetes cohorts comprising 29 and 6 patients, respectively, are used to train the generative model. This study shows that the generated blood glucose values are statistically similar to the real blood glucose values, mimicking the time-in-range results for each of the standard blood glucose ranges in type 1 diabetes management and obtaining similar means and variability outcomes. Furthermore, the causal relationship between the plasma insulin values and the generated blood glucose conforms to the same relationship observed in real patients. These results herald the aptness of deep generative models for the generation of virtual patients with diabetes This work was partially supported by the Spanish Ministry of Universities, the European Union through Next GenerationEU (Margarita Salas), the Spanish Ministry of Science and Innovation through grant PID2019107722RBC22/AEI/10.13039/501100011033, PID2020-117171RA-I00 funded by MCIN/AEI/10.13039/501100011033 and the Government of Catalonia under 2017SGR1551 and 2020 FI_B 00965

El caràcter d’aquest treball està encarat de forma explícita a la recerca i té com a objectiu determinar si és possible fer una predicció de la viabilitat dels tractaments de radioteràpia a través de la informació que s’obté de les diferents mètriques que defineixen aquests tractaments. Amb aquest enfocament es van definir dues etapes del projecte. La primera consistia en l’estudi de les mètriques i la segona en el desenvolupament d’un model predictiu que fos capaç de predir amb les mètriques seleccionades en l’estudi de l’etapa anterior quins tractaments són els que superen els índexs deviabilitat establerts. A trets generals podem dir que els objectius marcats s’han assolit: S’han estudiat les mètriques i s’han extret conclusions interessants que marquen un nou full de ruta per seguir la investigació. La principal conclusió és que la correlació entre la viabilitat del tractament i les mètriques de complexitat analitzades indica que, d’existir relació entre elles, aquesta no és lineal, per tant s’hauria de seguir la investigació utilitzant regressions no lineals o altres mètodes. A més, s’ha descobert una característica important de les dades que és la multicolinealitat i s’ha determinat que intentar-la solventar pot comportar en la majoria dels casos la pèrdua d’informació rellevant. És per això que s’ha establert que per a futurs estudis, és imprescindible fer ús de totes les mètriques. Amb la utilització dels mètodes de Machine Learning inclosos dins la llibreria de Scikitlearn de Python s’ha aconseguit obtenir un model que és capaç de predir en un 72% d’efectivitat quins són els tractaments que passen els índexs de viabilitat per poder ser utilitzats en pacients sense necessitat de fer proves amb maniquí. Tot i no ser un bon resultat donada la finalitat del predictor, ja que l’existència de falsos positius n’impossibiliten el seu ús en casos reals, s’ha pogut demostrar que no és possible fer una predicció de la viabilitat dels tractaments de radioteràpia utilitzant els mètodes de boscos aleatoris, màquines de suport vectorial, o KNN. Una aposta per seguir treballant en el desenvolupament d’un model que sigui capaç d’identificar els tractaments òptims per a ser utilitzats en pacients seria ampliar les fronteres del marc en què s’engloba aquest treball i fer ús d’altres tècniques en intel·ligència artificial com poden ser les xarxes neurals

Objectius de Desenvolupament Sostenible::3 - Salut i Benestar

BACKGROUND Endometriosis is an inflammatory oestrogen-dependent condition characterized by the presence of endometrial tissue outside the uterus. It affects women during childbearing age and has an association with pelvic pain and infertility. Despite a range of symptoms, diagnosis of endometriosis is often delayed due to a lack of non-invasive, definitive tools for the diagnosis of endometriosis. Traditionally, endometriosis was diagnosed by an exploratory laparoscopy with posterior histology analysis. Nowadays new diagnostic strategies are arising such as Transvaginal ultrasound (TVUS) due to its availability and low cost. Although ultrasound can diagnose most locations, its limited sensitivity for posterior lesions does not allow management decisions in all patients. MRI has shown high accuracies for anterior and posterior pelvic endometriosis and enables complete lesion mapping before surgery. Moreover, adding an artificial intelligence (AI) analysis system to the MRI description can provide the expert review that lacks in tertiary centres or other centres where there are no specialised radiologists in endometriosis. OBJECTIVE: To evaluate the accuracy of artificial intelligence analysis combined with MRI in comparison with the Gold Standard technique, exploratory laparoscopy, in the diagnosis and staging of endometriosis. METHODS: This protocol study is designed as a cross-sectional retrospective study. Pre-surgical MRIs of 30 female patients who have been diagnosed with endometriosis by laparoscopy in Girona during 2005 -2021 will be re-analysed by an AI analysis system in order to compare the AI + MRI with the Gold Standard, diagnostic laparoscopy

The current research focus in Robot-Assisted Minimally Invasive Surgery (RAMIS) is directed towards increasing the level of robot autonomy, to place surgeons in a supervisory position. Although Learning from Demonstrations (LfD) approaches are among the preferred ways for an autonomous surgical system to learn expert gestures, they require a high number of demonstrations and show poor generalization to the variable conditions of the surgical environment. In this work, we propose an LfD methodology based on Generative Adversarial Imitation Learning (GAIL) that is built on a Deep Reinforcement Learning (DRL) setting. GAIL combines generative adversarial networks to learn the distribution of expert trajectories with a DRL setting to ensure generalisation of trajectories providing human-like behaviour. We consider automation of tissue retraction, a common RAMIS task that involves soft tissues manipulation to expose a region of interest. In our proposed methodology, a small set of expert trajectories can be acquired through the da Vinci Research Kit (dVRK) and used to train the proposed LfD method inside a simulated environment. Results indicate that our methodology can accomplish the tissue retraction task with human-like behaviour while being more sample-efficient than the baseline DRL method. Towards the end, we show that the learnt policies can be successfully transferred to the real robotic platform and deployed for soft tissue retraction on a synthetic phantom., Accepted in IEEE International Symposium of Medical Robotics (ISMR 2021)

The use of automated insulin delivery systems has become a reality for people with type 1 diabetes (T1D), with several hybrid systems already on the market. One of the particularities of this technology is that the patient is in the loop. People with T1D are the plant to control and also a plant operator, because they may have to provide information to the control loop. The most immediate information provided by patients that affects performance and safety are the announcement of meals and exercise. Therefore, to ensure safety and performance, the human factor impact needs to be addressed by designing fault monitoring strategies. In this paper, a monitoring system is developed to diagnose potential patient modes and faults. The monitoring system is based on the residual generation of a bank of observers. To that aim, a linear parameter varying (LPV) polytopic representation of the system is adopted and a bank of Kalman filters is designed using linear matrix inequalities (LMI). The system uncertainty is propagated using a zonotopic-set representation, which allows determining confidence bounds for each of the observer outputs and residuals. For the detection of modes, a hybrid automaton model is generated and diagnosis is performed by interpreting the events and transitions within the automaton. The developed system is tested in simulation, showing the potential benefits of using the proposed approach for artificial pancreas systems This work was partially supported by the Spanish Ministry of Science and Innovation through grant PID2019-107722RB-C22, in part by the Autonomous Government of Catalonia under Grant 2017 SGR 1551, in part by the Ministerio de Educación, Cultura y Deporte under Grant FPU0244 2015, and in part EU through FEDER funds

HTE 3.0 aims to support clinicians in the detection of patients with dyslipidemia, especially patients with familial hypercholesterolemia (FH), and in the recommendation of personalized lipid-lowering treatments. The core of HTE 3.0 is a clinical decision support system in which several knowledge-based systems are serialized: patient detection, therapeutic target setting, personalized treatment assessment, and treatment combination and prioritization, according to different criteria. The experimental evaluation of HTE 3.0 shows that the use of HTE 3.0 would mean increasing the capacity to detect FH by 5.7 times compared with usual clinical practice. Regarding the lipid-lowering treatment, a comparison of 18 cases among seven lipidologists shows that the differences between treatments provided by HTE 3.0 and human lipidologists are smaller than the differences between human experts This work has been developed with the support of the research group SITES awarded with distinction by the Generalitat de Catalunya(2017 SGR 1551) Open Access funding provided thanks to the CRUE-CSIC agreement with Wiley

La Inteligencia Artificial se ha vuelto una disciplina que trata cualquier ámbito, en este trabajo se ha realizado una aplicación en el ámbito de la salud. Se ha creado un modelo de aprendizaje automático, que clasifica imágenes de rayos-X del tórax, en positivas o negativas frente al virus COVID-19, previamente clasificados. El objetivo de este trabajo es encontrar un modelo que realice dicha clasificación correctamente, mediante redes neuronales. Los datos utilizados son proporcionados por la página web kaggle[12] y se trata de datos desequilibrados, se han estudiado los parámetros, métodos y métricas a tener en cuenta a la hora de clasificar datos desequilibrados. Se han entrenado varios modelos, un primer modelo con los datos desequilibrados, un segundo modelo añadiendo un ajuste al modelo anterior y finalmente, se ha realizado un modelo a partir de un subconjunto de las muestras para poder tratar con datos equilibrados y mejorar la clasificación. Los resultados obtenidos no han sido los esperados, debido a la similitud entre las imágenes y la complicación que conlleva tratar con datos desequilibrados, sin embargo, se han obtenido resultados significativos. El modelo que mejor ha realizado la clasificación ha conseguido detectar un 45,5 % de las muestras con COVID-19 Artificial Intelligence has become a discipline that deals with any field, in this work an application has been made in the field of health. A deep learning model has been created, which classifies X-ray images of the chest, as positive or negative against the COVID-19 virus, previously classified. The objective of this work is to find a model that performs this classification correctly, using neural networks. The data used are provided by the kaggle [12] website and it is about unbalanced data. The parameters, methods and metrics to take into account when classifying unbalanced data have been studied. Several models have been trained, a first model with unbalanced data, a second model adding an adjustment to the previous model and finally, a model has been made from a subset of the samples to be able to deal with balanced data and improve the classification. The results obtained have not been as expected, due to the similarity between the images and the complication involved in dealing with unbalanced data, however, significant results have been obtained. The model that performed the best classification detected 45.5 % of the samples with COVID-19 Objectius de Desenvolupament Sostenible::3 - Salut i Benestar

The time spent in glucose ranges is a common metric in type 1 diabetes (T1D). As the time in one day is finite and limited, Compositional Data (CoDa) analysis is appropriate to deal with times spent in different glucose ranges in one day. This work proposes a CoDa approach applied to glucose profiles obtained from six T1D patients using continuous glucose monitor (CGM). Glucose profiles of 24-h and 6-h duration were categorized according to the relative interpretation of time spent in different glucose ranges, with the objective of presenting a probabilistic model of prediction of category of the next 6-h period based on the category of the previous 24-h period. A discriminant model for determining the category of the 24-h periods was obtained, achieving an average above 94% of correct classification. A probabilistic model of transition between the category of the past 24-h of glucose to the category of the future 6-h period was obtained. Results show that the approach based on CoDa is suitable for the categorization of glucose profiles giving rise to a new analysis tool. This tool could be very helpful for patients, to anticipate the occurrence of potential adverse events or undesirable variability and for physicians to assess patients’ outcomes and then tailor their therapies This project has been partially supported by the Spanish Government, Ministerio de Economía y Competitividad (MINECO) (Grants DPI2016-78831-C2-1-R, DPI2016-78831-C2-2-R, RTI2018-095518-B-C21), Agencia Estatal de Investigación (PID2019-107722RB-C21/ AEI/10.13039/501100011033), the National Council of Technological and Scientific Development, CNPq Brazil through Grants 202050/2015-7, 207688/2014-1 and EU through FEDER funds

Preclinical testing and validation of therapeutic strategies developed for patients with type 1 diabetes (T1D) require a cohort of virtual patients (VPs). However, current simulators provide a limited number of VPs, lack real-life scenarios, and inadequately represent intra- and inter-day variability in insulin sensitivity and blood glucose (BG) profile. The generation of a realistic scenario was achieved by using the meal patterns, insulin profiles (basal and bolus), and exercise sessions estimated as disturbances using clinical data from a cohort of 14 T1D patients using the Medtronic 640G insulin pump provided by the Hospital Clínic de Barcelona. The UVa/Padova’s cohort of adult patients was used for the generation of a new cohort of VPs. Insulin model parameters were optimized and adjusted in a day-by-day fashion to replicate the clinical data to create a cohort of 75 VPs. All primary and secondary outcomes reflecting the BG profile of a T1D patient were analyzed and compared to the clinical data. The mean BG 166.3 versus 162.2 mg/dL (p = 0.19), coefficient of variation 32% versus 33% (p = 0.54), and percent of time in range (70 to 180 mg/dL) 59.6% versus 66.8% (p = 0.35) were achieved. The proposed methodology for generating a cohort of VPs is capable of mimicking the BG metrics of a real cohort of T1D patients from the Hospital Clínic de Barcelona. It can adopt the inter-day variations in the BG profile, similar to the observed clinical data, and thus provide a benchmark for preclinical testing of control techniques and therapy strategies for T1D patients.

Multiple myeloma (MM) is a highly heterogeneous disease of malignant plasma cells. Diagnosis and monitoring of MM patients is based on bone marrow biopsies and detection of abnormal immunoglobulin in serum and/or urine. However, biopsies have a single-site bias; thus, new diagnostic tests and early detection strategies are needed. Matrix-Assisted Laser Desorption/Ionization Time-of Flight Mass Spectrometry (MALDI-TOF MS) is a powerful method that found its applications in clinical diagnostics. Artificial intelligence approaches, such as Artificial Neural Networks (ANNs), can handle non-linear data and provide prediction and classification of variables in multidimensional datasets. In this study, we used MALDI-TOF MS to acquire low mass profiles of peripheral blood plasma obtained from MM patients and healthy donors. Informative patterns in mass spectra served as inputs for ANN that specifically predicted MM samples with high sensitivity (100%), specificity (95%) and accuracy (98%). Thus, mass spectrometry coupled with ANN can provide a minimally invasive approach for MM diagnostics.

People with type 1 diabetes must control their blood glucose level through insulin infusion either with several daily injections or with an insulin pump. However, estimating the required insulin dose is not easy. Recommender systems, mainly based on Case-Based Reasoning (CBR), are being developed to provide recommendations to users. These systems are designed to keep the experiences or cases of the user in a case-base, which requires maintenance to keep system's response accurate and efficient. This paper proposes a case-base maintenance methodology that combines case-base redundancy reduction and attribute weight learning. Contrary to previous approaches designed for classification problems, the maintenance methodology presented in this paper deals with numerical recommendations. It can manage a potentially huge case-base due to the combinatorial derived from the number of attributes used to represent a case. The proposed approach has been tested using the UVA/PADOVA type 1 diabetes simulator and the results demonstrate that it can accomplish better levels of accuracy than other insulin recommender systems mentioned in the literature, when a large number of attributes is considered This project has received funding from the European Union Horizon 2020 research and innovation programme under grant agreement No. 689810, www.pepper.eu.com/, PEPPER, and the grant of the University of Girona 20162018 (MPCUdG2016). The work has been developed with the support of the research group SITES awarded with distinction by the Generalitat de Catalunya (SGR 20142016)

In personalized healthcare, an ecosystem for the manipulation of reliable and safe private data should be orchestrated. This paper describes an approach for the generation of synthetic electrocardiograms (ECGs) based on Generative Adversarial Networks (GANs) with the objective of anonymizing users’ information for privacy issues. This is intended to create valuable data that can be used both in educational and research areas, while avoiding the risk of a sensitive data leakage. As GANs are mainly exploited on images and video frames, we are proposing general raw data processing after transformation into an image, so it can be managed through a GAN, then decoded back to the original data domain. The feasibility of our transformation and processing hypothesis is primarily demonstrated. Next, from the proposed procedure, main drawbacks for each step in the procedure are addressed for the particular case of ECGs. Hence, a novel research pathway on health data anonymization using GANs is opened and further straightforward developments are expected.

The high infectivity of SARS-CoV-2 makes it essential to develop a rapid and accurate diagnostic test so that carriers can be isolated at an early stage. Viral RNA in nasopharyngeal samples by RT-PCR is currently considered the reference method although it is not recognized as a strong gold standard due to certain drawbacks. Here we develop a methodology combining the analysis of from human nasopharyngeal (NP) samples by matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) with the use of machine learning (ML). A total of 236 NP samples collected in two different viral transport media were analyzed with minimal sample preparation and the subsequent mass spectra data was used to build different ML models with two different techniques. The best model showed high performance in terms of accuracy, sensitivity and specificity, in all cases reaching values higher than 90%. Our results suggest that the analysis of NP samples by MALDI-TOF MS and ML is a simple, safe, fast and economic diagnostic test for COVID-19. SAUN—Santander Universidades-CRUE (PEDIEC from FONDO SUPERA COVID-19) 5.093 JCR (2020) Q1, 28/167 Medicine, General & Internal 1.388 SJR (2020) Q1, 304/2446 Medicine (miscellaneous) No data IDR 2020 UEM

Unas décadas atrás, nació Deep Learning, una técnica que revolucionó el mundo de la Inteligencia Artificial tal y como se conocía. Esta imita el funcionamiento neuronal de un cerebro humano para clasificar conjuntos de datos y encontrar patrones entre ellos. Gracias a este sistema computacional, en este trabajo se ha podido diseñar un modelo clasificatorio de imágenes que detecte la neumonía. Para ello, se ha hecho uso de la metodología Transfer Learning. Esta se basa en utilizar modelos entrenados previamente para crear el nuevo modelo. Además se ha utilizado las redes neuronales convolucionales como extractoras de patrones. Se ha decidido identificar esta patología porque la neumonía es una infección respiratoria que, si se complica, puede llegar a ser letal, sobre todo en niños menores de cinco años. Del mismo modo, se ha creado este modelo con el objetivo de que se pueda utilizar en estudios clínicos para así poder encontrar nuevos tratamientos o controlar la enfermedad. Para crear el modelo, se ha utilizado la librería Fast.ai. debido a que tiene componentes de alto nivel que permiten proporcionar resultados de forma más sencilla y rápida. El conjunto de datos utilizados para crear el modelo se ha extraído de la página web Kaggle. Como se quería obtener el mejor rendimiento posible, se han realizado cuatro modelos distintos a partir de cuatro modelos pre entrenados; ResNet 18, ResNet 34, ResNet 50 y AlexNet. Al final, se han acabado obteniendo resultados muy buenos a partir de todos los modelos, incluso en tres de ellos se ha alcanzado una precisión mayor del 99%. Aun así, ha sido a partir de ResNet 50 que se ha obtenido el mejor rendimiento y con el que se ha creado el programa que consigue detectar la neumonía A few decades ago, Deep Learning appeared revolutionizing the world of Artificial Intelligence as it was known. This technique mimics the neural functioning of a human brain to classify data sets and to find patterns among them. Thanks to this computational system, it has been possible to design through this project a classifying model of images that detects pneumonia. Transfer Learning has been the methodology used to achieve so. This method is based on using previously trained models to create the new model. Furthermore, convolutional neural networks have been used as pattern extractors. It was decided to identify this pathology because pneumonia is a respiratory infection that, if it gets bad, can be lethal, especially in children under five years old. Likewise, this model has been created with the aim of being used in clinical studies in order to find new treatments and to control the disease. To create the model, the Fast.ai. library has been used because it has high-level components that can quickly and easily provide results. The dataset used to create the model was extracted from the Kaggle website. As it was wanted to obtain the best possible performance, four different models have been made from four pre-trained models; ResNet 18, ResNet 34, ResNet 50 and AlexNet. In the end, very good results have been obtained from all the models, even in three of them an accuracy of more than 99% has been achieved. However, the best performance was obtained from ResNet 50. It was with this model that the program that detects pneumonia was created Objectius de Desenvolupament Sostenible::3 - Salut i Benestar::3.4 - Per a 2030, reduir en un terç la mortalitat prematura per malalties no transmissibles, mitjançant la prevenció i el tractament, i promoure la salut mental i el benestar

Skin cancer is a major public health issue with millions of cases diagnosed around the world. Melanoma is the most mortal form of skin cancer and causes the highest percentage of deaths. Although the mortality rate for melanoma is significant, when an early diagnosis is made more than the 95% of people survive. Therefore, a neural network image classification would be significantly helpful during diagnosis. Artificial neural networks have become a popular machine learning model due to the ability to make nonlinear and complex relationships between thousands of variables. In fact, on occasion convolutional neural networks pattern detection abilities surpass their human counterparts. The main objective of this thesis is to design a convolutional neural network which is able to distinguish melanoma from other skin lesions with a high rate of accuracy. Neural networks were trained and tested using images from HAM10000 dataset. Using a replicated factorial fractional design of experiments, we have realized the importance of balance in classification. Using different oversampling methods we have obtained a F1-score of 84% in the test data. As a result, we have demonstrated that it is possible to classify melanoma with greater precision. In fact, these results appear to be better than the usual results achieved through multi-class classification. Furthermore, we believe that with more computational capabilities better results could be reached El càncer de pell és un problema important de salut pública, amb milions de casos diagnosticat sarreu del món. El melanoma és la forma més mortal de càncer de pell, responsable de la immensa majoria de les morts per càncer de pell. Tot i que la mortalitat és significativa, quan es detecta precoçment, la supervivència del melanoma supera el 95%. És per aquest motiu que la classificació d'imatges mitjançant xarxes neuronals pot ser de gran ajuda per als diagnòstics.Les xarxes neuronals són un model d'aprenentatge automàtic molt utilitzat actualment per la seva capacitat d'establir relacions complexes i no lineals entre milers de variables, i arriben a superar-les capacitats humanes en detecció de patrons. L'objectiu principal d'aquest projecte és crear una xarxa neuronal convolucional capaç de distingir els melanomes entre diferents lesions cutànies amb gran precisió.Les xarxes neuronals han sigut entrenades i provades fent servir les imatges de la basede dades HAM10000. A partir d'un disseny d'experiments factorial fraccionat de 26-2 amb rèpliques s'ha obtingut el coneixement necessari per centrar-se en el balanceig de les classes en la classificació binària. S'ha obtingut una F1-score de 84% en el seu test. Aquests resultats demostren que és possible fer aquesta classificació amb gran precisió d'aquestes imatges dels melanomes. De fet, sembla que aquests resultats són millors que els que s'acostumen a obtenir mitjançant una classificació de múltiples classes i es creu que, amb més capacitats computacionals, es podria obtenir encara millors resultats Objectius de Desenvolupament Sostenible::3 - Salut i Benestar

Background: The Patient Empowerment through Predictive Personalized Decision Support (PEPPER) system provides personalized bolus advice for people with type 1 diabetes. The system incorporates an adaptive insulin recommender system (based on case-based reasoning, an artificial intelligence methodology), coupled with a safety system, which includes predictive glucose alerts and alarms, predictive low-glucose suspend, personalized carbohydrate recommendations, and dynamic bolus insulin constraint. We evaluated the safety and efficacy of the PEPPER system compared to a standard bolus calculator. Methods: This was an open-labeled multicenter randomized controlled crossover study. Following 4-week run-in, participants were randomized to PEPPER/Control or Control/PEPPER in a 1:1 ratio for 12 weeks. Participants then crossed over after a washout period. The primary end-point was percentage time in range (TIR, 3.9-10.0 mmol/L [70-180 mg/dL]). Secondary outcomes included glycemic variability, quality of life, and outcomes on the safety system and insulin recommender. Results: Fifty-four participants on multiple daily injections (MDI) or insulin pump completed the run-in period, making up the intention-to-treat analysis. Median (interquartile range) age was 41.5 (32.3-49.8) years, diabetes duration 21.0 (11.5-26.0) years, and HbA1c 61.0 (58.0-66.1) mmol/mol. No significant difference was observed for percentage TIR between the PEPPER and Control groups (62.5 [52.1-67.8] % vs. 58.4 [49.6-64.3] %, respectively, P = 0.27). For quality of life, participants reported higher perceived hypoglycemia with the PEPPER system despite no objective difference in time spent in hypoglycemia. Conclusions: The PEPPER system was safe, but did not change glycemic outcomes, compared to control. There is wide scope for integrating PEPPER into routine diabetes management for pump and MDI users. Further studies are required to confirm overall effectiveness. Clinical trial registration: ClinicalTrials.gov NCT03849755 The work outlined in this manuscript was developed with the support of the awarded distinction by the Generalitat de Catalunya 2017 SGR 1551, and the European Union’s Horizon 2020 research and innovation programme under grant agreement No 689810 (PEPPER)

Skin cancer is a major public health issue with millions of cases diagnosed around the world. Melanoma is the most mortal form of skin cancer and causes the highest percentage of deaths. Although the mortality rate for melanoma is significant, when an early diagnosis is made more than the 95% of people survive. Therefore, a neural network image classification would be significantly helpful during diagnosis. Artificial neural networks have become a popular machine learning model due to the ability to make nonlinear and complex relationships between thousands of variables. In fact, on occasion convolutional neural networks pattern detection abilities surpass their human counterparts. The main objective of this thesis is to design a convolutional neural network which is able to distinguish melanoma from other skin lesions with a high rate of accuracy. Neural networks were trained and tested using images from HAM10000 dataset. Using a replicated factorial fractional design of experiments, we have realized the importance of balance in classification. Using different oversampling methods we have obtained a F1-score of 84% in the test data. As a result, we have demonstrated that it is possible to classify melanoma with greater precision. In fact, these results appear to be better than the usual results achieved through multi-class classification. Furthermore, we believe that with more computational capabilities better results could be reached El càncer de pell és un problema important de salut pública, amb milions de casos diagnosticat sarreu del món. El melanoma és la forma més mortal de càncer de pell, responsable de la immensa majoria de les morts per càncer de pell. Tot i que la mortalitat és significativa, quan es detecta precoçment, la supervivència del melanoma supera el 95%. És per aquest motiu que la classificació d'imatges mitjançant xarxes neuronals pot ser de gran ajuda per als diagnòstics.Les xarxes neuronals són un model d'aprenentatge automàtic molt utilitzat actualment per la seva capacitat d'establir relacions complexes i no lineals entre milers de variables, i arriben a superar-les capacitats humanes en detecció de patrons. L'objectiu principal d'aquest projecte és crear una xarxa neuronal convolucional capaç de distingir els melanomes entre diferents lesions cutànies amb gran precisió.Les xarxes neuronals han sigut entrenades i provades fent servir les imatges de la basede dades HAM10000. A partir d'un disseny d'experiments factorial fraccionat de 26-2 amb rèpliques s'ha obtingut el coneixement necessari per centrar-se en el balanceig de les classes en la classificació binària. S'ha obtingut una F1-score de 84% en el seu test. Aquests resultats demostren que és possible fer aquesta classificació amb gran precisió d'aquestes imatges dels melanomes. De fet, sembla que aquests resultats són millors que els que s'acostumen a obtenir mitjançant una classificació de múltiples classes i es creu que, amb més capacitats computacionals, es podria obtenir encara millors resultats Objectius de Desenvolupament Sostenible::3 - Salut i Benestar

BACKGROUND Diabetic ketoacidosis (DKA) is one of the most serious diabetes complications. It is characterized by hyperglycemia, anion gap metabolic acidosis and increased total body ketone concentration. DKA is the most common cause of death in youth type 1 diabetic patients and an important cause of morbimortality in type 2 diabetic patients. Its management is complex, however, if it is well performed DKA has a good prognosis. Lately, the need to standardize DKA treatment has become an important issue in our Department. OBJECTIVE Our main objective is to decrease hospital length of stay of DKA patients treated in Hospital Universitari Josep Trueta’s Emergency Department (ED) with the application of a computer decision support and electronic order sets (CDS&EOS) to minimize security errors. DESIGN Quasi-experimental study designed as a before-and-after evaluation of the application of a CDS&EOS in ED’s SILICON® to standardize DKA management. PARTICIPANTS A consecutive non-probabilistic model will be used to select DKA patients aged 18 years old or older treated in Hospital Universitari Josep Trueta’s ED. METHODS The study will include 134 participants in total, 67 for each group (pre and post- intervention). Each sample will be selected in an 18-month period, with a washout period between them. Data will be collected prospectively between May 2021 and October 2024. The association between the independent and dependent variables will be adjusted to avoid possible confounding factors