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

Diagnostic assistant systems or preventive detectors of health risks are nowadays possible thanks to the amount of digital healthcare data available and to machine learning techniques, which are promoting the development of knowledge discovery in databases. Extracting temporal relationships between medical events is essential to reveal hidden patterns that can help physicians find optimal treatments, diagnose illnesses, detect drug adverse reactions, and predict risk conditions in multimorbid patients. This project presents an approach to analyze medical text from clinical histories written in Catalan and/or Spanish and automatically extract patient evolution patterns that can aid in diagnosis prediction. This thesis proposes a framework to extract temporal patterns based on a combination of node embeddings and a Dynamic Time Warp (DTW) algorithm. First, the electronic health records of each patient are represented as a graph, where each node corresponds to a medical instance. Random walks are performed in the graphs to obtain sentences of medical concepts. These sequences are fed into a word2vec model to learn node representations for each medical concept. The node embeddings are subsequently used to compare sequences that embody the development of some given diagnoses in a DTW approach. Using a DTW-based algorithm, temporal patterns can be extracted from this collection of sequences. To demonstrate the effectiveness of this methodology, it has been applied to extract patient evolution patterns from clinical histories of multimorbid patients who suffered from one of the following most frequent diseases: type II diabetes, ischemic cardiopathy, acute myocardial infarction, colon cancer and lung cancer. Furthermore, the work is not limited to the aforementioned diagnoses, but it extends the exploration to any type of diagnosis. Due to the real-world and error-prone condition of the data, this framework proposes a validation process of the results performed by expert practitioners in primary care. Despite the limited size of the dataset, the high percentage of patterns deemed correct and relevant by the experts is promising. The insights gained from these patterns can inform preventive measures and help detect risk factors, ultimately leading to better treatments and outcomes for patients. The algorithm proposed in this thesis is able to successfully extract a set of meaningful and relevant temporal patterns, especially for the specific type of multimorbid patients considered. These patterns were evaluated by experts and demonstrated the ability to predict risk factors that are commonly associated with certain diseases. Moreover, the average time gap between the occurrence of medical events provided critical insight into the term of these risk factors. This information holds significant value in the context of primary healthcare and preventive medicine, highlighting the potential of this methodology to serve as a valuable medical tool.

In the modern working environment, people are overloaded with data. We collect, store, and process vast amounts of information for business, safety, and legal purposes. Managing this information manually is increasingly challenging, and employees spend significant time performing repetitive tasks. Consequently, automation offers a promising breakthrough that could substantially reduce costs. This project leverages data mining on email data, initially focusing on emails with the potential to extend findings to other areas. We aim to identify patterns within email interactions to automate repetitive tasks such as reading, responding, attaching files, forwarding information, and managing spam. This automation could significantly enhance productivity and user satisfaction by reducing the manual effort involved in email management. Our approach involves analyzing a large dataset of business emails to detect recurring interaction patterns. To reconstruct email chains and convert them into structured data, we use a systematic methodology relying on the emails' metadata and content. After automatically processing the texts, we generate embeddings to convert the text into numerical representations. A time-aware distance metric assesses sequence similarity to cluster the emails, revealing potential automation opportunities. The results demonstrate the feasibility of extracting processes and similar interactions from emails using the proposed solution. This serves as a model pipeline for future projects, where specific steps can be adapted to meet different task requirements, improve performance, and adapt to other data formats.

The main objective of this project is to use Large Language Models (LLM) to carry out the extraction of information from medical texts using learning techniques such as zero shot and few-shot learning. The information from these medical texts will come from the DDI corpus dataset, composed of the Drug Bank and MEDLINE sets, which consists of a semantically annotated corpus of documents describing interactions between drugs. From these data, we will extract information to identify the different pharmacological substances present in these texts using smaller-sized language models such as GPT2 and GPTNeo. With these models, an exploration of the different complexities available will also be conducted to study the computational resources required for predictions and the time needed for these tasks. To use these techniques and models, it is necessary to have GPUs and CPUs with a certain power and their specifications should remain consistent across different experiment reproductions to maintain reproducibility. Therefore, we will rely on the Google Colab platform, which provides a simple and intuitive environment for conducting our research.

This study focuses on named entity recognition (NER), an essential task in natural language processing (NLP). The main objective is to develop a customized solution for named entity recognition for the business sector of the insurance company. The work is divided into three parts: a theoretical framework that reviews the concepts and techniques used in the field, the design and implementation of the NER solution based on the GPT model, and a comparative analysis that evaluates the performance of the proposed solution in comparison with other tools and algorithms. The results of this study could have significant implications for the adaptation of these natural language processing techniques within the company and satisfactory results will be shown for both the model with Conditional Random Fields and one of the alternatives presented with GPT.

In this project, a tutor has been developed for students preparing for university entrance exams. Using the Transformers architecture, a GPT model has been trained to reliably answer questions about the university entrance exams, specifically those related to the announcements, schedules, and regulations. A BERT model has also been trained to detect the degree of similarity between two texts. This model would be used to conduct a similarity search, in case a student wants to get the answer to a problem, using a photograph or a text of the problem. The model is capable of searching the database for the most similar problem and returning its solution. All the trainings have been carried out on a local 24 GB graphics card. Finally, everything has been packaged into a web application programmed in Vue 3 and Flask.

In this project, the behavior of the routing algorithms existing in the Lightning Network, the second layer of Bitcoin payments, is studied, along with their analysis and comparison with other models proposed by the student. The report explains the context in which the project is developed, its scope, and everything related to its management, such as its temporal planning, work methodology, resource management, budget, and sustainability. An introduction with its basic concepts is also provided, detailing the design of the routing system proposed by the student and its subsequent implementation. Furthermore, the working environment is described, along with the design of the experiments we wanted to carry out, their implementation, and, finally, the results are analyzed by comparing the different models available through the data obtained by visualizing some of them in graphs. Finally, a comprehensive evaluation of the project and its results is made, and paths for further investigation are proposed.

Currently, many companies do not use the information contained in their emails, yet it is a data set that is full of information and could be very useful. This thesis report focuses on email data analysis and task automation, particularly in the area of email-based process mining. The state of the art section reviews existing research on extracting information from email content using techniques such as lexical analysis, language detection, semantic analysis and machine learning methods. It explores different areas of process mining, including process pattern discovery, anomaly discovery, and process extraction from texts. The objectives of this research are to assess the feasibility of extracting candidate processes from emails, to develop human-understandable metrics to classify processes, to propose a system to identify automation opportunities in email templates and explore possibilities for automation in email interactions. To do this, we carried out different steps such as data preparation, chains detection, text representation, distance matrix calculation and grouping methods.

myReach is a startup born in 2018 dedicated to offering a centralized and smart data storage and management system to the public, based on graph data storage. Its public mobile application needs a smart recommender of tags for the nodes that represent webpages. Currently, no publicly available tool designed for this specific task exists. Thus, the goal of this thesis is to implement an url tag categorizer that can give accurate recommendations to the user on the type of webpage of any saved url. Even though the end results are far from ideal, most of the work done can be reused with better datasets or techniques which could not be applied in this thesis.

The asset market has grown significantly in the last few years as well as the number of young people interested in investing and making easy money. This increase in demand, however, has not been met with an increase in formation or education, resulting in lots of avoidable mistakes and the loss of considerable amounts of money. A common approach on minimizing losses consists of using social media to look for the guidance of specialists or try to identify the latest trends. The use of artificial intelligence can be highly beneficial to automate the steps followed by this approach along with the use of metrics such as the number of mentions or the sentiment of the comments. The purpose of this project is to develop an app capable of obtaining and processing data from social media in real time, automating the process of identifying trends and promoting conscious investment against speculation. Besides the AI engine behind the app, this project also focuses on the efficiency and performance of the chosen database and the implementation of a webpage where users can see the processed data and generated metrics.

Finding automatization opportunities for email interactions can have positive effects for several industries, especially in tasks such as reading, receiving, writing and responding emails, categorizing emails or even to prevent loss of productivity and financial loses by dealing with spam, or improve users' satisfaction; even improving automatic categorization systems can mitigate negative impacts on personal and organization performance. Furthermore, people who work in companies spend around 28 % of their time reading and answering emails. In this project we proposed a methodology based on NLP and Unsupervised Machine Learning to look for opportunities of automation arising from recurrent email patterns found in email texts. We intent to facilitate the linguistic analysis in order to retrieve interaction patterns that can trigger automation actions. We proposed CRISP-DM methodology that lays the groundwork for detection of automatization opportunities in tasks relates. We compared the unsupervised machine learning methods K-Means, DBSCAN, and HDBSCAN with four clustering metrics applied to the Enron e-mails dataset transformed into paragraph vectors and performed several experiments with Word Mover's Distance, Euclidean Distance, L2-Norm and Cosine Similarity. Although our process yielded limited results in the detection of email interactions, we found that DBSCAN combined with Euclidean Distance was the best method among all scores. This project also contributes to the parameterization literature of said clustering algorithms as well as showing which methods, distances and scores settings are relevant for unsupervised email mining.

Currently, most business sit on rich depositories of data that they aren't utilizing to their fullest, among these, is most usually, business emails. This trove of data is rarely exploited, but it holds data that could enhance business insights, for this purpose, in this work we propose to extract candidate processes from these emails, that, with the aid of an analyst well-versed in the knowledge of their business, can turn into processes and business insights. The solution proposed is unsupervised learning with a modified instance of the Earth Mover's Distance via HDBScan, to allow for a computationally cheap yet robust enough method to obtain candidate processes and metrics from large amounts of emails. The results support the feasibility of extracting processes and business insights from emails with our proposed solution, though improvements should be pursued to extract better insights and remove the need of a human analyst.

Formally defining business processes in standards such as BPMN 2.0 is a crucial step in the digital transformation of organizations. Many organizations have their processes documented in text, and creating formal models has a high cost of specialized staff. The goal of the project is to create a set of annotated data that relate texts to the formal models that describe them, so that they can be used to train Deep Learning neural systems to do this task.

In the recent years, neural networks are very popular in the field of the artificial intelligence (AI). We decided to design the project on the basis of it. Neural network is the branch of the machine learning that uses the different layers to represent the data. Data are transformed to different or multi layers and generate the output. Our project is User ticketing system with automatic resolution suggestions, in which data is text based and with the help of Natural language processing (NLP) we can solve the problem. NLP is a sub-field of artificial intelligence which creates interactions between computers and human language. The goal is a computer capable of understanding the contents of documents, including the contextual nuances of the language within them. In this project the main goal is measuring the similarity between the texts. We design different models and use different layers and hyper-parameters later we will discuss in details. Our models are based on the Siamese neural network (SNN). SNN (sometimes called a twin neural network) is an artificial neural network that uses the same weights while working in tandem on two different input vectors to compute comparable output vectors. In this project, three main approaches have been used. Firstly, on the basis of the pre-trained model we get the output. Secondly, we follow the transfer learning approach. The transfer learning consists of taking features learned on one problem, and leveraging them on a new, similar problem. Thirdly, the model is fine tunned and designed on the basis of mutilingual (Spanish and English).