Nowadays, colorectal cancer is one of the most common cancers, and early detection would greatly help improve patient survival. The current methods used by physicians to detect it are based on the visual detection of polyps in colonoscopy, a task that can be tackled by means of semantic segmentation methods. However, the amount of data necessary to train deep learning models for these problems is a barrier for their adoption. In this work, we study the application of different semi-supervised learning techniques to this problem when we have a small amount of annotated data. In this study, we have used the Kvasir-SEG data set, taking only 60 and 120 annotated images and studying the behaviour of the Data Distillation, Model Distillation, and Data & Model distillation methods in both cases, using 10 different architectures. The results show that as we increase the number of initially annotated data, most models obtained better results, but two of them performed worse in the baseline case. Furthermore, we can conclude that the Data Distillation method increases the performance of the models a 48.6% and 30.6% on average using 60 and 120 annotated images respectively. Finally, using only 12% of the annotated data and applying Data Distillation, the results obtained are not very far from those obtained by training the models with the fully annotated dataset. For all these reasons, we conclude that the Data Distillation method is a good tool in semantic segmentation problems when the number of initially annotated images is small.
This study explores prodromal Parkinson’s Disease (PD) by leveraging data from the Parkinson’s Progression Markers Initiative (PPMI). The main goal was to discriminate between prodromals that phenoconverted to PD in 7 years to those that did not. Through feature selection, the system identified key first visit predictors of PD phenoconversion, encompassing demographic, clinical, and structural magnetic resonance imaging (MRI) data. Employing seven machine learning algorithms in standard and balanced forms, we find Support Vector Machine (balanced) as most effective for demographic and clinical data, and Logistic Regression (balanced) when adding thicknesses and volumes of MRI data. The metrics were improve in the second case (AUC ROC of 0.84). Significant predictors include olfactory dysfunction, motor symptoms, psychomotor speed, and third ventricle dilation.
This paper contributes with an alternative to the multivariate Analogue Method (AM) version, using a preprocessing stage carried out by an Autoencoder (AE). The proposed method (MvAE-AM) is applied to reconstruct France’s 2003, Balkans’ 2007 and Russia 2010 mega heat waves. Using divers such as geopotential height of the 500hPA (Z500), mean sea level pressure (MSL), soil moisture (SM), and potential evaporation (PEva), the AE extracts the most relevant information into a smaller univariate latent space. Then, the classic univariate AM is applied to search for similar situations in the past over the latent space, with a minimum distance to the heat wave under evaluation. We have compared the proposed method’s performance with that of a classical multivariate AM (MvAM), showing that the MvAE-AM approach outperforms the MvAM in terms of accuracy ( + 1.1257 C), while reducing the problem’s dimensionality.
E-commerce has become an essential aspect of modern life, providing consumers globally with convenience and accessibility. However, the high volume of short and noisy product descriptions in text streams of massive e-commerce platforms translates into an increased number of clusters, presenting challenges for standard model-based stream clustering algorithms. Standard LDA-based methods often lead to clusters dominated by single elements, effectively failing to manage datasets with varied cluster sizes. Our proposed Community-Based Topic Modeling with Contextual Outlier Handling (CB-TMCOH) algorithm introduces an approach to outlier detection in text data using transformer models for similarity calculations and graph-based clustering. This method efficiently separates outliers and improves clustering in large text datasets, demonstrating its utility not only in e-commerce applications but also proving effective for news and tweets datasets.
Embryo selection is an indispensable step to ensure the success of in vitro fertilization. There are two techniques to perform embryo selection: preimplantation genetic screening and embryo morphological grading. However, even with these techniques, the embryo implantation probability is barely 65% making extremely difficult to evaluate their implantation potential. This is mainly due to the lack of markers, and the subjectivity associated with experience, judgment, and training of the embryologists. Computer vision and deep learning methods can help to automatically identify those markers with methods such as the segmentation of the embryo structures to offer detailed, quantitative, and objective assessments; and with that, information to predict the pregnancy outcome of embryos. In this paper, we present different methods capable of segmenting the components of an embryo (namely, the Trophectoderm, the Inner Cell Mass and the Zona Pellucida) with Dice scores ranging from 0.85 to 0.89, and openly release the code so that anyone can use it and replicate the results. These models are a first step towards a more objective evaluation of the embryos’ implantation potential.
Suicide is a major health and social issue worldwide; therefore, a simple access to reliable sources of information that can be used by family members or friends of people who have suicidal ideation can be a valuable resource. This information can be provided by means of chatbot tools; however, the reliability and topicality of the chatbot’s answers should be ensured. In this work, we present an architecture to build a chatbot with the aim of providing reliable suicide information in Spanish. The architecture consists of two text classification models (one to check that a user’s question is related to suicidal content, and another to decide whether the user is looking for information or if the question should be derived to a human), and a retrieval augmented generation system that, using as a basis a corpus of documents filtered by experts, generates an answer to the user question. In addition, all the components of the architecture have been automatically tested to prove their suitability to be incorporated to the chatbot. The developed system is a step towards helping in one of the greatest global public health concerns.
Machine learning techniques have recently transformed the way we analyze competitive games. However, accurately detecting the impact of different insights on match outcomes remains a challenge. This study focuses on League of Legends, a popular multiplayer online battle arena game known for its strategic depth and teamwork requirements. We aim to understand how various actions and strategies influence match results, using a dataset from professional tournaments. Factors like “building damage”, “total gold”, and “assists” are analyzed as predictors. We employ tree-based and linear models to predict outcomes, supplemented by SHapley Additive exPlanations for explaining both local and global model outcomes. Our article offers a generalizable match analysis approach, compares explainable methods, and delves into key determinants of victory. The results, showcasing a remarkable 98.8% accuracy with the top-performing model, provide strong support for our conclusions, underlining their reliability.
The notion of bond in formal concept analysis arose as a mechanism for aggregating contexts, preserving the main information of the original ones. This notion can also be fundamental in the inverse process, that is, in the factorization of contexts, which will allow the computation of the information of a real context from smaller subcontexts (distributed computing). This paper considers the flexible fuzzy multi-adjoint framework in order to introduce the first definition of bond in this setting and presents the first properties and examples of this definition.
This comprehensive review explores the rapidly advancing field of Spiking Neural Networks (SNNs), particularly emphasizing their computational capabilities and potential for energy-efficient computing. SNNs distinguish themselves from traditional neural networks by skillfully processing complex, time-sensitive binary inputs through intricate encoding strategies and dynamic learning algorithms. This paper discusses various encoding techniques and evaluates several neuron models integral to SNN architecture, such as the Leaky Integrate-and-Fire, Hodgkin-Huxley, and Izhikevich models. These models are appraised for their trade-offs between computational simplicity and biological plausibility. Additionally, we examine the energy-saving expertise of SNNs relative to their traditional counterparts, identifying challenges in scaling and the intricacy of training. The review explores a spectrum of training techniques for SNNs, including supervised, unsupervised, and reinforcement learning approaches. This paper culminates by highlighting imperative future research directions in SNNs. It underscores the pressing need for developing sophisticated training algorithms and customizing models to augment efficiency and versatility in energy-conscious computing. These focal points are suggested as pivotal for driving the field forward and unlocking the full potential of SNNs in real-world applications.
Time Series Ordinal Classification (TSOC) is a yet unexplored field with a substantial projection in following years given its applicability to numerous real-world problems and the possibility to obtain more consistent prediction than nominal Time Series Classification (TSC). Specifically, TSOC involves time series data along with an ordinal categorical output. That is, there is a natural order relationship among the labels associated with the time series. TSOC is a subfield of nominal TSC, with the main distinction being that TSOC exploits the ordinality of the labels to boost the performance. Two categories within the TSC taxonomy are dictionary-based and convolution-based methodologies, each representing competing approaches presented in the literature. In this study, we adapt the Hybrid Dictionary-Rocket Architecture (Hydra) approach, which incorporates elements from the two previous categories, to TSOC, resulting in O-Hydra. For the experiments, we have included a collection of 21 ordinal problems sourced from two well-known archives. O-Hydra has been benchmarked against its nominal counterpart, Hydra, as well as against two state-of-the-art approaches in the two previous categories, TDE and ROCKET, including their ordinal counterparts, O-TDE and O-ROCKET, respectively. The results achieved by the ordinal versions significantly outperformed those of current nominal TSC techniques. This underscores the significance of incorporating the label ordering when addressing such problems.
The partial consolidated tree bagging (PCTBagging) was presented as a multiple classifier that, based on a parameter, the consolidation percentage, can exploit more the possibilities of the inner ensembles, and obtain higher levels of interpretability, or can exploit more the possibilities of the ensembles, and obtain higher discriminant capacity. Thus, at the extreme values, with a consolidation percentage of 100 × Size, and, Level by level. The results show that the use of different criteria affects the discriminant capacity of the classifier for the same level of interpretability, and that this effect is greater the higher the percentage of consolidation is.
The use of Neural Networks (NN) within Combinatorial Optimization (CO) marks a significant shift in the paradigm, moving towards automatically learning heuristic strategies in deterministic and local search frameworks. NNs are capable of learning relevant patterns and symmetries of various CO problems. Despite their potential, the practical application of NNs in both academic and real-world optimization problems has not yet reached the levels of traditional exact solvers or metaheuristic approaches. This study primarily focuses on the Maximum Cut problem to investigate the capabilities and limitations of NN models within the CO domain. We introduce a series of research questions aimed at examining the generalization capabilities, reliability, and computational costs associated with these models. Our findings reveal that: (1) NN models exhibit better modeling capabilities and generalizability when trained on a diverse set of instances, (2) the model’s level of uncertainty can act as an indicator of its performance, and (3) employing a unified representation framework, wherein models concurrently learn from diverse tasks or instance types offers a significant training-speedup.
This article explores the use of Large Language Models (LLM) as transformative tools for teaching Artificial Intelligence (AI) and Robotics concepts at the master’s level. LLMs, exemplified by models like ChatGPT, present a unique opportunity to revolutionize the pedagogical landscape by offering advanced capabilities in any service robot. The study investigates the integration of LLMs in the instructional framework, through the llama_ros tool, capable of replacing different classic cognitive functions in a transversal project across different subjects of an official master’s degree. The research presents as an example the creation of an LLM-based chatbot on an open hardware platform called Mini Pupper. The reader will find how to emphasize the potential of LLMs to shape their inclusion in bachelor’s or master’s programs.
The Bi-Objective Double Floor Corridor Allocation Problem is one of the most recent incorporation to the family of Facility Layout Problems. This problem, which has been a challenge for exact and metaheuristic approaches, involves optimizing the layout of the given facilities to minimize material handling cost and the length of the corridor considering more than one floor. This paper introduces a new approach based on the combination of two greedy methods and a path relinking implementation to tackle this problem. The experimental results show the superiority of our proposal in relation to the current state-of-the-art under different multi-objective metrics.
One of the crucial aspects of solving robust optimization over time (ROOT) problems is to efficiently approximate the robustness of the solutions. However, current progress in this area has been scarce to date. To help bridge this gap, this paper proposes an alternative approach to one of the predominant frameworks in this field. Specifically, we decouple the fit and prediction of future environments that occur for each fitness evaluation by just evaluating previously fitted surrogate models. In this way, we globally approximate the robustness of the solutions by learning fitness functions, rather than point-wise predicting values during the execution of the algorithm. Preliminary results obtained from computational experiments indicate that this approach can achieve significantly superior performances to the existing framework, especially for specific surrogate model configurations. Furthermore, we show that in certain cases where our algorithms are less efficient than the existing approach, such inefficiency is compensated by improvements in error.
The exceptional performance of Convolutional Neural Networks (CNNs) entails increasing requirements in computing power and storage. While several efficient compression methods have been developed, there is no consideration on which features are removed or preserved, which can affect pruning. In this paper, we propose a novel filter pruning strategy, named Layer Factor Analysis one to one (LFA1-1), that, relying on explainability, selects the filters that best retain the essential features underlying convolutional layers. We provide insights about the relevance of preserving these features and verify its relationship with compressed network’s performance. The explanatory analysis carried out allows us to justify pruning efficiency and detect problematic parts. Experiments with VGG-16 on CIFAR-10 are conducted in order to validate our approach. Quantitative and qualitative comparisons with methods in the literature uncover pruning properties and prove the effectiveness of our proposal, which reaches a 89.1% parameters and 83.8% FLOPs reduction with the lowest accuracy drop.
This chapter is devoted to the problems of personalized medicine using mass monitoring of the population based on electrocardiograms with a reduced number of leads. Artificial intelligence is used by the authors in the basic sense of the term. The algorithms for segmenting the elements of the cardiocycle are developed on the principles of the bionic approach, namely, on the basis of already existing knowledge about the features of the human visual system. Variable resolution segmentation algorithms, which are described in the chapter, have a number of advantages over classical methods for processing cyclic signals, as well as binary image contours. The authors believe that the proposed approaches will be useful in other subject areas where information technologies are used to process visual information. Then, the results of various experimental studies in the field of occupational medicine are presented in the chapter. The conducted studies confirm the possibility of estimating the physiological cost of a military officer's and medical doctor's professional activity using portable electrocardiogram (ECG) devices. Monitoring the functional and psychoemotional state of servicemen will allow to objectively determine the individual operational readiness of each serviceman to perform tasks, prevent a critical exceedance of his physiological and functional capabilities, as well as predict readiness to participate in long-term combat operations. Wearable wireless superminiature electrocardiographic devices in combination with modern analytical platforms potentially provide an opportunity to objectively monitor important physiological parameters directly during the performance of duties of various nature. The proposed methods and means can be considered as the main tool to support the decision-making of the commander regarding the ability of personnel from the point of view of their functional state to perform combat tasks.
Obstructive sleep apnea (OSA), a prevalent nocturnal breathing disorder, affects more than a billion people overall. OSA causes frequent disruptions in breathing due to the collapse of the upper airway, resulting in sleep fragmentation and it eventually leads to several abnormalities such as asphyxia, fatigue, cardiac arrhythmias, hypertension, and many other cardiovascular disorders. This can turn into a chronic disorder and even result in the death of the person if neglected, misdiagnosed, or undertreated. Hence accurate, early, and patient-friendly diagnosis is necessary. Polysomnography is the conventional method for OSA diagnosis that involves recording of different physiological signals during sleep. This is a very costly and inconvenient procedure. Disadvantages of this standard procedure through led to a research interest in diagnosing OSA using lesser signals like ECG, EEG, and SpO2 signals, which are impacted because of OSA. Of all the signals, it is observed that ECG has potential changes due to OSA. Hence, OSA diagnosis using ECG signals became a research interest and moreover to improve the accuracy and diagnose early, automated diagnosis using machine learning and deep learning methods became popular in recent times. This chapter includes a brief introduction to OSA, its pathogenesis, pathophysiology, and recent advances in automated OSA diagnosis through single-lead ECG signals using ML and DL methods.
Growing incidence, prevalence, and risk make atrial fibrillation (AF), a significant arrhythmia in the cohort with highest mortality, morbidity, and healthcare expenses. It is because its early stages largely remain undetected due to its abrupt, rapid period of incidence. Atria tends to fibrillate fast and irregularly during AF at approximately 400 beats/minute. AF manifests itself due to various factors, which can be modified and nonmodified. Daily monitoring of modifiable factors, such as body mass index, blood sugar level, blood pressure level, and lifestyle, may be regarded as the best therapy for AF risk modulation, decreasing mortality, and to prevent healthcare expenditures. Primary and secondary measures for AF prevention comprise individual, healthcare, and social interventions. Recent advances in eHealth devices and applications, and artificial intelligence (AI) research will be indispensable for refining AF risk predictors and communication, which will contribute to the AF prevention and management.
