We present a novel contribution to Spanish clinical Natural Language Processing (NLP) by introducing the largest publicly available clinical corpus, ClinText-SP , along with a state-of-the-art clinical encoder language model. Our corpus was meticulously curated from diverse open sources, including clinical cases from medical journals and annotated corpora from shared tasks, providing a rich and diverse dataset that was previously difficult to access. Our model, developed through domain-adaptive pretraining on this comprehensive dataset, significantly outperforms existing models on multiple clinical NLP benchmarks. By publicly releasing both the dataset and the model, we aim to empower the research community with robust resources that can drive further advancements in clinical NLP and ultimately contribute to improved healthcare applications.

Data quality or evaluation can sometimes be a task as important as collecting a large volume of data when it comes to generating accurate artificial intelligence models. Being able to evaluate the data can lead to a larger database that is better suited to a particular problem because we have the ability to filter out data of dubious quality obtained automatically. In this paper, we present RLBoost (Anguiano et al., 2024) [1] , an algorithm that uses deep reinforcement learning strategies to evaluate a particular dataset and obtain a model capable of estimating the quality of any new data in order to improve the final predictive quality of a supervised learning model. This solution has the advantage of being agnostic regarding the supervised model used and, through multi-attention strategies, takes into account the data in its context and not only individually. The results of the article show that this model obtains better and more stable results than other state-of-the-art algorithms such as Leave One Out, Shapley or Data Valuation using Reinforcement Learning.

OBJECTIVES: This comparative analysis aims to assess the efficacy of encoder Language Models for clinical tasks in the Spanish language. The primary goal is to identify the most effective resources within this context. IMPORTANCE: This study highlights a critical gap in NLP resources for the Spanish language, particularly in the clinical sector. Given the vast number of Spanish speakers globally and the increasing reliance on electronic health records, developing effective Spanish language models is crucial for both clinical research and healthcare delivery. Our work underscores the urgent need for specialized encoder models in Spanish that can handle clinical data with high accuracy, thus paving the way for advancements in healthcare services and biomedical research for Spanish-speaking populations. MATERIALS AND METHODS: We examined 17 distinct corpora with a focus on clinical tasks. Our evaluation centered on Spanish Language Models and Spanish Clinical Language models (both encoder-based). To ascertain performance, we meticulously benchmarked these models across a curated subset of the corpora. This extensive study involved fine-tuning over 3000 models. RESULTS: Our analysis revealed that the best models are not clinical models, but general-purpose models. Also, the biggest models are not always the best ones. The best-performing model, RigoBERTa 2, obtained an average F1 score of 0.880 across all tasks. DISCUSSION: Our study demonstrates the advantages of dedicated encoder-based Spanish Clinical Language models over generative models. However, the scarcity of diverse corpora, mostly focused on NER tasks, underscores the need for further research. The limited availability of high-performing models emphasizes the urgency for development in this area. CONCLUSION: Through systematic evaluation, we identified the current landscape of encoder Language Models for clinical tasks in the Spanish language. While challenges remain, the availability of curated corpora and models offers a foundation for advancing Spanish Clinical Language models. Future efforts in refining these models are essential to elevate their effectiveness in clinical NLP.

We study TV regularization, a widely used technique for eliciting structured sparsity. In particular, we propose efficient algorithms for computing prox-operators for $\ell_p$-norm TV. The most important among these is $\ell_1$-norm TV, for whose prox-operator we present a new geometric analysis which unveils a hitherto unknown connection to taut-string methods. This connection turns out to be remarkably useful as it shows how our geometry guided implementation results in efficient weighted and unweighted 1D-TV solvers, surpassing state-of-the-art methods. Our 1D-TV solvers provide the backbone for building more complex (two or higher-dimensional) TV solvers within a modular proximal optimization approach. We review the literature for an array of methods exploiting this strategy, and illustrate the benefits of our modular design through extensive suite of experiments on (i) image denoising, (ii) image deconvolution, (iii) four variants of fused-lasso, and (iv) video denoising. To underscore our claims and permit easy reproducibility, we provide all the reviewed and our new TV solvers in an easy to use multi-threaded C++, Matlab and Python library.

In this work we will analyze and apply to the prediction of wind energy some of the best known regularized linear regression algorithms, such as Ordinary Least Squares, Ridge Regression and, particularly, Lasso, Group Lasso and Elastic-Net that also seek to impose a certain degree of sparseness on the final models. To achieve this goal, some of them introduce a non-differentiable regularization term that requires special techniques to solve the corresponding optimization problem that will yield the final model. Proximal Algorithms have been recently introduced precisely to handle this kind of optimization problems, and so we will briefly review how to apply them in regularized linear regression. Moreover, the proximal method FISTA will be used when applying the non-differentiable models to the problem of predicting the global wind energy production in Spain, using as inputs numerical weather forecasts for the entire Iberian peninsula. Our results show how some of the studied sparsity-inducing models are able to produce a coherent selection of features, attaining similar performance to a baseline model using expert information, while making use of less data features.

Sequential Minimal Optimization (SMO) can be regarded as the state-of-the-art approach in non-linear Support Vector Machines training, being the method of choice in the successful LIBSVM software. Its optimization procedure is based on updating only a couple of the problem coefficients per iteration, until convergence. In this paper we notice that this strategy can be interpreted as finding the sparsest yet most useful updating direction per iteration. We present a modification of SMO including a new approximate momentum term in the updating direction which captures information from previous updates, and show that this term presents a trade-off between sparsity and suitability of the chosen direction. We show how this novelty is able to provide substantial savings in practice in SMO's number of iterations to convergence, without increasing noticeably its cost per iteration. We study when this saving in iterates can result in a reduced SVM training times, and the behavior of this new technique when combined with caching and shrinking strategies.

Scaled Convex Hulls (SCHs) have been recently proposed by Liu et al. as the basis of a method to build linear classifiers that, when extended to kernel settings, provides an alternative approach to more established methods such as SVMs. Here we show how to adapt the Mitchell-Dem'yanov-Malozemov (MDM) algorithm to build such SCH-based classifiers by solving a concrete nearest point problem. We shall discuss two possible approaches to do so and show that they produce the same updates; we shall also prove that the resulting algorithm converges to the optimal solution. Moreover, our experiments also show that MDM's complexity is better than that of the SK method proposed in Liu's work. However, while the SCH classifiers often give competitive accuracies, further work is needed, particularly to obtain the scaling parameter #, to ensure good classifier accuracies for general problems.