The majority of clinical information stored in Spanish healthcare systems is found as unstructured text in electronic health records (EHRs). The automatic extraction of valuable information contained in these documents is a critical task. Valuable information for oncology clinical analysis units and Real-World Evidence studies includes the location of the neoplasm presented by a patient. This location, included in the ICD-10 coding category, can be extracted from the texts by natural language processing (NLP). This study set out to explore the classification of medical documents in Spanish for the purpose of extracting the location of the patient's primary neoplasm. A private corpus composed of 23,704 real clinical EHRs was utilised. The prediction problem was approached through a classification of 12 primary organ groupings and 29 specific locations. In order to achieve this, four NLP methodologies were developed: traditional machine learning (ML); ensemble ML; recurrent neural networks (RNN); and Transformers-based models. Our findings demonstrate that traditional ML models exhibit superior performance when compared to RNNs and Transformers. Models such as XGBoost and SVMs demonstrate remarkable efficacy, attaining an F1-score of 0.938 for the 12-class classification and an F1-score of 0.838 for the 29-class specific classification, respectively. The pre-trained RoBERTa-Base-Biomed model, which incorporates medical and clinical corpora in Spanish, demonstrates an F1-score of 0.808 for the 29-location problem. However, the Transformer model exhibits superior performance than the rest of the approaches when dealing with an external corpus, indicating a higher generalisation capacity.

BACKGROUND AND OBJECTIVE: Automatic clinical coding is a crucial task in the process of extracting relevant information from unstructured medical documents contained in Electronic Health Records (EHR). However, most of the existing computer-based methods for clinical coding act as "black boxes", without giving a detailed description of the reasons for the clinical-coding assignments, which greatly limits their applicability to real-world medical scenarios. The objective of this study is to use transformer-based models to effectively tackle explainable clinical-coding. In this way, we require the models to perform the assignments of clinical codes to medical cases, but also to provide the reference in the text that justifies each coding assignment. METHODS: We examine the performance of 3 transformer-based architectures on 3 different explainable clinical-coding tasks. For each transformer, we compare the performance of the original general-domain version with an in-domain version of the model adapted to the specificities of the medical domain. We address the explainable clinical-coding problem as a dual medical named entity recognition (MER) and medical named entity normalization (MEN) task. For this purpose, we have developed two different approaches, namely a multi-task and a hierarchical-task strategy. RESULTS: For each analyzed transformer, the clinical-domain version significantly outperforms the corresponding general domain model across the 3 explainable clinical-coding tasks analyzed in this study. Furthermore, the hierarchical-task approach yields a significantly superior performance than the multi-task strategy. Specifically, the combination of the hierarchical-task strategy with an ensemble approach leveraging the predictive capabilities of the 3 distinct clinical-domain transformers, yields the best obtained results, with f1-score, precision and recall of 0.852, 0.847 and 0.849 on the Cantemist-Norm task and 0.718, 0.566 and 0.633 on the CodiEsp-X task, respectively. CONCLUSIONS: By separately addressing the MER and MEN tasks, as well as by following a context-aware text-classification approach to tackle the MEN task, the hierarchical-task approach effectively reduces the intrinsic complexity of explainable clinical-coding, leading the transformers to establish new SOTA performances for the predictive tasks considered in this study. In addition, the proposed methodology has the potential to be applied to other clinical tasks that require both the recognition and normalization of medical entities.

After being synthesized by ribosomes in the cells, proteins can suffer from post-translational modifications (PTM) that affect their functionality. One of the most studied PTMs is phosphorylation. Mass-spectrometry methods aimed at identifying phosphorylation sites in proteins are arduous and expensive. For these reasons, numerous studies propose the use of machine leaning techniques to predict this PTM. Like phosphorylation, methionine oxidation is another important PTM. Recently, we have proposed a machine learning approach that extracts a set of features from the primary and tertiary structure of the proteins to predict methionine oxidation sites. However, this work had an important limitation that impairs feature extraction, since the 3D structure of many proteins is not fully resolved. In this study, we present MetODeep, a deep learning approach to predict methionine oxidation. Unlike phosphorylation, for which datasets with several hundred thousands samples are available to train effective predictive models, methionine oxidation counts on small datasets, which could lead a deep neural network to experiment over-fitting issues. The recently evidenced existence of a cross-talk between phosphorylation and methionine oxidation, has motivated our transfer-learning approach. Thus, on the basis of a deep convolutional neural network (CNN) pre-trained with phosphorylation data, MetODeep is fine-tuned to predict methionine oxidation. The resulting CNN architecture allows us to omit manual feature extraction, since it accepts raw protein sequences as input data. The final model gives performance results (AUC 0.8267±0.0174) that surpass state-of-art of computational models for the prediction of methionine oxidation.

Deep learning provides a variety of neural network based models, known as Deep Neural Networks (DNNs), which are being successfully used in several domains to build highly accurate predictors from data. In particular, the predictive performance of a dense fully-connected multi-layer neural networks may vary depending on some factors. In this paper, 18 synthetic datasets were used to test the effect of data dimension and data structure on the predictive performance of a standard DNN and an architecture-constrained DNN (c-DNN) based on problem-specific information. The results of the analysis showed that a c-DNN clearly outperforms a standard DNN in most of the cases considered. Moreover, it suggested that both adding constraints to the network architecture and having the lowest number of input features possible which are relevant to the problem addressed may have a positive impact in terms of reducing overfitting and getting better prediction results.

Clinicians are interested in the estimation of robust and relevant genetic signatures from gene sequencing data. Many machine learning approaches have been proposed trying to address well-known issues of this complex task (feature or gene selection, classification or model selection, and prediction assessment). Addressing this problem often requires a deep knowledge of these methods and some of them demand high computational resources that may not be affordable. In this paper, an exhaustive study that includes different types of feature selection methods and classifiers is presented, providing clinicians an useful insight of the most suitable methods for this purpose. Predictions assessment is performed using a bootstrap cross-validation strategy as an honest validation scheme. The results of this study for six benchmark datasets show that filter or embedded methods are preferred, in general, to wrapper methods according to their better statistical significant results, in terms of accuracy, and lower demand for computational resources.

The well-known backpropagation learning algorithm is implemented in a field-programmable gate array (FPGA) board and a microcontroller, focusing in obtaining efficient implementations in terms of a resource usage and computational speed. The algorithm was implemented in both cases using a training/validation/testing scheme in order to avoid overfitting problems. For the case of the FPGA implementation, a new neuron representation that reduces drastically the resource usage was introduced by combining the input and first hidden layer units in a single module. Further, a time-division multiplexing scheme was implemented for carrying out product computations taking advantage of the built-in digital signal processor cores. In both implementations, the floating-point data type representation normally used in a personal computer (PC) has been changed to a more efficient one based on a fixed-point scheme, reducing system memory variable usage and leading to an increase in computation speed. The results show that the modifications proposed produced a clear increase in computation speed in comparison with the standard PC-based implementation, demonstrating the usefulness of the intrinsic parallelism of FPGAs in neurocomputational tasks and the suitability of both implementations of the algorithm for its application to the real world problems.

A two-dimensional Ising model with nearest-neighbors ferromagnetic interactions is implemented in a Field Programmable Gate Array (FPGA) board. Extensive Monte Carlo simulations were carried out using an efficient hardware representation of individual spins and a combined global-local LFSR random number generator. Consistent results regarding the descriptive properties of magnetic systems, like energy, magnetization and susceptibility are obtained while a speed-up factor of approximately six times is achieved in comparison to previous FPGA-based published works and almost$10^4$times in comparison to a standard CPU simulation. A detailed description of the logic design used is given together with a careful analysis of the quality of the random number generator used. The obtained results confirm the potential of FPGAs for analyzing the statistical mechanics of magnetic systems.

Competitive majority network trained by error correction (C-Mantec), a recently proposed constructive neural network algorithm that generates very compact architectures with good generalization capabilities, is implemented in a field programmable gate array (FPGA). A clear difference with most of the existing neural network implementations (most of them based on the use of the backpropagation algorithm) is that the C-Mantec automatically generates an adequate neural architecture while the training of the data is performed. All the steps involved in the implementation, including the on-chip learning phase, are fully described and a deep analysis of the results is carried on using the two sets of benchmark problems. The results show a clear increase in the computation speed in comparison to the standard personal computer (PC)-based implementation, demonstrating the usefulness of the intrinsic parallelism of FPGAs in the neurocomputational tasks and the suitability of the hardware version of the C-Mantec algorithm for its application to real-world problems.

A new constructive approach for creating all linearly separable functions is introduced. Balanced and unate functions on N+1 variables are created and then projected to N variables permitting to create straightforward all the linearly separable functions without needing to check for linearly separability. The method is supported by the demonstration of a theorem and numerical simulation results for small number of variables. If the results extrapolates the method may permit to test the linear separability of any function on N variables by checking monotonicity and unaleness in N+1 dimension. Furthermore, the generalization complexity of the linear threshold functions is analyzed.