Explainable clinical coding with in-domain adapted transformers.

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.

Transfer learning with convolutional neural networks for cancer survival prediction using gene-expression data.

Precision medicine in oncology aims at obtaining data from heterogeneous sources to have a precise estimation of a given patient's state and prognosis. With the purpose of advancing to personalized medicine framework, accurate diagnoses allow prescription of more effective treatments adapted to the specificities of each individual case. In the last years, next-generation sequencing has impelled cancer research by providing physicians with an overwhelming amount of gene-expression data from RNA-seq high-throughput platforms. In this scenario, data mining and machine learning techniques have widely contribute to gene-expression data analysis by supplying computational models to supporting decision-making on real-world data. Nevertheless, existing public gene-expression databases are characterized by the unfavorable imbalance between the huge number of genes (in the order of tenths of thousands) and the small number of samples (in the order of a few hundreds) available. Despite diverse feature selection and extraction strategies have been traditionally applied to surpass derived over-fitting issues, the efficacy of standard machine learning pipelines is far from being satisfactory for the prediction of relevant clinical outcomes like follow-up end-points or patient's survival. Using the public Pan-Cancer dataset, in this study we pre-train convolutional neural network architectures for survival prediction on a subset composed of thousands of gene-expression samples from thirty-one tumor types. The resulting architectures are subsequently fine-tuned to predict lung cancer progression-free interval. The application of convolutional networks to gene-expression data has many limitations, derived from the unstructured nature of these data. In this work we propose a methodology to rearrange RNA-seq data by transforming RNA-seq samples into gene-expression images, from which convolutional networks can extract high-level features. As an additional objective, we investigate whether leveraging the information extracted from other tumor-type samples contributes to the extraction of high-level features that improve lung cancer progression prediction, compared to other machine learning approaches.

BLASSO: integration of biological knowledge into a regularized linear model.

BACKGROUND: In RNA-Seq gene expression analysis, a genetic signature or biomarker is defined as a subset of genes that is probably involved in a given complex human trait and usually provide predictive capabilities for that trait. The discovery of new genetic signatures is challenging, as it entails the analysis of complex-nature information encoded at gene level. Moreover, biomarkers selection becomes unstable, since high correlation among the thousands of genes included in each sample usually exists, thus obtaining very low overlapping rates between the genetic signatures proposed by different authors. In this sense, this paper proposes BLASSO, a simple and highly interpretable linear model with l1-regularization that incorporates prior biological knowledge to the prediction of breast cancer outcomes. Two different approaches to integrate biological knowledge in BLASSO, Gene-specific and Gene-disease, are proposed to test their predictive performance and biomarker stability on a public RNA-Seq gene expression dataset for breast cancer. The relevance of the genetic signature for the model is inspected by a functional analysis. RESULTS: BLASSO has been compared with a baseline LASSO model. Using 10-fold cross-validation with 100 repetitions for models' assessment, average AUC values of 0.7 and 0.69 were obtained for the Gene-specific and the Gene-disease approaches, respectively. These efficacy rates outperform the average AUC of 0.65 obtained with the LASSO. With respect to the stability of the genetic signatures found, BLASSO outperformed the baseline model in terms of the robustness index (RI). The Gene-specific approach gave RI of 0.15±0.03, compared to RI of 0.09±0.03 given by LASSO, thus being 66% times more robust. The functional analysis performed to the genetic signature obtained with the Gene-disease approach showed a significant presence of genes related with cancer, as well as one gene (IFNK) and one pseudogene (PCNAP1) which a priori had not been described to be related with cancer. CONCLUSIONS: BLASSO has been shown as a good choice both in terms of predictive efficacy and biomarker stability, when compared to other similar approaches. Further functional analyses of the genetic signatures obtained with BLASSO has not only revealed genes with important roles in cancer, but also genes that should play an unknown or collateral role in the studied disease.

A machine learning approach for predicting methionine oxidation sites.

BACKGROUND: The oxidation of protein-bound methionine to form methionine sulfoxide, has traditionally been regarded as an oxidative damage. However, recent evidences support the view of this reversible reaction as a regulatory post-translational modification. The perception that methionine sulfoxidation may provide a mechanism to the redox regulation of a wide range of cellular processes, has stimulated some proteomic studies. However, these experimental approaches are expensive and time-consuming. Therefore, computational methods designed to predict methionine oxidation sites are an attractive alternative. As a first approach to this matter, we have developed models based on random forests, support vector machines and neural networks, aimed at accurate prediction of sites of methionine oxidation. RESULTS: Starting from published proteomic data regarding oxidized methionines, we created a hand-curated dataset formed by 113 unique polypeptides of known structure, containing 975 methionyl residues, 122 of which were oxidation-prone (positive dataset) and 853 were oxidation-resistant (negative dataset). We use a machine learning approach to generate predictive models from these datasets. Among the multiple features used in the classification task, some of them contributed substantially to the performance of the predictive models. Thus, (i) the solvent accessible area of the methionine residue, (ii) the number of residues between the analyzed methionine and the next methionine found towards the N-terminus and (iii) the spatial distance between the atom of sulfur from the analyzed methionine and the closest aromatic residue, were among the most relevant features. Compared to the other classifiers we also evaluated, random forests provided the best performance, with accuracy, sensitivity and specificity of 0.7468±0.0567, 0.6817±0.0982 and 0.7557±0.0721, respectively (mean ± standard deviation). CONCLUSIONS: We present the first predictive models aimed to computationally detect methionine sites that may become oxidized in vivo in response to oxidative signals. These models provide insights into the structural context in which a methionine residue become either oxidation-resistant or oxidation-prone. Furthermore, these models should be useful in prioritizing methinonyl residues for further studies to determine their potential as regulatory post-translational modification sites.

Methionine residues around phosphorylation sites are preferentially oxidized in vivo under stress conditions.

Protein phosphorylation is one of the most prevalent and well-understood protein modifications. Oxidation of protein-bound methionine, which has been traditionally perceived as an inevitable damage derived from oxidative stress, is now emerging as another modification capable of regulating protein activity during stress conditions. However, the mechanism coupling oxidative signals to changes in protein function remains unknown. An appealing hypothesis is that methionine oxidation might serve as a rheostat to control phosphorylation. To investigate this potential crosstalk between phosphorylation and methionine oxidation, we have addressed the co-occurrence of these two types of modifications within the human proteome. Here, we show that nearly all (98%) proteins containing oxidized methionine were also phosphoproteins. Furthermore, phosphorylation sites were much closer to oxidized methionines when compared to non-oxidized methionines. This proximity between modification sites cannot be accounted for by their co-localization within unstructured clusters because it was faithfully reproduced in a smaller sample of structured proteins. We also provide evidence that the oxidation of methionine located within phosphorylation motifs is a highly selective process among stress-related proteins, which supports the hypothesis of crosstalk between methionine oxidation and phosphorylation as part of the cellular defence against oxidative stress.

Sulphur Atoms from Methionines Interacting with Aromatic Residues Are Less Prone to Oxidation.

Methionine residues exhibit different degrees of susceptibility to oxidation. Although solvent accessibility is a relevant factor, oxidation at particular sites cannot be unequivocally explained by accessibility alone. To explore other possible structural determinants, we assembled different sets of oxidation-sensitive and oxidation-resistant methionines contained in human proteins. Comparisons of the proteins containing oxidized methionines with all proteins in the human proteome led to the conclusion that the former exhibit a significantly higher mean value of methionine content than the latter. Within a given protein, an examination of the sequence surrounding the non-oxidized methionine revealed a preference for neighbouring tyrosine and tryptophan residues, but not for phenylalanine residues. However, because the interaction between sulphur atoms and aromatic residues has been reported to be important for the stabilization of protein structure, we carried out an analysis of the spatial interatomic distances between methionines and aromatic residues, including phenylalanine. The results of these analyses uncovered a new determinant for methionine oxidation: the S-aromatic motif, which decreases the reactivity of the involved sulphur towards oxidants.

Efficient detection of wound-bed and peripheral skin with statistical colour models.

A pressure ulcer is a clinical pathology of localised damage to the skin and underlying tissue caused by pressure, shear or friction. Reliable diagnosis supported by precise wound evaluation is crucial in order to success on treatment decisions. This paper presents a computer-vision approach to wound-area detection based on statistical colour models. Starting with a training set consisting of 113 real wound images, colour histogram models are created for four different tissue types. Back-projections of colour pixels on those histogram models are used, from a Bayesian perspective, to get an estimate of the posterior probability of a pixel to belong to any of those tissue classes. Performance measures obtained from contingency tables based on a gold standard of segmented images supplied by experts have been used for model selection. The resulting fitted model has been validated on a training set consisting of 322 wound images manually segmented and labelled by expert clinicians. The final fitted segmentation model shows robustness and gives high mean performance rates [(AUC: .9426 (SD .0563); accuracy: .8777 (SD .0799); F-score: 0.7389 (SD .1550); Cohen's kappa: .6585 (SD .1787)] when segmenting significant wound areas that include healing tissues.

A web-based e-learning application for wound diagnosis and treatment.

Pressure ulcers (PrU) are considered as one of the most challenging problems that Nursing professionals have to deal with in their daily practice. Nowadays, the education on PrUs is mainly based on traditional lecturing, seminars and face-to-face instruction, sometimes with the support of photographs of wounds being used as teaching material. This traditional educational methodology suffers from some important limitations, which could affect the efficacy of the learning process. This current study has been designed to introduce information and communication technologies (ICT) in the education on PrU for undergraduate students, with the main objective of evaluating the advantages an disadvantages of using ICT, by comparing the learning results obtained from using an e-learning tool with those from a traditional teaching methodology. In order to meet this major objective, a web-based learning system named ePULab has been designed and developed as an adaptive e-learning tool for the autonomous acquisition of knowledge on PrU evaluation. This innovative system has been validated by means of a randomized controlled trial that compares its learning efficacy with that from a control group receiving a traditional face-to-face instruction. Students using ePULab gave significantly better (p<0.01) learning acquisition scores (from pre-test mean 8.27 (SD 1.39) to post-test mean 15.83 (SD 2.52)) than those following traditional lecture-style classes (from pre-test mean 8.23 (SD 1.23) to post-test mean 11.6 (SD 2.52)). In this article, the ePULab software is described in detail and the results from that experimental educational validation study are also presented and analyzed.

Effectiveness of an e-learning tool for education on pressure ulcer evaluation.

AIMS AND OBJECTIVES: To evaluate the effectiveness of information and communication technologies in the undergraduate students' pressure ulcer training as a learning tool, compared with traditional teaching methods. BACKGROUND: Pressure ulcers constitute one of the great challenges faced by nursing professionals. Currently, pressure ulcer training is based on traditional on-campus teaching, involving lecture-style classes with frequent use of photographs of the wounds. This traditional training has some important weaknesses that can put the efficacy of the training at risk. DESIGN: A randomised controlled trial was developed including undergraduate nursing students. METHODS: The intervention group used an adaptive self-learning e-learning tool developed by the research team (ePULab) for pressure ulcer assessment and treatment. The control group received a traditional on-campus class on the same topic. Pretest and post-test questionnaires were designed to assess the students' ability in pressure ulcer diagnosis and treatment. RESULTS: The educational intervention based on the use of the ePULab tool produced significantly better learning acquisition results than those obtained by traditional lecture-style classes: the total score improved in the control group from 8·23 (SD 1·23)-11·6 (SD 2·52) after the lecture, whereas in the intervention group, the knowledge score changed from 8·27 (SD 1·39)-15·83 (SD 2·52) (p = 0·01) with the use of ePULab. CONCLUSIONS: The results show a higher effectiveness of the devised e-learning approach for education on management of pressure ulcers. RELEVANCE TO CLINICAL PRACTICE: Our results reveal the suitability of the ePULab e-learning tool as an effective instrument for training on assessment of and treatment for pressure ulcers and its potential impact on clinical decision-making.

Imprecise correlated activity in self-organizing maps of spiking neurons.

How neurons communicate with each other to form effective circuits providing support to functional features of the nervous system is currently under debate. While many experts argue the existence of sparse neural codes based either on oscillations, neural assemblies or synchronous fire chains, other studies defend the necessity of a precise inter-neural communication to arrange efficient neural codes. As it has been demonstrated in neurophysiological studies, in the visual pathway between the retina and the visual cortex of mammals, the correlated activity among neurons becomes less precise as a direct consequence of an increase in the variability of synaptic transmission latencies. Although it is difficult to measure the influence of this reduction of correlated firing precision on the self-organization of cortical maps, it does not preclude the emergence of receptive fields and orientation selectivity maps. This is in close agreement with authors who consider that codes for neural communication are sparse. In this article, integrate-and-fire neural networks are simulated to analyze how changes in the precision of correlated firing among neurons affect self-organization. We observe how by keeping these changes within biologically realistic ranges, orientation selectivity maps can emerge and the features of neuronal receptive fields are significantly affected.

Factors determining the precision of the correlated firing generated by a monosynaptic connection in the cat visual pathway.

Across the visual pathway, strong monosynaptic connections generate a precise correlated firing between presynaptic and postsynaptic neurons. The precision of this correlated firing is not the same within thalamus and visual cortex. While retinogeniculate connections generate a very narrow peak in the correlogram (peak width < 1 ms), the peaks generated by geniculocortical and corticocortical connections have usually a time course of several milliseconds. Several factors could explain these differences in timing precision such as the amplitude of the monosynaptic EPSP (excitatory postsynaptic potential), its time course or the contribution of polysynaptic inputs. While it is difficult to isolate the contribution of each factor in physiological experiments, a first approximation can be done in modelling studies. Here, we simulated two monosynaptically connected neurons to measure changes in their correlated firing as we independently modified different parameters of the connection. Our results suggest that the precision of the correlated firing generated by strong monosynaptic connections is mostly determined by the EPSP time course of the connection and much less by other factors. In addition, we show that a polysynaptic pathway is unlikely to emulate the correlated firing generated by a monosynaptic connection unless it generates EPSPs with very small latency jitter.