1,720,981 research outputs found
Applying Self-Interaction Attention for Extracting Drug-Drug Interactions
No full textDiscovering the effect of the simultaneous assumption of drugs is a very important field in medical research that could improve the effectiveness of healthcare and avoid adverse drug reactions which can cause health problems to patients. Although there are several pharmacological databases containing information on drugs, this type of information is often expressed in the form of free text. Analyzing sentences in order to extract drug-drug interactions was the objective of the DDIExtraction-2013 task. Despite the fact that the challenge took place six years ago, the interest on this task is still active and several new methods based on Recurrent Neural Networks and Attention Mechanisms have been designed. In this paper, we propose a model that combines bidirectional Long Short Term Memory (LSTM) networks with the Self-Interaction Attention Mechanism. Experimental analysis shows how this model improves the classification accuracy reducing the tendency to predict the majority class resulting in false negatives, over several input configurations
The Impact of Self-Interaction Attention on the Extraction of Drug-Drug Interactions
No full textSince a large amount of medical treatments requires the assumption of multiple drugs, the discovery of how these interact with each other, potentially causing health problems to the patients, is the subject of a huge quantity of documents. In order to obtain this information from free text, several methods involving deep learning have been proposed over the years. In this paper we introduce a Recurrent Neural Network-based method combined with the Self-Interaction Attention Mechanism. Such a method is applied to the DDI2013-Extraction task, a popular challenge concerning the extraction and the classification of drug-drug interactions. Our focus is to show its effect over the tendency to predict the majority class and how it differs from the other types of attention mechanisms
Machine Learning Models for Predicting Short-Long Length of Stay of COVID-19 Patients
No full textDuring 2020 and 2021, managing limited healthcare resources and hospital beds has been a fundamental aspect of the fight against the COVID-19 pandemic. Predicting in advance the length of stay, and in particular identifying whether a patient is going to stay in the hospital longer or less than a week, can provide important support in handling resources allocation. However, there have been significant changes in terms of containment measures, virus diffusion, new treatments, vaccines, and new variants of SARS-CoV-2 during the last period. These changes pose several conceptual drift issues that can limit the usefulness of machine learning in this context. In this work, we present a machine learning system trained and tested using data from more than 6000 hospitalised patients in northern Italy, distributed over almost two years of pandemic. We show how machine learning can be effective even by analysing data over this long period of time, also exploiting a model that predicts the patient's outcome in terms of discharge or death. Furthermore, learning from data that also consider deceased patients is a common issue in predicting the length of stay because they have severe conditions similar to patients with a long stay period, but may actually have a very short duration of hospitalisation. For this purpose, we present a method for handling data from alive and deceased patients, exploiting more patient records, increasing the robustness of the model and its performance in this task. Finally, we investigate the features that are most relevant to the prediction of the simplified length of stay
Performance robustness of AI planners in the 2014 International Planning Competition
Full text link. Solver competitions have been used in many areas of AI to assess the current state of the art and guide future research and development. AI planning is no exception, and the International Planning Competition (IPC) has been frequently run for nearly two decades. Due to the organisational and computational burden involved in running these competitions, solvers are generally compared using a single homogeneous hardware and software environment for all competitors. To what extent does the specific choice of hardware and software environment have an effect on solver performance, and is that effect distributed equally across the competing solvers?In this work, we use the competing planners and benchmark instance sets from the 2014 IPC to investigate these two questions. We recreate the 2014 IPC Optimal and Agile tracks on two distinct hardware environments and eight distinct software environments. We show that solver performance varies significantly based on the hardware and software environment, and that this variation is not equal for all planners. Furthermore, the observed variation is sufficient to change the competition rankings, including the top-ranked planners for some tracks
Multi-task Learning Applied to Biomedical Named Entity Recognition Task
No full textRecent deep learning techniques have shown significant improvements in biomedical named entity recognition task. However, such techniques are still facing challenges; one of them is related to the limited availability of annotated text data. In this perspective, with a multi-task approach, simultaneously training different related tasks enables multi-task models to learn common features among different tasks where they share some layers with each other. It is desirable to used stacked long-short term memories (LSTMs) in such models to deal with a large amount of training data and to learn the underlying hidden structure in the data. However, the stacked LSTMs approach also leads to the vanishing gradient problem. To alleviate this limitation, we propose a multi-task model based on convolution neural networks, stacked LSTMs, and conditional random fields and use embedding information at different layers. The model proposed shows results comparable to state-of-the-art approaches. Moreover, we performed an empirical analysis of the proposed model with different variations to see their impact on our model
Two-Oracle Optimal Path Planning on Grid Maps
Full text linkPath planning on grid maps has progressed significantly in recent years, partly due to the Grid-based Path Planning Competition GPPC. In this work we present an optimal approach which combines features from two modern path planning systems, SRC and JPS+, both of which were among the strongest entrants at the 2014 edition of the competition. Given a current state s and a target state t, SRC is used as an oracle to provide an optimal move from s towards t. Once a direction is available we invoke a second JPS-based oracle to tell us for how many steps that move can be repeated, with no need to query the oracles between these steps. Experiments on a range of grid maps demonstrate a strong improvement from our combined approach. Against SRC, which remains an optimal solver with state-of-the-art speed, the performance improvement of our new system ranges from comparable to more than one order of magnitude faster.</p
Leveraging Multi-task Learning for Biomedical Named Entity Recognition
No full textBiomedical named entity recognition (BioNER) is the task of categorizing biomedical entities. Due to the specific characteristics of the names of biomedical entities, such as ambiguity among different concepts or different ways of referring to the same entity, the BioNER task is usually considered more challenging compared to standard named entity recognition tasks. Recent techniques based on deep learning not only significantly reduce the hand crafted feature engineering phase but also determined relevant improvements in the BioNER task. However, such systems are still facing challenges. One of them is the limited availability of annotated text data. Multi-task learning approaches tackle this problem by training different related tasks simultaneously. This enables multi-task models to learn common features among different tasks where they share some layers. To explore the advantages of the multi-task learning, we propose a model based on convolution neural networks, long-short term memories, and conditional random fields. The model we propose shows comparable results to state-of-the-art approaches. Moreover, we present an empirical analysis considering the impact of different word input representations (word embedding, character embedding, and case embedding) on the model performance
Recurrent Neural Networks for Daily Estimation of COVID-19 Prognosis with Uncertainty Handling
Full text linkMost ML-based applications for COVID-19 assess the general conditions of a patient trained and tested on cohorts of patients collected over a short period of time and are capable of providing an alarm a few days in advance, helping clinicians in emergency situations, monitor hospitalised patients and identify potentially critical situations at an early stage. However, the pandemic continues to evolve due to new variants, treatments, and vaccines; considering datasets over short periods could not capture this aspect. In addition, these applications often avoid dealing with the uncertainty associated with the prediction provided by machine learning models, potentially causing costly mistakes. In this work, we present a system based on Recurrent Neural Networks (RNN) for the daily estimate of the prognosis of COVID-19 patients that is built and tested using data collected over a long period of time. Our system achieves high predictive performance and uses an algorithm to effectively determine and discard those patients for whom RNN cannot predict the prognosis with sufficient confidence
Path Planning with CPD Heuristics
Full text linkCompressed Path Databases (CPDs) are a leading technique for optimal pathfinding in graphs with static edge costs. In this work we investigate CPDs as admissible heuristic functions and we apply them in two distinct settings: problems where the graph is subject to dynamically changing costs, and anytime settings where deliberation time is limited. Conventional heuristics derive cost-to-go estimates by reasoning about a tentative and usually infeasible path, from the current node to the target. CPD-based heuristics derive cost-to-go estimates by computing a concrete and usually feasible path. We exploit such paths to bound the optimal solution, not just from below but also from above. We demonstrate the benefit of this approach in a range of experiments on standard gridmaps and in comparison to Landmarks, a popular alternative also developed for searching in explicit state-spaces.</p
- …
