Setting the standard for high-quality studies using open health datasets

BackgroundGovernments worldwide are facing growing pressure to increase transparency, as citizens demand greater insight into decision-making processes and public spending. An example is the release of open healthcare data to researchers, as healthcare is one of the top economic sectors. Significant information systems development and computational experimentation are required to extract meaning and value from these datasets. We use a large open health dataset provided by the New York State Statewide Planning and Research Cooperative System (SPARCS) containing 2.3 million de-identified patient records. One of the fields in these records is a patient’s length of stay (LoS) in a hospital, which is crucial in estimating healthcare costs and planning hospital capacity for future needs. Hence it would be very beneficial for hospitals to be able to predict the LoS early. The area of machine learning offers a potential solution, which is the focus of the current paper.MethodsWe investigated multiple machine learning techniques including feature engineering, regression, and classification trees to predict the length of stay (LoS) of all the hospital procedures currently available in the dataset. Whereas many researchers focus on LoS prediction for a specific disease, a unique feature of our model is its ability to simultaneously handle 285 diagnosis codes from the Clinical Classification System (CCS). We focused on the interpretability and explainability of input features and the resulting models. We developed separate models for newborns and non-newborns.ResultsThe study yields promising results, demonstrating the effectiveness of machine learning in predicting LoS. The best R2 scores achieved are noteworthy: 0.82 for newborns using linear regression and 0.43 for non-newborns using catboost regression. Focusing on cardiovascular disease refines the predictive capability, achieving an improved R2 score of 0.62. The models not only demonstrate high performance but also provide understandable insights. For instance, birth-weight is employed for predicting LoS in newborns, while diagnostic-related group classification proves valuable for non-newborns.ConclusionOur study showcases the practical utility of machine learning models in predicting LoS during patient admittance. The emphasis on interpretability ensures that the models can be easily comprehended and replicated by other researchers. Healthcare stakeholders, including providers, administrators, and patients, stand to benefit significantly. The findings offer valuable insights for cost estimation and capacity planning, contributing to the overall enhancement of healthcare management and delivery.

Saliency detection has been widely used as the pre-processing of the computer-vision tasks. Existing propagation based saliency detection methods simply select a k-regular graph for saliency propagation, which usually leads to the mistaken highlighting of the long-range smooth background regions. In this paper, we design a novel graph for label propagation based saliency detection by considering the local consistency and the global symmetry of the image scene and updating the graph model based on smoothness assumption and cluster assumption. Then, we label the reliable seeds and propagate the saliency value through the designed graph. On two widely used large open benchmark data sets, the proposed method significantly outperforms thirteen state-of-the-arts under either quantitative or qualitative evaluation.

Integrating Open Government Data with Stratosphere for More Transparency

ImportanceThe PAWPER tape system is one of the three most accurate paediatric weight estimation systems in the world. The latest version of the tape, which does not rely on a subjective assessment of habitus, is the PAWPER XL-MAC method which uses length and mid-arm circumference (MAC) to estimate weight. It was derived and validated in a population in the USA and has not yet been fully validated in a population from a resource-limited setting.ObjectiveThe objective of this study was to evaluate the performance of the PAWPER XL-MAC tape weight estimation system in a large dataset sample of children from resource-limited settings.MethodsThis was a “virtual” study in which weight estimates were generated using the PAWPER XL-MAC tape and Broselow tape 2007B and 2011A editions in a very large open access dataset. The dataset contained anthropometric information of children aged 6 to 59 months from standardised nutritional surveys in 51 low- and middle-income countries. The performance of PAWPER XL-MAC method was compared with the Broselow tape and a new length- and habitus-based tape, the Ralston method.Main outcomes and measuresThe bias of the weight estimation methods was assessed using the mean percentage error (MPE) and precision using the 95% limits of agreement (LOA) of the MPE. The overall accuracy was denoted by the percentage of weight estimates falling within 10% and 20% of actual weight (abbreviated as p10 and p20 respectively).ResultsThe MPE (LOA) for the PAWPER XL-MAC tape, the Broselow 2007B and 2011A and Ralston method were 1.9 (-15.3, 19.2), 5.4 (-15.9, 26.7), 7.7 (-13.3, 30.5) and -0.7 (-20.2, 19.3) respectively. The p10 and p20 for each method were 79.3% and 96.9% for the PAWPER XL-MAC tape, 64.3% and 91.0% for the Broselow tape 2007B, 55.5% and 85.9% for the Broselow tape 2011A and 67.4 and 94.0% for the Ralston method respectively. The PAWPER XL-MAC system was statistically significantly more accurate than the Broselow tape 2011A, the Broselow tape 2007B and the Ralston method. The relative difference in accuracy (p10) was 43% (odds ratio 4.4 (4.4, 4.5), p<0.001), 23% (odds ratio 2.9 (2.8, 2.9), p<0.001) and 18% (odds ratio 1.8 (1.8, 1.8), p<0.001) compared to each method, respectively.Conclusions and relevanceThe PAWPER XL-MAC tape performed well in this study and was statistically significantly more accurate than both the Broselow tape editions and the Ralston method. This difference was substantial and clinically important. The tape did not perform as well at extremes of habitus-type, however, and might benefit from recalibration.

Convolutional neural networks (CNN) are a powerful deep learning method for medical image segmentation. However performance in clinical practice often deteriorates when attempting to generalize models trained from a particular source domain to a different target domain (e.g. different vendor, acquisition parameters, protocols). To address this issue, domain adaptation has attracted increasing attention because it can minimize distribution differences among different but related domains. Extending from this prior work, we introduce Co-Unet-GAN, a co-learning domain adaptation and segmentation model addressing the domain shift problem. In this model, we train a Unet segmentation network and an image translation generative adversarial network (GAN) together to generalize performance across domains given supervised data only in the source domain. We evaluate our model on two large open echocardiography datasets, using the CAMUS set as supervised source domain and EchoNet-Dynamic as the unsupervised target. In this context, we obtain mean absolute error on ejection fraction of 9.67% on Co-Unet-GAN compared to 11.28% for a previously published Unet-GAN. Our Co-Unet-GAN for image translation and segmentation is a promising solution to the domain shift problem.

Spontaneous fluctuations of resting-state functional connectivity have been studied in many ways, but grasping the complexity of brain activity has been difficult. Dimensional complexity measures, which are based on Eigenvalue (EV) spectrum analyses (e.g., Ω entropy) have been successfully applied to EEG data, but have not been fully evaluated on functional MRI recordings, because only through the recent introduction of fast multiband fMRI sequences, feasable temporal resolutions are reached. Combining the Eigenspectrum normalization of Ω entropy and the scalable architecture of the so called Multivariate Principal Subspace Entropy (MPSE) leads to a new complexity measure, namely normalized MPSE (nMPSE). It allows functional brain complexity analyses at varying levels of EV energy, independent from global shifts in data variance. Especially the restriction of the EV spectrum to the first dimensions, carrying the most prominent data variance, can act as a filter to reveal the most discriminant factors of dependent variables. Here we look at the effects of healthy aging on the dimensional complexity of brain activity. We employ a large open access dataset, providing a great number of high quality fast multiband recordings. Using nMPSE on whole brain, regional, network and searchlight approaches, we were able to find many age related changes, i.e., in sensorimotoric and right inferior frontal brain regions. Our results implicate that research on dimensional complexity of functional MRI recordings promises to be a unique resource for understanding brain function and for the extraction of biomarkers.

Sharing research artifacts is known to help people to build upon existing knowledge, adopt novel contributions in practice, and increase the chances of papers receiving attention. In Model-Driven Engineering (MDE), openly providing research artifacts plays a key role, even more so as the community targets a broader use of AI techniques, which can only become feasible if large open datasets and confidence measures for their quality are available. However, the current lack of common discipline-specific guidelines for research data sharing opens the opportunity for misunderstandings about the true potential of research artifacts and subjective expectations regarding artifact quality. To address this issue, we introduce a set of guidelines for artifact sharing specifically tailored to MDE research. To design this guidelines set, we systematically analyzed general-purpose artifact sharing practices of major computer science venues and tailored them to the MDE domain. Subsequently, we conducted an online survey with 90 researchers and practitioners with expertise in MDE. We investigated our participants' experiences in developing and sharing artifacts in MDE research and the challenges encountered while doing so. We then asked them to prioritize each of our guidelines as essential, desirable, or unnecessary. Finally, we asked them to evaluate our guidelines with respect to clarity, completeness, and relevance. In each of these dimensions, our guidelines were assessed positively by more than 92\% of the participants. To foster the reproducibility and reusability of our results, we make the full set of generated artifacts available in an open repository at \texttt{\url{https://mdeartifacts.github.io/}}.

The black-box nature of machine learning models hinders the deployment of some high-accuracy medical diagnosis algorithms. It is risky to put one's life in the hands of models that medical researchers do not fully understand or trust. However, through model interpretation, black-box models can promptly reveal significant biomarkers that medical practitioners may have overlooked due to the surge of infected patients in the COVID-19 pandemic. This research leverages a database of 92 patients with confirmed SARS-CoV-2 laboratory tests between 18th January 2020 and 5th March 2020, in Zhuhai, China, to identify biomarkers indicative of infection severity prediction. Through the interpretation of four machine learning models, decision tree, random forests, gradient boosted trees, and neural networks using permutation feature importance, partial dependence plot, individual conditional expectation, accumulated local effects, local interpretable model-agnostic explanations, and Shapley additive explanation, we identify an increase in N-terminal pro-brain natriuretic peptide, C-reaction protein, and lactic dehydrogenase, a decrease in lymphocyte is associated with severe infection and an increased risk of death, which is consistent with recent medical research on COVID-19 and other research using dedicated models. We further validate our methods on a large open dataset with 5644 confirmed patients from the Hospital Israelita Albert Einstein, at São Paulo, Brazil from Kaggle, and unveil leukocytes, eosinophils, and platelets as three indicative biomarkers for COVID-19.

e13604 Background: Cancer is a complex and polymorphic disease. Large open datasets are currently available, though unstructured and heterogeneous. As such, expertise from various fields is needed to curate and analyze these datasets. Epidemium was set up as an innovative multidisciplinary community-based program to tackle these challenges and offer a new approach to cancer epidemiology research using big data. Methods: Epidemium combines cancer knowledge and data science expertise with an open, inclusive and collaborative approach. It invites any willing participant to submit a project or to join an existing one, regardless of his background. To foster synergies within the open community, regular conferences, meetings, and data challenges were set up, combined with a dedicated web platform - built using open source software - hosting selected resources and open data sets (WHO, Roche, FAO, World Bank, ILO, etc.). A six-month community-oriented Challenge allowed teams to make contributions that were reviewed by an Independent Ethics Committee and a Scientific Committee. Results: As of June 2016, Epidemium had gathered 678 members with different backgrounds: 19% Health Professionals, 28% Computer Scientists, and 53% Data Scientists. 15 projects were submitted to the Challenge and 3 were awarded: Baseline, CancerViz, and ELSE, focusing respectively on risk and protective factors of cancer, distribution of cancer in time and space, and environmental changes and cancer. All contributions (3.600 on the wiki) were shared in open source. Conclusions: Epidemium provides a proof of concept showing that collective intelligence can be leveraged to tackle complex problems such as cancer epidemiology and big data management. The proposed framework accelerates research initiatives through interconnected tools, allowing real time analysis and collaboration within an open multidisciplinary ecosystem.

Abstract. Dryland salinity remains a major global natural resource management concern, and which is amplified in Australia. However, limited detailed space-time data sets with observations of stream and groundwater salinity has constrained a deep understanding of the range of processes that can lead to dryland salinity problems in landscapes. The aim of this study is to report on the open dataset (https://doi.org/10.25910/m0wp-8890, Vervoort et al., 2025) resulting from a 14 year data collection effort in a subcatchment of the Murrumbidgee catchment in New South Wales, Australia. Over a 14 year period a series of different sampling campaigns has resulted in a large dataset with hydrogeochemical data which includes both in-situ (field) data and post laboratory analysis of major anions and cations. This data is augmented with observed groundwater levels and publicly available streamflow and climate data (https://doi.org/10.17605/OSF.IO/BEUWK, Vervoort and Akter, 2024). The data set covers 23 groundwater sample sites and 39 surface water sites. Because the data was collected by four distinct groups and over many years, we investigate whether this has caused a bias in the dataset. In addition, we show the major spatial and temporal trends to provide an overview of the dataset. The dataset is made open access to encourage further research and the current paper shows the richness of the collected data and opportunities for further research.

Monitoring electricity consumption in residential buildings is an important way to help reduce energy usage. Nonintrusive load monitoring is a technique to separate the total electrical load of a single household into specific appliance loads. This problem is difficult because we aim to extract the energy consumption of each appliance by only using the total electrical load. Deep transfer learning is expected to solve this problem. This paper proposes a deep neural network model based on an attention mechanism. This model improves the traditional sequence-to-sequence model with a time-embedding layer and an attention layer so that it can be better applied in nonintrusive load monitoring. In particular, the improved model abandons the recurrent neural network structure and shortens the training time, which means it is more appropriate for use in model pretraining with large datasets. To verify the validity of the model, we selected three open datasets and compared them with the current leading model. The results show that transfer learning can effectively improve the prediction ability of the model, and the model proposed in this study has a better performance than the most advanced available model.

Big Data usually refers to the very large datasets generated by scientists, to the many petabytes of held by companies like Facebook and Google, and to analyzing real-time assets like the stream of twitter messages emerging from events around the world. Key areas of interest include technologies to manage much larger datasets, technologies for the visualization and analysis of databases, cloud-based management and mining algorithms. Recently, however, we have begun to see the emergence of another, and equally compelling challenge -- that of the Broad data that emerges from millions and millions of raw datasets available on the World Wide Web. For broad the new challenges that emerge include Web-scale search and discovery, rapid and potentially ad hoc integration of datasets, visualization and analysis of only-partially modeled datasets, and issues relating to the policies for use, reuse and combination. In this talk, we present the broad challenge and discuss potential starting points for solutions including those arising from research in the Semantic Web area. We illustrate these approaches using from a meta-catalog of over 1,000,000 open datasets that have been collected from about two hundred governments around the world.

Is power everything? What can we learn from large data sets

Open Data philosophy is becoming more popular among scientists. Open Data approach aims to transform science by making high-quality and well-documented scientific data open to everybody in order to promote collaboration and transparency. In diffuse optical and near-infrared spectroscopy community, a large measurement dataset collected with state-of-the-art instrumentation applied on well-defined phantoms is still missing. Within that context, several European labs from BitMap network1 have collected diffuse optical data on standard phantoms involving the largest set of diffuse optics instruments published until now. In this work, we present a running project on the open dataset and associated reporting tools.

Biomedical Instrumentation is one of the fastest health emerging innovative technologies with proven contribution towards interdisciplinary medicine, it helps physicians to diagnose complex medical problems and provide treatment to patients precisely and safely. With this technological trend, explainable artificial intelligence, biomedical image processing and augmented intelligence can provide a tool that can help pediatricians, pulmonology and otolaryngology physicians, epidemiologists and pediatric practitioners to interpretably and reliably diagnose chronic and acute respiratory disorders in children, adolescents and infants. Unfortunately, the reliability of digital image processing for pulmonary disease diagnosis often depends on availability of large chest X-ray image datasets. This work presents a reliable interpretable deep transfer learning approach for pediatric pulmonary health evaluation regardless of the scarcity and limited annotated pediatric chest X-ray Image dataset sizes. This approach leverages a combination of computer vision tools and techniques to reduce child morbidity and mortality through predictive and preventive medicine with reduced surveillance risks and affordability in low resource settings. With open datasets, the deep neural networks classified the generated augmented images into 4 classes namely; Normal, Covid-19, Tuberculosis and Pneumonia at an accuracy of 97%, 97%, 70%, and 73% respectively with recall of 100% for Pneumonia and overall accuracy of 79% at only 10 epochs for both regular and transferred learning.