Research increasingly relies on interrogating large-scale data resources. The NIH National Heart, Lung, and Blood Institute developed the NHLBI BioData CatalystⓇ (BDC), a community-driven ecosystem where researchers, including bench and clinical scientists, statisticians, and algorithm developers, find, access, share, store, and compute on large-scale datasets. This ecosystem provides secure, cloud-based workspaces, user authentication and authorization, search, tools and workflows, applications, and new innovative features to address community needs, including exploratory data analysis, genomic and imaging tools, tools for reproducibility, and improved interoperability with other NIH data science platforms. BDC offers straightforward access to large-scale datasets and computational resources that support precision medicine for heart, lung, blood, and sleep conditions, leveraging separately developed and managed platforms to maximize flexibility based on researcher needs, expertise, and backgrounds. Through the NHLBI BioData Catalyst Fellows Program, BDC facilitates scientific discoveries and technological advances. BDC also facilitated accelerated research on the coronavirus disease-2019 (COVID-19) pandemic.

OBJECTIVE: This article describes the implementation of a privacy-preserving record linkage (PPRL) solution across PCORnet®, the National Patient-Centered Clinical Research Network. MATERIAL AND METHODS: Using a PPRL solution from Datavant, we quantified the degree of patient overlap across the network and report a de-duplicated analysis of the demographic and clinical characteristics of the PCORnet population. RESULTS: There were ∼170M patient records across the responding Network Partners, with ∼138M (81%) of those corresponding to a unique patient. 82.1% of patients were found in a single partner and 14.7% were in 2. The percentage overlap between Partners ranged between 0% and 80% with a median of 0%. Linking patients' electronic health records with claims increased disease prevalence in every clinical characteristic, ranging between 63% and 173%. DISCUSSION: The overlap between Partners was variable and depended on timeframe. However, patient data linkage changed the prevalence profile of the PCORnet patient population. CONCLUSIONS: This project was one of the largest linkage efforts of its kind and demonstrates the potential value of record linkage. Linkage between Partners may be most useful in cases where there is geographic proximity between Partners, an expectation that potential linkage Partners will be able to fill gaps in data, or a longer study timeframe.

MOTIVATION: As the number of public data resources continues to proliferate, identifying relevant datasets across heterogenous repositories is becoming critical to answering scientific questions. To help researchers navigate this data landscape, we developed Dug: a semantic search tool for biomedical datasets utilizing evidence-based relationships from curated knowledge graphs to find relevant datasets and explain why those results are returned. RESULTS: Developed through the National Heart, Lung and Blood Institute's (NHLBI) BioData Catalyst ecosystem, Dug has indexed more than 15 911 study variables from public datasets. On a manually curated search dataset, Dug's total recall (total relevant results/total results) of 0.79 outperformed default Elasticsearch's total recall of 0.76. When using synonyms or related concepts as search queries, Dug (0.36) far outperformed Elasticsearch (0.14) in terms of total recall with no significant loss in the precision of its top results. AVAILABILITY AND IMPLEMENTATION: Dug is freely available at https://github.com/helxplatform/dug. An example Dug deployment is also available for use at https://search.biodatacatalyst.renci.org/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Image-based cell counting is a fundamental yet challenging task with wide applications in biological research. In this paper, we propose a novel unified deep network framework designed to solve this problem for various cell types in both 2D and 3D images. Specifically, we first propose SAU-Net for cell counting by extending the segmentation network U-Net with a Self-Attention module. Second, we design an extension of Batch Normalization (BN) to facilitate the training process for small datasets. In addition, a new 3D benchmark dataset based on the existing mouse blastocyst (MBC) dataset is developed and released to the community. Our SAU-Net achieves state-of-the-art results on four benchmark 2D datasets - synthetic fluorescence microscopy (VGG) dataset, Modified Bone Marrow (MBM) dataset, human subcutaneous adipose tissue (ADI) dataset, and Dublin Cell Counting (DCC) dataset, and the new 3D dataset, MBC. The BN extension is validated using extensive experiments on the 2D datasets, since GPU memory constraints preclude use of 3D datasets. The source code is available at https://github.com/mzlr/sau-net.

This paper proposes a novel deep learning architecture involving combinations of Convolutional Neural Networks (CNN) layers and Recurrent neural networks (RNN) layers that can be used to perform segmentation and classification of 5 cardiac rhythms based on ECG recordings. The algorithm is developed in a sequence to sequence setting where the input is a sequence of five second ECG signal sliding windows and the output is a sequence of cardiac rhythm labels. The novel architecture processes as input both the spectrograms of the ECG signal as well as the heartbeats' signal waveform. Additionally, we are able to train the model in the presence of label noise. The model's performance and generalizability is verified on an external database different from the one we used to train. Experimental result shows this approach can achieve an average F1 scores of 0.89 (averaged across 5 classes). The proposed model also achieves comparable classification performance to existing state-of-the-art approach with considerably less number of training parameters.

Roadway safety, especially in rural areas, is one of the most critical components in transportation planning. In collaboration with North Carolina Department of Transportation (NCDOT), UNC Highway Safety Research Center (HSRC), and DOT Volpe National Transportation Systems Center, UNC Renaissance Computing Institute (RENCI) developed a roadside feature detection solution leveraging multiple convolutional neural networks. The solution used an iterative active learning (AL) computer vision model training pipeline integrated into an AI tool to detect safety features such as guardrails and utility poles in geographically distributed NC rural roads. We utilized transfer learning by adopting the Xception neural network architecture [1] as the feature extraction backbone which was then used in an iterative AL process supported by a web-based annotation tool. The annotation tool not only allowed for the collection of annotations through an iterative AL process for multiple safety features, it also enabled visual analysis and assessment of model prediction performance in the geospatial context. AL techniques were used to direct human annotators to label images that would most effectively improve the model aimed at minimizing the number of required training labels while maximizing the model’s performance. The iterative AL process combined with a common feature extraction backbone allowed fast model inference on millions of images in the AL sampling space. This enabled a rapid transition between AL rounds while also reducing the computing requirements for each round. Model feature extraction weights were then fine-tuned in the last round of AL to obtain the best accuracy. Since only about 2.7% of 2.6 million unlabeled images in the AL sampling space contain guardrails, there is a significant class imbalance problem that must be addressed in our AL sampling strategies for the guardrail classification model. In this paper, we present our AI tool processing pipeline and methodology and discuss our AL results and future work. Our AI tool can be used to detect roadside safety features and be extended to also locate them for assessing roadside hazards.

Humans understand videos from both the visual and audio aspects of the data. In this work, we present a self-supervised cross-modal representation approach for learning audio-visual correspondence (AVC) for videos in the wild. After the learning stage, we explore retrieval in both cross-modal and intra-modal manner with the learned representations. We verify our experimental results on the VGGSound dataset [1], and our approach achieves promising results.

OBJECTIVE: This study aimed to develop a novel, regulatory-compliant approach for openly exposing integrated clinical and environmental exposures data: the Integrated Clinical and Environmental Exposures Service (ICEES). MATERIALS AND METHODS: The driving clinical use case for research and development of ICEES was asthma, which is a common disease influenced by hundreds of genes and a plethora of environmental exposures, including exposures to airborne pollutants. We developed a pipeline for integrating clinical data on patients with asthma-like conditions with data on environmental exposures derived from multiple public data sources. The data were integrated at the patient and visit level and used to create de-identified, binned, "integrated feature tables," which were then placed behind an OpenAPI. RESULTS: Our preliminary evaluation results demonstrate a relationship between exposure to high levels of particulate matter ≤2.5 µm in diameter (PM2.5) and the frequency of emergency department or inpatient visits for respiratory issues. For example, 16.73% of patients with average daily exposure to PM2.5 >9.62 µg/m3 experienced 2 or more emergency department or inpatient visits for respiratory issues in year 2010 compared with 7.93% of patients with lower exposures (n = 23 093). DISCUSSION: The results validated our overall approach for openly exposing and sharing integrated clinical and environmental exposures data. We plan to iteratively refine and expand ICEES by including additional years of data, feature variables, and disease cohorts. CONCLUSIONS: We believe that ICEES will serve as a regulatory-compliant model and approach for promoting open access to and sharing of integrated clinical and environmental exposures data.

OBJECTIVE: Record linkage to integrate uncoordinated databases is critical in biomedical research using Big Data. Balancing privacy protection against the need for high quality record linkage requires a human-machine hybrid system to safely manage uncertainty in the ever changing streams of chaotic Big Data. METHODS: In the computer science literature, private record linkage is the most published area. It investigates how to apply a known linkage function safely when linking two tables. However, in practice, the linkage function is rarely known. Thus, there are many data linkage centers whose main role is to be the trusted third party to determine the linkage function manually and link data for research via a master population list for a designated region. Recently, a more flexible computerized third-party linkage platform, Secure Decoupled Linkage (SDLink), has been proposed based on: (1) decoupling data via encryption, (2) obfuscation via chaffing (adding fake data) and universe manipulation; and (3) minimum information disclosure via recoding. RESULTS: We synthesize this literature to formalize a new framework for privacy preserving interactive record linkage (PPIRL) with tractable privacy and utility properties and then analyze the literature using this framework. CONCLUSIONS: Human-based third-party linkage centers for privacy preserving record linkage are the accepted norm internationally. We find that a computer-based third-party platform that can precisely control the information disclosed at the micro level and allow frequent human interaction during the linkage process, is an effective human-machine hybrid system that significantly improves on the linkage center model both in terms of privacy and utility.

We present a framework for the estimation of driver behavior at intersections, with applications to autonomous driving and vehicle safety. The framework is based on modeling the driver behavior and vehicle dynamics as a hybrid-state system (HSS), with driver decisions being modeled as a discrete-state system and the vehicle dynamics modeled as a continuous-state system. The proposed estimation method uses observable parameters to track the instantaneous continuous state and estimates the most likely behavior of a driver given these observations. This paper describes a framework that encompasses the hybrid structure of vehicle-driver coupling and uses hidden Markov models (HMMs) to estimate driver behavior from filtered continuous observations. Such a method is suitable for scenarios that involve unknown decisions of other vehicles, such as lane changes or intersection access. Such a framework requires extensive data collection, and the authors describe the procedure used in collecting and analyzing vehicle driving data. For illustration, the proposed hybrid architecture and driver behavior estimation techniques are trained and tested near intersections with exemplary results provided. Comparison is made between the proposed framework, simple classifiers, and naturalistic driver estimation. Obtained results show promise for using the HSS-HMM framework.

Population informatics is the systematic study of populations via secondary analysis of massive data collections about people, called the social genome. A major challenge in building the social genome is the difficulty in data integration of heterogeneous and uncoordinated data while protecting the confidentiality of the data subjects. Here, we present our work in designing a flexible computerized third party linkage platform, Secure Decoupled Linkage (SDLink), which can provide both privacy protection and accurate high quality integrated data using a hybrid human-machine data integration system. Our evaluation results show that chaffing used in combination with universe manipulation is very effective in blocking inferences during the clerical review process.

To monitor and diagnose bottlenecks on network paths used for large-scale data transfers, there is an increasing trend to deploy measurement frameworks such as perfSONAR. These deployments use Web-services to expose vast data archives of current and historic measurements, which can be queried across end-to-end multi-domain network paths. Consequently, there has arisen a need to develop automated techniques and intuitive tools that help analyze these measurements for detecting and notifying prominent network anomalies such as plateaus in both real-time and offline manner. In this paper, we present a dynamically adaptive plateau-detection (APD) scheme and its implementation in our “OnTimeDetect” tool to enable consumers of perfSONAR measurements within the data-intensive scientific communities in overcoming their existing limitations of network anomaly detection and notification. We empirically evaluate our APD scheme in terms of accuracy, agility and scalability by using measurement traces collected by OnTimeDetect tool from worldwide perfSONAR deployments in HPC communities.

In this paper, a model of a humanoid biped is developed. The dynamics are formulated to simulate responses to a sudden backwards translational disturbance of the support surface. The effect of joint stiffnesses, the role of vestibular and proprioceptive sensory apparatus in the maintenance of balance, and the involvement of the centers of mass and pressure are taken into consideration and shown in a number of simulations. Toward this end, a three-link sagittal biped with three muscle pairs at the ankle, knee, and hip, and two pairs of two-jointed muscles corresponding to quadriceps-hamstring and the gastrocnemius-antagonist group is subjected to computational experiments. Excursions of the center of gravity and the center of pressure are compared under different conditions. Comparisons to biological results are also discussed.

Remote desktop access (RDA) applications have become vital for users involved in tasks such as tele-commuting, distance learning, and remote instrumentation. To deliver optimum user quality of experience (QoE), existing RDA applications are "network-aware" i.e., they employ online encoding adaptation that is based on network quality of service (QoS) measurement between the client and server ends. In this paper, we propose and evaluate an online RDA encoding adaptation scheme that is "human-and-network aware" i.e., our novel adaptation considers quality of application (QoA) performance perceived by the user in addition to the network QoS measured by the application. Owing to our offline QoA performance modeling strategy that uses polynomial regression and bootstrap sampling, our scheme does not require input from actual users for the online encoding adaptation. Our performance validation results show that our proposed human-network-aware adaptation outperforms the network-aware adaptation in terms of user QoE for a wide-variety of degraded network QoS conditions.

Computer-controlled scientific instruments such as electron microscopes, spectrometers, and telescopes are expensive to purchase and maintain. Also, they generate large amounts of raw and processed data that has to be annotated and archived. Cyber-enabling these instruments and their data sets using remote instrumentation cyberinfrastructures can improve user convenience and significantly reduce costs. In this paper, we discuss our experiences in gathering technical and policy requirements of remote instrumentation for research and training purposes. Next, we describe the cyberinfrastructure solutions we are developing for supporting related multi-user workflows. Finally, we present our solution-deployment experiences in the form of case studies. The case studies cover both technical issues (bandwidth provisioning, collaboration tools, data management, system security) and policy issues (service level agreements, use policy, usage billing). Our experiences suggest that developing cyberinfrastructures for remote instrumentation requires: (a) understanding and overcoming multi-disciplinary challenges, (b) developing reconfigurable-and-integrated solutions, and (c) close collaborations between instrument labs, infrastructure providers, and application developers.

We present a survey of modern parallel MATLAB techniques. We concentrate on the most promising and well supported techniques with an emphasis in SIP applications. Some of these methods require writing explicit code to perform inter-processor communication while others hide the complexities of communication and computation by using higher level programming interfaces. We cover each approach with special emphasis given to performance and productivity issues.

This work on the automatic generation of gestural scores from articulatory data is a follow-up to Jung (1996). The a priori information used to place gestures on the gestural score has been replaced by a statistical method based on the temporal decomposition reconstruction weights. Comparisons with Jung's gestural scores are reported.

This paper describes the application of artificial neural networks to acoustic-to-phonetic mapping. The experiments described are typical of problems in speech recognition in which the temporal nature of the input sequence is critical. The specific task considered is that of mapping formant contours to the corresponding CVC' syllable. We performed experiments on formant data extracted from the acoustic speech signal spoken at two different tempos (slow and normal) using networks based on the Elman simple recurrent network model. Our results show that the Elman networks used in these experiments were successful in performing the acoustic-to-phonetic mapping from formant contours. Consequently, we demonstrate that relatively simple networks, readily trained using standard backpropagation techniques, are capable of initial and final consonant discrimination and vowel identification for variable speech rates.

The application of artificial neural networks for phonetic-to-acoustic mapping is described. The specific task considered is that of mapping consonant-vowel-consonant (CVC) syllables to the corresponding formant values at different speech tempos. The performances of two different networks, the Elman recurrent network and a single hidden layer feedforward network, are compared. The results indicate that the recurrent network is able to generalize from the training set and produce valid formant contours for new CVC syllables that are not a part of the training set. It is shown that by choosing the proper input representation, the feedforward network is also capable of learning this mapping.>

The authors investigate the use of two types of neural networks, multilayer perceptrons (MLP) and learning vector quantizers (LVQ), as applied to isolated speaker-independent vowel recognition as a typical classification task. The LVQ algorithm used is a modification called the frequency-sensitive competitive-learning (FSCL) LVQ. The performance of each of these networks for different input feature sets is evaluated and compared. The effects of different distortion measures on recognition performance are also studied. The results show that the choice of the input feature set and the distortion measure can significantly affect recognition performance. It is shown that, while both the backpropagation (BP) and FSCL-LVQ algorithms can be applied to a set of vowel-recognition tasks, the FSCL-LVQ procedure offers an advantage over the MLP approach. The FSCL-LVQ algorithm allows the use of any appropriate distortion measure for particular input features, while the BP algorithm optimizes the weights by minimizing the squared errors between the actual and desired outputs. Consequently, for some tasks, the LVQ architecture can perform more accurate classification

Using neural networks for vector quantization (VQ) is described. The authors show how a collection of neural units can be used efficiently for VQ encoding, with the units performing the bulk of the computation in parallel, and describe two unsupervised neural network learning algorithms for training the vector quantizer. A powerful feature of the new training algorithms is that the VQ codewords are determined in an adaptive manner, compared to the popular LBG training algorithm, which requires that all the training data be processed in a batch mode. The neural network approach allows for the possibility of training the vector quantizer online, thus adapting to the changing statistics of the input data. The authors compare the neural network VQ algorithms to the LBG algorithm for encoding a large database of speech signals and for encoding images.>