No abstract

At the end of 2019, Chinese authorities alerted the World Health Organization (WHO) of the outbreak of a new strain of the coronavirus, called SARS-CoV-2, which struck humanity by an unprecedented disaster a few months later. In response to this pandemic, a publicly available dataset was released on Kaggle which contained information of over 63,000 papers. In order to facilitate the analysis of this large mass of literature, we have created a knowledge graph based on this dataset. Within this knowledge graph, all information of the original dataset is linked together, which makes it easier to search for relevant information. The knowledge graph is also enriched with additional links to appropriate, already existing external resources. In this paper, we elaborate on the different steps performed to construct such a knowledge graph from structured documents. Moreover, we discuss, on a conceptual level, several possible applications and analyses that can be built on top of this knowledge graph. As such, we aim to provide a resource that allows people to more easily build applications that give more insights into the COVID-19 pandemic.

Medicines based on messenger RNA (mRNA) hold immense potential, as evidenced by their rapid deployment as COVID-19 vaccines. However, worldwide distribution of mRNA molecules has been limited by their thermostability, which is fundamentally limited by the intrinsic instability of RNA molecules to a chemical degradation reaction called in-line hydrolysis. Predicting the degradation of an RNA molecule is a key task in designing more stable RNA-based therapeutics. Here, we describe a crowdsourced machine learning competition (‘Stanford OpenVaccine’) on Kaggle, involving single-nucleotide resolution measurements on 6,043 diverse 102–130-nucleotide RNA constructs that were themselves solicited through crowdsourcing on the RNA design platform Eterna. The entire experiment was completed in less than 6 months, and 41% of nucleotide-level predictions from the winning model were within experimental error of the ground truth measurement. Furthermore, these models generalized to blindly predicting orthogonal degradation data on much longer mRNA molecules (504–1,588 nucleotides) with improved accuracy compared with previously published models. These results indicate that such models can represent in-line hydrolysis with excellent accuracy, supporting their use for designing stabilized messenger RNAs. The integration of two crowdsourcing platforms, one for dataset creation and another for machine learning, may be fruitful for other urgent problems that demand scientific discovery on rapid timescales.

In industry, dashboards are often used to monitor fleets of assets, such as trains, machines or buildings. In such industrial fleets, the vast amount of sensors evolves continuously, new sensor data exchange protocols and data formats are introduced, new visualization types may need to be introduced and existing dashboard visualizations may need to be updated in terms of displayed sensors. These requirements motivate the development of dynamic dashboarding applications. These, as opposed to fixed-structure dashboard applications, allow users to create visualizations at will and do not have hard-coded sensor bindings. The state-of-the-art in dynamic dashboarding does not cope well with the frequent additions and removals of sensors that must be monitored—these changes must still be configured in the implementation or at runtime by a user. Also, the user is presented with an overload of sensors, aggregations and visualizations to select from, which may sometimes even lead to the creation of dashboard widgets that do not make sense. In this paper, we present a dynamic dashboard that overcomes these problems. Sensors, visualizations and aggregations can be discovered automatically, since they are provided as RESTful Web Things on a Web Thing Model compliant gateway. The gateway also provides semantic annotations of the Web Things, describing what their abilities are. A semantic reasoner can derive visualization suggestions, given the Thing annotations, logic rules and a custom dashboard ontology. The resulting dashboarding application automatically presents the available sensors, visualizations and aggregations that can be used, without requiring sensor configuration, and assists the user in building dashboards that make sense. This way, the user can concentrate on interpreting the sensor data and detecting and solving operational problems early.

The Matrix Profile is a state-of-the-art time series analysis technique that can be used for motif discovery, anomaly detection, segmentation and others, in various domains such as healthcare, robotics, and audio. Where recent techniques use the Matrix Profile as a preprocessing or modelling step, we believe there is unexplored potential in generalizing the approach. We derived a framework that focuses on the implicit distance matrix calculation. We present this framework as the Series Distance Matrix (SDM). In this framework, distance measures (SDM-generators) and distance processors (SDM-consumers) can be freely combined, allowing for more flexibility and easier experimentation. In SDM, the Matrix Profile is but one specific configuration. We also introduce the Contextual Matrix Profile (CMP) as a new SDM-consumer capable of discovering repeating patterns. The CMP provides intuitive visualizations for data analysis and can find anomalies that are not discords. We demonstrate this using two real world cases. The CMP is the first of a wide variety of new techniques for series analysis that fits within SDM and can complement the Matrix Profile.

Manufacturers can plan predictive maintenance by remotely monitoring their assets. However, to extract the necessary insights from monitoring data, they often lack sufficiently large datasets that are labeled by human experts. We suggest combining knowledge-driven and unsupervised data-driven approaches to tackle this issue. Additionally, we present a dynamic dashboard that automatically visualizes detected events using semantic reasoning, assisting experts in the revision and correction of event labels. Captured label corrections are immediately fed back to the adaptive event detectors, improving their performance. To the best of our knowledge, we are the first to demonstrate the synergy of knowledge-driven detectors, data-driven detectors and automatic dashboards capturing feedback. This synergy allows a transition from detecting only unlabeled events, such as anomalies, at the start to detecting labeled events, such as faults, with meaningful descriptions. We demonstrate all work using a ventilation unit monitoring use case. This approach enables manufacturers to collect labeled data for refining event classification techniques with reduced human labeling effort.

No abstract

Integrating Internet of Things (IoT) sensor data from heterogeneous sources with domain knowledge and context information in real-time is a challenging task in IoT healthcare data management applications that can be solved with semantics. Existing IoT platforms often have issues with preserving the privacy of patient data. Moreover, configuring and managing context-aware stream processing queries in semantic IoT platforms requires much manual, labor-intensive effort. Generic queries can deal with context changes but often lead to performance issues caused by the need for expressive real-time semantic reasoning. In addition, query window parameters are part of the manual configuration and cannot be made context-dependent. To tackle these problems, this paper presents DIVIDE, a component for a semantic IoT platform that adaptively derives and manages the queries of the platform’s stream processing components in a context-aware and scalable manner, and that enables privacy by design. By performing semantic reasoning to derive the queries when context changes are observed, their real-time evaluation does require any reasoning. The results of an evaluation on a homecare monitoring use case demonstrate how activity detection queries derived with DIVIDE can be evaluated in on average less than 3.7 seconds and can therefore successfully run on low-end IoT devices.