POSTER ABSTRACTS   from Third Annual Public Meeting: Mobilizing Computable Biomedical Knowledge (MCBK 2020)

identified


Medical Informatics Corporation, Houston, Texas
We present Sickbay, an FDA-cleared software platform for remote patient monitoring and alarm distribution. It passively records timesynchronized waveforms, vitals, alarms, and settings from bedside medical devices along with labs, medications and observations. Sharing and reuse of biomedical data is critical to enhance research reproducibility and increase efficiency in translational biomedical sciences. [1][2][3] This requires biomedical data and processes to be findable, accessible, interoperable and reusable according to the FAIR guiding principles. 4 The capture of sufficient metadata is a key requirement for successful data harmonization and integration, knowledge presentation and research process management.
In this presentation we describe a generalizable approach to managing metadata for diverse informatics applications in the translational research spectrum. 5 OpenFurther's metadata repository 6,7  Considering the data and process complexity within translation research, we conceptually divided metadata management into three categories: • Data Metadata: Describes the data output resulting from an observation or measurement. This could include sensor measurements, output of computational models, clinical observations, genomic sequence annotations, socio-behavioral data, and participant report data among others.
• Process Metadata: Describes research or data processes within informatics infrastructure. These include among others sequences of steps followed in different computational models in order to generate outputs, and data transformation and integration workflows to harmonize source data as events or into analytical models. An example of a research process is sensor deployment.
• Knowledge Resources: Metadata that describes the source or instrument used to collect, measure or derive data. These could include sensor devices, electronic medical records or study specific data collection instruments.
We have developed and evaluated the MDR in different technologies including relational, 6 graph and document stores 8,9 of different use cases including data federation, 6 integration, 10 data quality assessment, 11 knowledge presentation, 12 and sensor-based exposomic research. 8,9 Future directions include developing methods automate metadata discovery process, 13,14 consume and store such metadata into the MDR, 15 maintenance of metadata provenance and trajectory using approaches like blockchain, and data and process orchestration in ultra large scale systems.
This poster presentation is particularly relevant to informaticians developing methods and tools for translational research.   Model is an epistemological framework that represents how hybrid reasoning can be employed in a clinical setting. Based on this framework and by leveraging semantic technologies, we design and implement an AI system that can support healthcare providers with their reasoning tasks. We focus on reasoning tasks, such as differential diagnosis, treatment planning, and plan critiquing, considering strategies clinicians commonly use. By providing clinicians with an evidence-based clinical decision-support system, this work has the potential to improve patient care.
How can we trust computable knowledge? Some UK perspectives Results: The discussion identified six key themes as crucial for trust: regulation, transparency, ease of use, confidence, evaluation and particular issues to resolve (multi-morbidity and missing data). Regulatory concerns included "off-label" algorithms, professional requirements in software engineering and safety-critical design approaches.
Conclusions: There is clinical and public expectation of suitable controls and safety methods in the design, implementation and governance of computable biomedical knowledge.

Boston, Massachusetts
Easy and consistent access to high-quality information is critical nowadays considering the COVID-19 pandemic. Stakeholders with diverse information needs use online sources to find guidance. In an effort to improve the "findability" of COVID-19 information developed by public health agencies, we are creating a taxonomy to index and classify available guidance using an inductive methodology. The ongoing effort to identify relevant MeSH terms has revealed gaps related to COVID-19. Over 60% of the taxonomy concepts were successfully matched to current MeSH terms, approximately 17% had related terms but meaning nuances relevant to COVID-19 were difficult to represent, and 2% of terms had no match. An example of a missing concept is "reopening procedure," made important given the quarantine period. Promptly establishing comprehensive concept coverage is critical to timely and effective information retrieval during and after a pandemic.  We are also targeting new use cases, especially those with crossspecialty appeal and a focus on pandemic protocols (eg, COVID-19 severity). Finally, ACEP is actively pursuing the implementation of developed use cases into EMR systems. Intelligence (AI) models and systems. However, current support for explainability may not address specific end-user needs that arise in actual use. To bridge this gap, we designed a semantic representation, informed by a literature review and user study, which system designers can leverage to connect explanation types to user needs.
Methods: We conducted a literature review in various domains to identify different explanation types and user goals they address (eg, for education, clarification, and exploration purposes). We then designed an ontology that includes system and user attributes and connects needs to the literature-derived explanation types. We validated and refined this ontology through user studies involving a concrete example in guideline-based healthcare.
Results: We identified semantic attributes needed for AI explainability design, including representing forms of knowledge and reasoning.
User studies confirmed the need for multiple explanation types, including contrastive, counterfactual, case-based, and contextual explanations in clinical reasoning, and informed our understanding of when each applies.
Conclusion: This poster is particularly relevant to system designers who may be able to leverage our ontology-enabled infrastructure to build explainable AI systems in clinical decision support settings.