MedMap

This project responds to the COVID-19 crisis by building on several longstanding lines work: the theory and practice of ontology and metaontology in our transpositional grammar.

“MedMap,” will be an Artificial Intelligence (AI) tool able to track with a high degree of semantic precision unique medical cases, and analyze multiple cases with unambiguous, standards-based labels. This may include the early cases of a novel pathogen, such as COVID-19, as well as the finely differentiated paths in disease progression and stratification that have made COVID-19 so perplexing for many medical professionals.

MedMap will be an ontology-suggestion and diagramming tool. It will leverage medical and everyday life ontologies, and the data they contain, suggesting labels with a high degree of semantic precision for supervised machine learning, and offering conjectures about possible labels in the case of unsupervised machine learning. In computer science, this project supplements the algorithmic power of machine and deep learning with under-exploited areas of data modelling, ontologies and semantics. Applied to the domain of medical informatics, this aligns with today’s objectives in the areas of precision or personalized medicine.

The semantic processes we are proposing to develop will enrich data and AI across populations at any scale by making datapoints smaller, more precise, and their relations specified with less ambiguity. Cross-mapping between ontologies by users will progressively yield further precision. As a novel event such as COVID-19 becomes a pandemic, the accumulated single cases will become a richer set of annotated training data to further enable the use of powerful supervised learning approaches to further extract knowledge and help address similar scenarios in the future.

The figure below illustrates a working prototype created to demonstrate how MedMap would work. On the left is medical data, which may be one of the three possible sites of application we discuss below: 1) a deeply documented, peer reviewed single clinical case; 2) a smart medical record; or 3) a raw dataset of any size extracted from a medical information system as structured or unstructured text.

Left side: multimodal case documentation with color-coded annotations according to ontology items in the visualization. Right side: A medical logic model visualization, where each node and relation is linked to a standard medical ontology, e.g. International Statistical Classification of Diseases (ICD); the International Classification of Functioning, Disability and Health (ICF); the Systematized Nomenclature of Medicine (SNOMED); Logical Observation Identifiers Names and Codes (LOINC); and The Drug Ontology (DRON), as well as everyday life ontologies operating at a high level of semantic precision, e.g. place (GeoNames), time and event (iCal), demographic profile (age, gender, race/ethnicity etc.), occupational classification (SIC: Standard Occupational Classification), or objects in the form of identifiable products (IAN: International Article Number) – and many others.

This proposal builds on twenty years of work with our research associates in the areas of semantic publishing, ontology design and application, and more recently, clinical case documentation in medical education. This work has been supported by US Department of Education, Gates Foundation and NSF grants.

Objectives

The objective of this project will be to design and prototype MedMap, a tool which will:

  • make suggestions based on standard medical and other ontologies, using AI to make smart recommendations based on unique configurations in single cases and patterns across multiple cases.
  • offer a software model for a distributed, AI-supported medical informatics system, able, with subject or patient permission, to share deidentified medical data in order to: 1) identify single cases that may be of concern, and 2) patterns of significance across small and larger populations.
  • create a clinical reasoning visualization tool that: 1) medical researchers and professionals can use to “map” a specific clinical case to support their clinical thinking and decision making, and 2) deepens multi-case comparisons, and the medical knowledge that underpins the suggestion system.

In supervised machine learning, users will identify semantically precise nodes and draw visualizations of clinical logic models or decision trees. In unsupervised machine learning, the system will offer label suggestions and node connections based on the semantic structure of the supporting ontologies and the previous actions of expert users according to the site of application: 1) medical researchers and students; 2) medical professionals; and 3) medical data analysts.

Among a number of possibilities, we anticipate the following potentially high-impact applications:

  1. Single clinical case documentation. This will fill a gap in medical science research and publication, which currently tends to prioritize larger population sizes. Semantically rich, single case documentation will highlight puzzling and worrisome cases, or cases of a kind frequently encountered but requiring more detailed documentation to differentiate variable disease causes and progressions. Using highly granular medical and everyday ontologies, multiple cases could be datamined for outlier warnings, as well as shared features.
  2. A smart medical record. This will be a shadow box overlay or parallel screen where: a) the content of a medical record can be more precisely specified by node selection from medical and everyday ontologies, then logic models created to assist the clinician in their thinking. The amount of extra work for the medical professional would be minimized. Effort would more than rewarded by smart node suggestions made by the system and the clinical logic and decision modeling by diagramming nodes in the visualization. In the COVID-19 crisis, the chief Scientist at the World Health organization, Soumya Swaminathan, has noted the difficulties of attaining and integrating case medical data. The MedMap tool would address this gap.
  3. Adding semantic precision to medical datasets. This application of the MedMap tool would support searching for already labelled datasets based on the salience of semantic connections; requesting or suggesting labels based on outlier alerts, or identifying semantically germane patterns in smaller or larger populations; and generating more semantically accurate topic models.

Potential Applications

The opportunities we envisage for each of these three potential applications are:

1) Biomedical science journal publishers and medical colleges

Pain Points: early alerts to new phenomena such as COVID-19 before they reach wider populations, and then tracing the varied impact on individuals.

Benefits: single case peer-reviewed documentation, where each case can be published in a semantically aware version of a journal article, and where highly granular, precision markup supports automated analysis across multiple cases. Not only would this be a tool for medical researchers and medical students, but also expand the possibilities for a wider range of medical practitioners to contribute in a more broadly distributed system of medical science.

2) Electronic medical record system providers

Pain Points: medical record keeping as an as an administrative task that offers limited value-add for the practitioner’s diagnosis and planning for patients.

Benefits: smart suggestion and decision tree tools to supplement medical records in support of precision and personalized medicine; distributed AI where, by sharing securely anonymized data, every medical professional contributes incrementally to the intelligence of the system.

3) Medical data analysts

Pain Points: medical data analysts, making sense of ambiguously or poorly labelled raw datasets, large and small.

Benefits: applying semantically deeper AI to analysis of data in supervised and unsupervised machine learning.