Personal Health Knowledge Graphs (PHKGs) are not yet ubiquitous, even though they have a great potential to enrich general knowledge captured in various Knowledge Graphs by adding personal contexts. This poster paper presents work in progress about a pipeline for generating PHKGs from tree-structured Electronic Health Record (EHR) data by applying a hierarchical ontological approach. This pipeline could also be applied to other domains of Personal Knowledge Graphs. Moreover, this pipeline targets the intersection between the symbolic representation of knowledge used for computational semantics and numeric graph data representation used for graph analysis and machine learning. We present the ifrst results from applying this pipeline to synthetic patient EHRs with the diagnosis of colorectal cancer (based on Synthea). The resulting numeric representation of PHKGs or their subgraphs can be used in many practical graph algorithms. Finally, our pipeline study uncovers future research on how this numeric representation of PHKGs should be embedded into continuous and low-dimensional vector space to utilize graph machine learning and deep learning methods.
Personal health knowledge graphs (PHKGs) represent structured information about entities
related to a patient’s health and well-being, attributes, and relations between them. Unlike
Knowledge Graphs (KGs), PHKGs are not yet ubiquitous. PHKGs should be generated for
individual patients from numerous information sources such as electronic health records (EHRs),
wearables and mobile health apps, sensors, and patient annotated texts and notes related to the
patient’s condition. In principle, PHKGs can be populated by all techniques mentioned in [
This poster paper presents a work in progress about a pipeline on how to populate a PHKG from several data sources. First, the data must be harmonized according to a flexible and helpful data model for data analysis. Then, we apply this approach to the PHKG population from data stored in EHRs. In addition, ongoing work will provide evidence that this approach generalizes to PKGs and other data sources, and enables linking PHKGs with KGs and generating new KGs.
The proposed representation of a PHKG is composed of two elements: a domain graph and
a mapping from the nodes and edge labels of the data graph to those of the domain graph in
which they are called entities and relation-types, respectively [
Ontologies such as SNOMED CT and LOINC (also a part of UMLS) give identity to our nodes
in PHKGs [
A PHKG is thus represented as a directed edge-labeled graph that enables querying and
reasoning. However, most graph data analysis techniques do not apply to this representation.
Therefore, we need to transform it into an undirected or directed graph without edge labels (i.e.,
predicate names). A directed graph is thus projected by optionally selecting a sub-graph from
the data graph from which all edge labels can be dropped. The proposed pipeline to populate a
PHKG from various data sources and use it for graph analysis is as follows:
1. Find or create an ontology for data harmonization: Our research suggests that the
mentioned Simple Event Model Ontology is a universal and useful upper-ontology that
can be extended by more specific models for various event types.
2. Choose a standard or format for harmonized data transformation into graph
dataset: As our chosen example of HL7 FHIR data was available in a tree-structured
format (JSON) and a tool to convert the data into HL7 FHIR RDF format existed (see
Section 3), the choice was straightforward. But for diferent categories of data other
formats such as KGTK [
CT code) and the begin timestamp. 6. Apply graph algorithms, such as operators, distance measures, and shortest paths to one or more PHKGs in their numeric form. Even if the edge labels and other edge and node meta data are dropped for further graph analysis, it is a good practice to create a possibility to retrieve this information even for the outputs of the graph analysis. 7. If needed, convert the result of the graph analysis back into the symbolic form to expand the original PHKG by new knowledge.
This choice of a strategy to convert any source PHKG into its symbolic representation for analysis is not trivial. It requires empirical validation and iterations of the steps 5 and 6 to develop new insights from the graph analysis. In addition, more study is needed to understand the efects of such strategies more generally on the results of diferent analytical techniques.
We applied the methods and pipeline proposed in the Section 2 on synthetic data generated
for the colorectal cancer diagnosis according to [
A PHKG (only a part is displayed) fhir:Patient
fhir:Condition rdf:type fhir:Patient pt01
sem: hasActor „malignant tumor of colon“@en
rdf:type fhir:Condition
c01 rdfs:label sem:eventType sem:hasSubEvent sem:hasBeginTimeStamp 2011-05-21T05:14-53+02:00 rdfs:label „partial resection of colon“@en sem:hasSubEvent
sem: sem: fhir:Procedure has fhir:Procedure has fhir:Procedure p01 ESvuebnt p02 ESvuebnt p03
rdfs:label rdfs:label
DBpedia:
Colorectal cancer
fhir:Observation
o01
Owlready2 with its PyMedTermino2 for easy access to domain ontologies3 and NetworkX
[
Once we had the numeric representation of PHKG subgraphs for our 325 synthetic patients, performed according to the Figure 1, we developed an algorithm for an ontological graph union to create a KG containing diferent patient pathways from the point of diagnosis to the outcome. This algorithm can contract nodes of the same type (in our case, the same SNOMED CT code of the underlying procedure) across patient PHKGs while preserving all node features. In theory, if the set of PHKGs were representative enough, this KG would cover all possible ways of treatment and their outcomes. We analyzed the most connected nodes - procedures with the help of the degree centrality. They represent the key decision points in the patients’ treatment or life events.
It is also possible to analyze the shortest simple paths from the principal diagnosis to the outcome, considering the weights. For our group of patients, the shortest simple path from the diagnosis to the unfortunate event of death is length four because the diagnosis happened at a very late stage of the disease.
In our future work, we will further explore the best combination of ontologies, knowledge, and data harmonization to create a universal pipeline for the PKG population. In addition, as we can see in Figure 1, the adjacency matrix is sparse (the sparsity equals 2/3 in this case), and this fact about the numeric representation holds in general. Therefore, a more eficient representation in continuous and low-dimensional vector space with the help of node and graph embedding should be researched.
Finally, in our pipeline, we found a more straightforward representation of the partial PHKG knowledge in which we could drop the edge labels (the nodes representing the events (procedures) were connected only by a sub-event relationship to capture the sequence of events in time). However, the more properties the graph embedder encodes, the better results can be retrieved in later tasks. Therefore, we can generate a more complex subgraph from our source PHKG with edge labels.