11/15/2018 Drug Side Effects Data Representation and Full Spectrum Inferencing using Knowledge Graphs in Intelligent Telehealth Presented on Student-Faculty Research Day, CSIS , Pace University , May 6th 2016 Researcher : Sar Jayaraman, DPS 2016 Research Advisor : Professor Lixin Tao S Jayaraman
Nearly 7 in 10 Americans take some form of Prescription Drugs – Mayo Clinic Survey. 11/15/2018
Growth of Reported Side effects - USA 11/15/2018
MedWatch approach 11/15/2018 There are nearly 2,000 medications, 94,450 health products and around 175,950 health packages with different active ingredients on the market in the U.S.[11] To handle the challenges associated with the data collection, reporting of these numerous adverse events, MedWatch system was implemented in 1993 Raw data from the MedWatch system, together with adverse drug reaction reports from manufacturers, are part of a public database.
Other Approaches – Literature survey – Drug Ontology By Samson et al – findings 11/15/2018 Mapping Needed Between CT and Patient Record Patient or Drug Class not included Not focused on Adverse data Hard to use by Domain experts Emulation of drug relationships
Limitations of the current approach 11/15/2018 Drug adverse reaction data contains important constraints about side-effects and conflict avoidance of component and compound drug. They are critically important in checking out prescriptions to avoid complications. Although MedWatch FAERS drug data are in XML, it doesn't have a proper knowledge representation mechanism to clearly specify all kinds of dependencies among the drug components and drugs. Therefore one has to depend on human interpretation to check prescriptions which can be error-prone.
Challenges with capturing the drugs and drug classes – relationship A drug may be classified by the “chemical type of the active ingredient” or by the way it is used to treat a particular condition. Each drug can be classified into one or more drug classes adding more complexity. Drug adverse events marked with a compound (parent drug class) applies to all its component drugs but the reverse is not true. All these relations are dynamic in nature requiring flexible approach to capture the knowledge. 11/15/2018
11/15/2018 Problem Statement: Drugs often inherit, important adverse reaction constraints from their compound drugs, which is not represented in the current format. - [Problem1 or P1] Drug side effect relationships should be kept updated current always, as new relationships emerge on component or compound drugs. Though this is possible with the current approach, it requires a very high effort for both updates and maintenance - [Problem 2 or P2] Drug side effect relationships should be interpreted easily by users (patients or doctors) allowing them to access the full spectrum of side effects. The current approach is error prone in its mechanism for patients to interpret the data. – [Problem 3 -P3]
Our Approach: Knowledge graph of drug side effects data using custom OWL relations 11/15/2018 Proposes to drastically simplify the syntax burden on DARs using the custom relations as explained in “Extending OWL to support custom relations” as well as providing a patient focused mechanism to extract the DARs dynamically. Relies heavily on linking the drug (component) vs drug class (compound) and drug vs side effects using custom OWL relationship based approach suggested in as well as providing a proof of concept application to explain DARs extraction process. Strives to prove how the DARs data can be represented in knowledge graph and demonstrate that it brings out meaningful interpretations.
Foundational Knowledge Representation Unit 11/15/2018
Drugs and Classes - knowledge representation using custom relationships. Class hierarchy dynamic relationship 11/15/2018
Pace protégé – with custom relationships 11/15/2018
Knowledge Graph - Drug Side Effects defined 11/15/2018
Scenarios of Doctor’s action during Patient visit for prescription 11/15/2018
D-GPR (Get Parents Relations) – An extension to the current capability of PaceJena 11/15/2018
In high level, D-GPR algorithm is designed in the following sequence 1) Input Drug Name D1 2) For the Drug D1, Enter a loop and Look for the Side effects S1-Sn 3) Then for Drug D1; Look for the Parent P1 4) For the Parent P1, Enter a loop and Look for the Side effects S11-Snn 5) Combine the Side effects S1-Sn and S11-Snn into a single set 6) Continue the Iteration until there are no more parents or siblings found 7) Output a combined Side effects list to display 11/15/2018
Drug Side effects identified using D-SERI model and D-GPR Algorithm 11/15/2018
Contribution Develop patient focused knowledge representation for drug adverse data Use custom relationships instead of object properties to extract meanings dynamically Derive drug adverse meanings quickly for doctors and patients to avoid human interpretation errors. Build knowledge graph framework which can be further extended to other drug domain information. Support adverse event “Predictive” modelling using data mining techniques. 11/15/2018
Questions 11/15/2018
Thank You 11/15/2018