PHUSE/FDA Data Science Innovation Challenge – Approach for Predicting Drug Interactions Challenge Abstracts

A current challenge of the PHUSE/FDA Data Science Innovation Challenge is 'Approach for Predicting Drug Interactions' which looks to overcome the limitations of traditional methods for mining post-market data sources or observational healthcare data to better predict drug interactions. Here you will find the abstracts of the accepted participants who have proposed a solution to this challenge.

Predicting Drug Interactions using Machine Learning and Data Visualisation
Submitted by: Jeffrey Philip

This proposal is based on the concept of developing a predictive model and data visualization solution that enables the identification of leading indicators of drugs likely to interact with other drugs, when administered in a pre-defined period.

De-identified patient-level data will be used together with a mixture of publicly available data sources and in-licensed real-world data databases. 


Prediction of Drug Interactions with Real-world Data
Submitted by: James Kim

This proposal intends to leverage real-world evidence data to extract relevant drugs or procedural information for a disease condition with the given target cohort definition.

An interactive tool will enable the user to visualize the findings using graphical displays, allowing them to identify disease conditions likely to be drug-drug interactions. Drug databases, such as DrugBank, could be used for verification of this methodology.


Real-world Evidence Drug-Drug Interaction (REDDI) Match
Submitted by: Kerry Deem

This proposal is intended to predict and/or identify side effects or adverse events (AEs) related to drug-drug interactions, using machine learning and pattern recognition models built on publicly available datasets combined with claims and electronic health record (EHR) data. 

The primary focus is on geriatric populations, patients undergoing immunosuppressive therapy and individuals receiving prescriptions from a high number of prescribers. 


Artificial Intelligence for Early Detection of Drug Interactions
Submitted by: Zhichao Liu

This proposal goes beyond predictive modeling using conventional machine learning methodologies, by implementing a deep-learning framework with autoencoder to predict DDIs through the integration of diverse biological drug profiles. 

It is expected for this solution to be used in early detection of DDIs.   


Approaches in Categorising and Predicting Drug-Drug Interactions using Artificial Intelligence and SAS Visual Text Analytics
Submitted by: Soundarya Palanisamy

This proposal is a hybrid approach, integrating the results from both rule-based and machine-learning methodologies. Factors such as renal and hepatic impairment, sex, age, concomitant intestinal and hepatic cytochrome P450 enzymes (CYP) will be used as features to the predictive machine learning models.

The fundamental idea is training Generative Adversarial Networks (GANs) for generator learning and creation of synthetic (artificial) molecules, based on the examples (from SMILES data) that the network sees in the training phase. This idea adopts a generative approach for predicting new DDIs.

Challenge Stream Chairs:

Beverly Hayes and Sanjay Sahoo

Posted by on

Categories: PHUSE News

Related Blogs

Add Your Comments

Thinking of joining PHUSE?

Already a member but not sure how you can benefit?

PHUSE is an expanding, global society with a global membership of clinical data scientists. It requires a large pool of resources to help with its running, and so there are many opportunities for members to become involved. Whether it's chairing a conference, presenting at an event, leading a working group or contributing to the quarterly online newsletter, we are always keen to hear from volunteers.

Find Out More