9:20 AM - 9:40 AM
[2L1-OS-25-02] Automatic Polypharmacy Detection Applying AI and Knowledge Graph
Keywords:polypharmacy, knowledge graph, medicine
Background
Japan is transitioning to an aging society, and the number of prescribed drugs per capita is increasing accordingly. There is limited time to determine polypharmacy in the outpatient setting. We report on an algorithm for detecting drug interactions based on the content of medications taken.
Subjects and Methods
From the database of ethical drug inserts of the Pharmaceuticals and Medical Devices Agency (PMDA), 4505 types and 19081 drugs were targeted. We extracted the part related to interactions, and the LLM was applied to extract 1,466 words that appeared in those interactions. We finally visually selected 1232 words that we considered necessary. A graph neural network was constructed with each drug and word as a node and connected by edges when relevant. When each drug was entered, the interaction was judged based on whether or not the drugs were connected to each other within two neighboring nodes. If an interaction was determined to exist, LLM determined the mechanism of action of each drug and output it as a sentence.
Results
For multiple drugs, the mechanism of action of each drug was automatically output as a sentence. The output time was about 10 seconds. The interaction of each drug can be automatically verified by simply inputting the drugs. The time required for verification was also sufficient for clinical use.
Japan is transitioning to an aging society, and the number of prescribed drugs per capita is increasing accordingly. There is limited time to determine polypharmacy in the outpatient setting. We report on an algorithm for detecting drug interactions based on the content of medications taken.
Subjects and Methods
From the database of ethical drug inserts of the Pharmaceuticals and Medical Devices Agency (PMDA), 4505 types and 19081 drugs were targeted. We extracted the part related to interactions, and the LLM was applied to extract 1,466 words that appeared in those interactions. We finally visually selected 1232 words that we considered necessary. A graph neural network was constructed with each drug and word as a node and connected by edges when relevant. When each drug was entered, the interaction was judged based on whether or not the drugs were connected to each other within two neighboring nodes. If an interaction was determined to exist, LLM determined the mechanism of action of each drug and output it as a sentence.
Results
For multiple drugs, the mechanism of action of each drug was automatically output as a sentence. The output time was about 10 seconds. The interaction of each drug can be automatically verified by simply inputting the drugs. The time required for verification was also sufficient for clinical use.
Authentication for paper PDF access
A password is required to view paper PDFs. If you are a registered participant, please log on the site from Participant Log In.
You could view the PDF with entering the PDF viewing password bellow.