Exploring The Impact of Motif-Driven Causal Temporal Analysis Using Graph Neural Network in Improving Large Language Model Performance for Pharmacovigilance

Pharmacovigilance analytics can lower the probability of negative outcomes when choosing a route of therapy by identifying pharmacological adverse effects in the past. Despite the abundance of resources such as medical reports, review websites, online platforms, etc. available for adverse signal detection, it has proven challenging to manually extract the pertinent data and relate the medicine to its side effects both when there is the prevalence of mass adverse effect from the same drug in various documents and areas. We can quickly identify the gap and propose a causal temporal motif-driven large language model to identify and extract the causal association between the medication and the events based on but not limited to the commencement of the event and ignoring the prior reaction which may be due to some other reason revealing is challenging and the current state of the art still relies on Named Entity Recognition (NER) approaches. These NER-based approaches also experience cold start issues, in which new side effects remain isolated and are not connected to the drug from which they originated. We propose an approach that uses a causal temporal graph neural network including motif embedding using local linear embedding under Centrality constrain followed by a motif-based similarity search for finding similar subgraph and causal relationships between the drug and the adverse events.