Healthcare Adverse Event Detection Using Amazon SageMaker and Large Language Models

Pharmaceutical customers faced a challenge of automating detection of adverse drug events from diverse sources like social media, calls, emails, and handwritten notes to support pharmacovigilance efficiently and cost-effectively. The solution involved fine-tuning transformer-based large language models using Amazon SageMaker and Hugging Face Transformers on the Adverse Drug Reaction dataset, employing synthetic data generation with Falcon models to address data skew, and ensuring HIPAA compliance with encryption. This ML solution improved classification accuracy of adverse events, enabled automated real-time detection supporting safer pharmacovigilance workflows for unnamed pharmaceutical customers, and was implemented using AWS Cloud Development Kit.

Industry
Pharma
Published
February 2024

Reported outcomes

Strategic outcomes

Risk & complianceBuilt a HIPAA-compliant detection systemSpeed & agilityAutomated real-time adverse event detectionBetter decisions & insightImproved pharmacovigilance monitoring reliability

Primary read

Use case focus

Showing 3 of 3

  • 1Healthcare NLP
  • 2Pharmacovigilance
  • 3ML model fine-tuning
Manual processing of adverse drug events from multiple unstructured data sources was slow, costly, and error-prone, limiting effective pharmacovigilance for pharmaceutical customers.
  • Fine-tuned multiple transformer NLP models on Amazon SageMaker using Hugging Face pre-trained models specialized in medical text like BioBERT.
  • Used synthetic data generation via Falcon models to augment training data and reduce class imbalance.
  • Implemented the solution in a HIPAA-compliant manner using the AWS Cloud Development Kit for deployment automation.
  • Built a scalable, HIPAA-compliant machine learning system that automates adverse event detection with high classification accuracy.
  • Improved efficiency of pharmacovigilance monitoring, enabling pharmaceutical customers to identify and respond to issues faster and more reliably.
Architecture

Used SageMaker PyTorch estimator and SageMaker JumpStart with Hugging Face models. Employed Falcon LLMs for synthetic data generation to mitigate class imbalance. Deployed via AWS Cloud Development Kit.

Sources & evidence1
Groundedness: 3/5

AI-generated summary. Verify important details with the linked sources before relying on this case.

Explore related AI use cases
This website uses cookies to enhance the user experience. Learn more.