PharmaSentinel

Drug Adverse Event Detection Pipeline

An end-to-end NLP pipeline that ingests FDA FAERS adverse event reports and classifies drug events by clinical severity. Built for cloud-native operation on ECS Fargate with a full CloudFormation footprint.

Methods & contributions

• FAERS-native ingestion — parses the authentic ICHICSR schema and pipe-delimited quarterly extracts (primaryid, drugname, reaction PT, outcome codes).
• Severity classifier — TF-IDF features feeding a logistic-regression model with stratified cross-validation and feature-importance attribution.
• Serverless data plane — S3 data lake, SQS event routing, and auto-scaling Fargate inference services.
• Observability — Datadog APM tracing across the pipeline with infrastructure defined entirely in CloudFormation.

Data flow

FAERS extract → S3 raw → ingestion (SQS) → NLP classifier (FastAPI) → severity store
                                  ↘ Datadog APM / CloudWatch

Technologies