Healthcare provider (anonymized)
Survey-to-AI Pipeline for Patient Insights
From survey to patient insight in minutes — an AI pipeline that processes responses in real time.
- AI classification accuracy, validated vs. hand-coded responses
- >90%
- AI classification accuracy, validated vs. hand-coded responses
The starting point
A healthcare provider surveyed patients after every visit, but the gap from response to action was weeks. Open-ended comments — the most valuable part — went unread, so a spike in wait-time complaints at one department might not surface for six weeks, long after the cause had passed and institutional memory had walked out the door.
The method
We built an end-to-end pipeline with no manual step: SurveyMonkey webhooks triggered AWS Lambda on each response, GCP Cloud Functions ran ChatGPT-API classification into clinical categories and sentiment, and results landed in Snowflake. We validated the classifier against 200 hand-coded responses past 90% accuracy, flagged low-confidence cases for human review, and fed reviewer corrections back into the prompts so accuracy improved without retraining.
The result
Insights that took weeks now reached department heads in hours, and the once-abandoned open-ended analysis became the most-valued feature. Department heads got rolling 7-day sentiment with drill-down; leadership got monthly trends with significance testing. The provider later extended the same infrastructure to post-procedure and staff surveys — feedback finally moving at the pace of care.