Before
Every claim was checked by hand, one at a time. Reviewers had no quick way to see whether an expensive test or procedure matched the doctor's prescription or the patient's history. Claims that didn't add up were often paid anyway.
What we built
A proof of concept that links each claim to the patient's medical records, history and prescriptions. Business rules and a scoring model compare what was claimed with what the record supports, and flag any claim where the test or procedure doesn't fit the diagnosis, the prescription or the patient's history.
What we found
Most unsupported claims weren't hidden. They were simply never compared with the patient's record. Joining the data was what made them visible.
After
- In testing on historical claims, the model caught around 30% of the suspicious claims before payout.
- Reviewers spend their time on the riskiest claims instead of checking every one.
- Each flag shows why: the claimed test next to the prescription and history it doesn't match.
Stack
- Databricks
- Python
- SQL
- Logistic regression
Illustrative