Digitization and analysis of medical reports.
A pipeline built to digitize, de-identify, and enrich medical reports with tags and KPIs. It is designed for a 600,000-document archive and an annual throughput of 200,000 reports. Interactive dashboards and a natural-language AI assistant complete the workflow.
The project, at a glance
200,000 documents a year that used to be processed by hand, one at a time.
The document becomes structured data: extraction and validation in a single flow.
Sensitive documents: direct identifiers are removed before downstream analysis.
Each lead is extracted and cleaned at image level, leaving an enhanced black-and-white trace. The system calculates angles and derivatives point by point, then checks the profile against percentile ranges derived from reports already validated in the archive. Each result is assigned to one of three groups: within range, borderline, or reject. Traces affected by poor electrode contact, noise, or other recording artifacts fall into the reject group.
Each report is enriched with tags and indicators, turning the archive into a queryable dataset.
Each report is classified with a confidence score: invalid, pediatric, or pathological. Staff retain final decision-making authority.
Interactive dashboards let staff explore the data and run natural-language queries.
How the system determines whether a trace is usable.
The reference distribution comes from reports the client has already validated, rather than manually chosen cutoffs. A trace proceeds only when its profile falls within the configured percentile band.
Each lead is cropped from the PDF at its original scale.
The grid and background are removed, leaving a black-and-white trace.
The system calculates the slope and normal at each point.
The profile is checked against percentiles from validated reports.
The workflow separates routine processing from the reports that need human review, reducing how much material staff must inspect manually.
Reports outside the expected ranges are isolated for manual validation rather than mixed into the automated flow.