Schedule of Activities Extractor

The problem

Clinical-trial protocols hide wide, multipage Schedule of Activities tables inside long PDFs. Flattening them to text loses hierarchy, footnote links, ambiguous cells, and page provenance—the exact details a downstream clinical workflow needs to trust.

What I built

I built a Python pipeline and upload UI that locates candidate pages, renders them for multimodal extraction, deterministically stitches multipage tables, and emits a graph schema that preserves headers, verbatim cell values, footnotes, provenance, and unresolved ambiguity.

Location and extraction are separate problems

A model cannot extract a table it never sees. The first stage finds likely Schedule of Activities pages across the entire protocol, and the extraction stage runs only on those images. Keeping the stages separate made it possible to measure page recall independently and reach 12/12 on the reference set and 8/8 on unseen holdouts.

Deterministic stitching protects the source

The table may repeat headers, continue rows across pages, or move footnotes to the final fragment. The stitcher joins extracted fragments with deterministic rules instead of prompting a model to summarize them, which preserves verbatim cell values and keeps every merge inspectable.

Ambiguity is data

The graph schema records unresolved structures and ambiguous cells rather than silently forcing them into a convenient shape. It also retains hierarchical headers and footnote targets. That makes the output suitable for review-heavy clinical workflows where an explicit unknown is safer than a clean but invented answer.