PPTX Slide and Speaker-Notes Redaction¶

OpenMed can extract text from PowerPoint slide shapes, table cells, and speaker notes through the shared multimodal document contract. The extractor returns a normalized text stream plus run-level provenance, so detections can be mapped back to the presentation without placing the original PHI in audit metadata.

Install the multimodal extra first:

uv pip install ".[multimodal]"

Extract text and offsets¶

from openmed.multimodal import extract_pptx

document = extract_pptx("clinical-case.pptx")

print(document.text)
for source in document.spans:
    print(
        source.page,
        source.metadata["part"],
        source.metadata["slide_start"],
        source.metadata["slide_end"],
    )

SourceSpan.page is the zero-based slide index. Each span records whether the text came from a slide shape, table cell, or speaker notes, together with its paragraph/run location and character offsets within that slide. The document metadata also includes a slide_offsets entry for every slide, including empty slides.

Project detections and write a redacted copy¶

Use offsets from document.text when supplying detected spans:

from openmed.multimodal import (
    map_text_spans_to_pptx_runs,
    write_redacted_pptx,
)

name = "Jane Doe"
start = document.text.index(name)
entities = [
    {
        "start": start,
        "end": start + len(name),
        "label": "PERSON",
    }
]

provenance = map_text_spans_to_pptx_runs(document, entities)
write_redacted_pptx(
    "clinical-case.pptx",
    "clinical-case.redacted.pptx",
    entities,
)

The default replacement is label-based, such as [PERSON]. A span may cross multiple styled runs; OpenMed inserts the replacement into the first covered run and removes the covered text from the remaining runs. The returned provenance contains offsets, labels, confidence values, source locations, and a SHA-256 digest of the removed text, but not the plaintext PHI.

Use the multimodal dispatcher¶

The .pptx handler is registered when openmed.multimodal is imported:

from openmed.multimodal import redact_document

result = redact_document(
    "clinical-case.pptx",
    models={"detector": detector},
    policy={"output_path": "clinical-case.redacted.pptx"},
    lang="en",
)

Without an output path, the dispatcher extracts text and returns projected redaction provenance without writing a presentation.

Scope and limitations¶

• Text in slide shapes, grouped text shapes, table cells, and speaker notes is supported.
• Embedded images are not inspected; use the image/OCR redaction pipeline for image-based PHI.
• Write-back preserves the deck structure and unaffected run formatting, but exact layout fidelity after replacement is not guaranteed.
• SmartArt, charts, OLE objects, comments, masters, and layout-template text are outside this adapter's scope.