Extracting lab tables from documents and scans

Lab results arrive as tables: a column of test names, a value column, units, a reference range, and an abnormal flag (H/L/Crit). To use them downstream you must recover that grid from a PDF, scan, or spreadsheet into clean rows, then code each test to LOINC, normalize units with UCUM, and flag abnormals.

This skill is the intake step: it OCRs/parses the table on-device with openmed.multimodal, de-identifies any embedded PHI, and emits structured rows. It pairs before OpenMed's clinical helpers, the LOINC/UCUM coding and the high/low/critical flag are downstream (see parsing-lab-values).

When to use

• You have a lab report as a scanned image / photo / PDF page and need the panel as rows, not pixels.
• The source is a CSV/TSV export and you need columns classified (which is the value, the unit, the range, the flag) and PHI columns redacted.
• You need machine-readable rows to feed LOINC mapping and a FHIR Observation/DiagnosticReport.

What OpenMed gives you here

openmed.multimodal ships the intake primitives (no heavy deps at import; the OCR backend loads lazily):

• openmed.multimodal.ocr.ocr(image, engine=...) → an OcrResult whose .words are OcrWord(text, bbox, confidence, page) and .text is the joined string. OcrResult.to_document() bridges each word (with its pixel bbox) into an ExtractedDocument so detected PHI can project back to the source location.
• read_table(...) → a TableView (headers, rows, delimiter, has_header, columns) for delimited text; classify_columns(...) labels each column; redact_table(...) → a RedactedTable with a PHI-safe manifest.

Engines: Tesseract (pip install "openmed[multimodal]" + the system binary) or PaddleOCR (pip install "openmed[ocr-paddle]"). ocr() auto-selects the first installed backend.

Quick start