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

from openmed.multimodal.ocr import ocr
from openmed.multimodal import read_table, classify_columns, redact_table

# A) Scanned / image lab report -> words with pixel boxes.
result = ocr("cbc_report.png")               # OcrResult
for w in result.words[:5]:
    print(repr(w.text), w.bbox, round(w.confidence, 2), "p", w.page)

doc = result.to_document()                   # ExtractedDocument; bbox preserved

# B) Delimited lab export (CSV/TSV) -> classified, PHI-redacted rows.
csv_text = (
    "PatientName,Test,Value,Unit,RefRange,Flag\n"
    "Jane Roe,Hemoglobin,9.1,g/dL,12.0-15.5,L\n"
    "Jane Roe,Glucose,148,mg/dL,70-99,H\n"
)
view = read_table(csv_text)                  # TableView
view = classify_columns(view)                # tag PHI vs data columns
redacted = redact_table(view)                # RedactedTable: PatientName redacted

for row in redacted.rows:
    print(row)                               # name column masked; lab data intact
for col in redacted.manifest:                # PHI-safe per-column audit manifest
    print(col["column_name"], col["assigned_class"], col["action"])

For an OCR'd (image) table, you reconstruct the grid yourself from word boxes (next section) — OCR yields positioned words, not a delimited table.

Workflow

1. Detect the source type. CSV/TSV → read_table. Image/scan → ocr(). PDF/DOCX are not directly parseable (they raise UnsupportedDocumentError); render PDF pages to images first, or extract their text layer, then OCR.
2. OCR with positions. ocr() returns OcrWords carrying bbox and page. Keep the boxes — they let you cluster words into rows/columns and project PHI redaction back to pixels.
3. Reconstruct the grid. Cluster words by their bbox y into rows, by x into columns. The header row names the columns; align body cells to those x bands. Confidence (OcrWord.confidence) flags shaky cells for review.
4. Identify the lab columns. Map headers to roles: test name, value, unit, reference range, flag. For delimited input, classify_columns tags PHI columns (name/MRN/DOB) so redact_table masks them.
5. De-identify embedded PHI. Patient name/MRN often sit in the table header or a leading column. Redact those columns (redact_table) and run free-text cells through openmed.deidentify before the rows leave the device.
6. Emit structured rows {test, value, unit, ref_range, flag} per result and hand off to LOINC/UCUM coding and parsing-lab-values.

Hand-off to / from OpenMed

• To parsing-lab-values (openmed.clinical.parse_reference_range, derive_abnormal_flag): pass the parsed value + ref_range (+ any explicit lab flag) to get a structured low/normal/high/critical signal.
• To mapping-loinc: code each test name to a LOINC code; normalize the unit with UCUM. OpenMed emits the row; the terminology binding is out-of-process.
• To FHIR (exporting-to-fhir): each row becomes an Observation (code=LOINC, valueQuantity with UCUM unit, referenceRange, interpretation) grouped under a DiagnosticReport.
• De-identify with deidentifying-clinical-text (openmed.deidentify) before export. OCR words carry pixel boxes so redaction maps back to the image.
• Everything here runs on-device; no scan or row leaves the process un-de-identified.

Edge cases & gotchas

• PDF/DOCX raise UnsupportedDocumentError. The multimodal dispatcher has no PDF/DOCX handler — rasterize PDF pages to PNG (or pull the text layer) before calling ocr(). Image formats (PNG/JPG/TIFF/…) and CSV/TSV are handled.
• OCR returns words, not a table. You must reconstruct rows/columns from bbox geometry. Multi-line cells, wrapped test names, and merged header cells break naive x/y bucketing — tune the clustering tolerance per template.
• Reference ranges are easy to mis-split. "12.0-15.5", "<5", "70 - 99", and en/em dashes must survive OCR and tokenization as one cell. Don't let a space or a misread dash fracture the range — parse_reference_range downstream expects it whole.
• Units belong to the value, not the range. Keep "9.1 g/dL" and the range "12.0-15.5" in separate fields; the value's unit must match the range's unit or the abnormal flag will be wrong (the flag helper is unit-agnostic).
• Low-confidence cells. Gate on OcrWord.confidence; a 0.4-confidence value in a lab table is a patient-safety risk — route it to human review, don't silently accept it.
• PHI hides in tables. Patient name, MRN, DOB, and accession numbers commonly occupy the header or first column. Classify and redact them; never log the raw table.
• Engine availability. ocr() raises a clear MissingDependencyError if no backend is installed — install Tesseract or PaddleOCR per the extras.

Standards & references

• LOINC — universal lab observation codes: https://loinc.org/
• UCUM — Unified Code for Units of Measure: https://ucum.org/
• HL7 FHIR R4 Observation (valueQuantity, referenceRange, interpretation): https://hl7.org/fhir/R4/observation.html
• HL7 FHIR R4 DiagnosticReport (lab grouping): https://hl7.org/fhir/R4/diagnosticreport.html
• Tesseract OCR: https://github.com/tesseract-ocr/tesseract
• PaddleOCR: https://github.com/PaddlePaddle/PaddleOCR
• OpenMed source: openmed/multimodal/ocr.py, openmed/multimodal/tabular_csv.py.