Parsing HL7 v2 Messages for OpenMed

HL7 v2.x is the workhorse of hospital interfacing — ADT (admit/discharge/ transfer), ORU (observation results), MDM (document management), and ORM (orders) messages flow continuously between EHR, lab, radiology, and ancillary systems. The clinical narrative you want for NLP is buried in OBX-5 (observation value) and NTE-3 (notes/comments) fields, wrapped in a pipe-and-caret encoding. This skill decodes that envelope and hands the free text to OpenMed.

When to use

• You receive HL7 v2 messages from an interface engine (Mirth/NextGen Connect, Rhapsody, Cloverleaf) and want to mine embedded note/result text.
• A lab feed (ORU^R01) carries impression/comment narrative in OBX/NTE.
• An MDM^T02 transcription message carries a full report in OBX-5.
• You need a de-identified, structured feed into openmed.analyze_text.

HL7 v2 structure in one minute

A message is segments separated by \r (carriage return). Each segment is 3-letter-named, then fields split by |, components by ^, repetitions by ~, sub-components by &, with \ as escape. The encoding characters are declared in MSH-1 (the field separator) and MSH-2 (^~\&). Field positions are one-based, and MSH is special: MSH-1 is the separator, so MSH-2 is the first real field.

MSH|^~\&|LAB|HOSP|EHR|HOSP|20240302101500||ORU^R01|MSG0001|P|2.5
PID|1||MRN12345^^^HOSP^MR||DOE^JANE^Q||19700115|F|||1 FAKE ST^^SPRINGFIELD^IL^62704
OBR|1||ORD9|CBC^Complete Blood Count
OBX|1|TX|IMPRESSION||Mild leukocytosis; clinically correlate.||||||F
NTE|1||Patient reports fatigue x1 week. Dr. Smith notified.

Quick start

Parse the envelope and pull narrative from OBX-5 / NTE-3, then hand off:

import openmed
from openmed.interop.hl7v2 import parse_hl7v2

raw = open("results.hl7", encoding="utf-8").read()
msg = parse_hl7v2(raw)               # -> HL7Message (segments preserved)

narrative_chunks = []
for seg in msg.segments:
    if seg.name == "OBX":
        # OBX-2 is the value type; OBX-5 is the observation value.
        value_type = seg.get_field(2)
        if value_type in {"TX", "FT", "CE", "ST"}:
            narrative_chunks.append(seg.get_field(5) or "")
    elif seg.name == "NTE":
        narrative_chunks.append(seg.get_field(3) or "")

# Decode component delimiters into plain text before NLP.
flat = "\n".join(c.replace("^", " ").replace("&", " ") for c in narrative_chunks if c)

# Hand the narrative to OpenMed.
deid = openmed.deidentify(flat, method="replace", policy="hipaa_safe_harbor")
result = openmed.analyze_text(deid.text, output_format="dict")

HL7Segment.get_field(position) uses one-based HL7 positions and returns None for absent fields. HL7Message.segment_names() lists segments in order.

Whole-message segment-aware de-identification

When you need to redact the entire message (structured PID/NK1/GT1 fields and OBX/NTE free text) while preserving HL7 framing, use the bundled redactor instead of hand-rolling it:

from openmed.interop.hl7v2 import redact_hl7v2

safe = redact_hl7v2("results.hl7")   # path or message text
# PID-3 hashed, PID-5 name surrogated, PID-7 DOB date-shifted, OBX-5/NTE-3
# free text masked via openmed.deidentify — delimiters and segment order kept.

redact_hl7v2 applies DEFAULT_FIELD_MAP (PID, PD1, NK1, GT1, IN1/IN2, OBX, NTE). Extend or override it with field_map={("ZPS", 4): {"action": "hash"}} for site-specific Z-segments, and pass date_shift_days= for a fixed, interval-preserving shift.

Workflow

1. Frame-split safely. Real feeds use \r, \r\n, or MLLP framing (\x0b…\x1c\r). parse_hl7v2 auto-detects the segment separator; strip MLLP control bytes before parsing.
2. Read encoding from MSH — never assume |^~\&. The adapter derives the delimiter set from MSH-1/MSH-2 (HL7V2Encoding.from_msh_segment).
3. Locate narrative. OBX-5 (gated by OBX-2 value type), NTE-3, and report-bearing segments. Concatenate repetitions (~) and components (^).
4. De-identify, then analyze with OpenMed.
5. Rejoin results to the patient/encounter via PID-3 (patient id) and PV1-19 (visit number) — but redact those identifiers in anything you persist.

Hand-off to / from OpenMed

• To OpenMed: flattened OBX-5/NTE-3 text → openmed.deidentify → openmed.analyze_text.
• Adapter: openmed.interop.hl7v2 provides parse_hl7v2, redact_hl7v2, HL7Message, HL7Segment, HL7V2Encoding, HL7FieldRule, and DEFAULT_FIELD_MAP for segment-aware de-id that preserves message framing. It is parse-and-redact only — not a conformance validator.
• Re-link by id, not by PHI: carry PID-3/PV1-19 as keys, but store hashed or surrogate values (the default redact_hl7v2 hashes PID-3).

Edge cases & gotchas

• MLLP wrapper. Messages off a TCP MLLP listener are framed with \x0b (start) and \x1c\r (end). Strip these before parse_hl7v2.
• Escape sequences. \F\, \S\, \T\, \R\, \E\ encode literal delimiters, and \.br\ is a line break inside OBX text. Unescape before NLP.
• Repeating OBX. A single result can span many OBX segments (one line each); reassemble in order before summarizing.
• Value types matter. Only treat OBX-5 as narrative when OBX-2 is a text type (TX, FT, ST, CE); numeric (NM) and coded-only values are not free text. The default redactor restricts free-text redaction to FT/TX.
• Z-segments. Site-defined Z* segments often carry extra PHI; add explicit field_map rules — they are not in the default map.
• Versions vary. v2.3 through v2.8 differ in field cardinality; resolve positions against MSH-12 (version id), don't hardcode across versions.

Standards & references

• HL7 v2 product overview: https://www.hl7.org/implement/standards/product_brief.cfm?product_id=185
• HL7 v2.5.1 message structures (ADT, ORU, MDM) — base reference: https://hl7-definition.caristix.com/v2/
• MLLP (Minimal Lower Layer Protocol) transport: https://www.hl7.org/documentcenter/public/wg/inm/mllp_transport_specification.PDF
• OBX/NTE segment definitions: https://hl7-definition.caristix.com/v2/HL7v2.5.1/Segments/OBX