Enforcing No-PHI Logging

Logs are a top breach vector: a clinical string lands in a log line, gets shipped to a centralized log store and an error tracker, and is now PHI sitting outside the de-id boundary. OpenMed's local-first stance says no raw PHI in logs, caches, or error reports — this skill enforces it with a redaction guard that runs before any record is emitted.

When to use this skill

• An OpenMed service logs request text, model output, or exception messages.
• You ship logs/traces to a centralized store or error tracker (Sentry, ELK).
• You need a logging.Filter (or OTel processor) that redacts PHI pre-emit.
• You want structured, no-PHI log fields (offsets, hashes, counts) for debugging.

Quick start — a redacting logging.Filter

import logging
import re
import openmed

# Cheap regex pre-filter for the highest-risk structured identifiers. This runs
# on every record, so keep it fast; the model is the fallback for free-text PHI.
_FAST_PATTERNS = [
    (re.compile(r"\b\d{3}-\d{2}-\d{4}\b"), "[SSN]"),
    (re.compile(r"\b\d{16}\b"), "[CARD]"),
    (re.compile(r"\b[\w.+-]+@[\w-]+\.[\w.-]+\b"), "[EMAIL]"),
    (re.compile(r"\b(?:\+?\d[\d().\-\s]{7,}\d)\b"), "[PHONE]"),
]

class NoPHIFilter(logging.Filter):
    """Redact PHI from a log record before it is emitted. Fail closed."""

    def __init__(self, model_name: str | None = None, use_model: bool = True):
        super().__init__()
        self.model_name = model_name
        self.use_model = use_model

    def filter(self, record: logging.LogRecord) -> bool:
        try:
            message = record.getMessage()
            record.msg = self._scrub(message)
            record.args = ()                 # message already rendered & scrubbed
        except Exception:
            # Never let the logger leak on error — drop the message, keep the level.
            record.msg = "[REDACTED: scrub error]"
            record.args = ()
        return True                          # keep the (now-clean) record

    def _scrub(self, text: str) -> str:
        for pattern, tag in _FAST_PATTERNS:
            text = pattern.sub(tag, text)
        if not self.use_model:
            return text
        # Model fallback for free-text PHI (names, locations, dates). Replace by
        # offset, right-to-left, so earlier offsets stay valid.
        spans = openmed.extract_pii(text, model_name=self.model_name) \
            if self.model_name else openmed.extract_pii(text)
        for e in sorted(spans.entities, key=lambda s: s.start, reverse=True):
            text = text[:e.start] + f"[{e.label}]" + text[e.end:]
        return text

# Attach to every handler that might emit clinical text.
handler = logging.StreamHandler()
handler.addFilter(NoPHIFilter(model_name="OpenMed/Privacy-PII-Detection"))
logging.getLogger("openmed.service").addHandler(handler)

Prefer structured, no-PHI fields

Don't log the note and scrub it — log about it without the text in the first place:

logger.info(
    "deidentified note",
    extra={
        "doc_id": doc_id,                       # opaque id, not the text
        "phi_entity_count": len(result.entities),
        "phi_labels": sorted({e.label for e in result.entities}),
        "char_len": len(text),
        # offsets/hashes for debugging; never the plaintext span
        "phi_offsets": [(e.start, e.end) for e in result.entities],
    },
)

Redaction is the safety net; not logging PHI is the actual fix.

OpenTelemetry / error trackers

• Spans: add a SpanProcessor.on_end (or attribute hook) that runs the same _scrub over string span attributes and events before export.
• Error trackers: register a before_send hook (e.g. Sentry) that scrubs exception messages, breadcrumbs, and request bodies. Stack traces often embed the offending input — scrub the message, not just the frames.

Workflow

1. Inventory sinks. List every place a clinical string can reach: app logs, access logs, OTel spans, error tracker, crash reports, request/response dumps.
2. Install the regex pre-filter for structured identifiers (SSN, card, email, phone) — fast, runs on every record.
3. Add the model fallback (openmed.extract_pii) for free-text PHI on the sinks that carry clinical narrative; skip it on hot paths where regex suffices.
4. Switch to structured fields. Replace "log the text" with "log counts, labels, offsets, ids".
5. Fail closed. On any scrub error, drop the message content, not the redaction.
6. Test it. Unit-test that known PHI strings never survive a round trip through the filter, including in exception messages.

Hand-off to / from OpenMed

• Uses openmed.extract_pii (and optionally the regex pre-filter) as the PHI detector — the same engine documented in extracting-pii-entities.
• From building-with-openmed: this is the runtime guard for the local-first, no-PHI-in-artifacts rule.
• Pairs with gating-deid-leakage: the gate proves the model doesn't leak; this guard proves your logs and traces don't leak.
• To auditing-deidentification-runs: route audit output through the same no-PHI discipline (offsets/hashes, never plaintext).

Edge cases & gotchas

• record.args must be cleared after scrubbing. If you rewrite record.msg but leave %s args, the formatter re-injects raw PHI downstream.
• Scrub before fan-out. Filters on one handler don't protect others — attach to every handler, or scrub at the record/formatter layer.
• Latency budget. The model fallback costs inference per record; gate it behind a level threshold or reserve it for narrative-bearing sinks.
• Regex alone is not de-id. It catches structured identifiers; names, locations, and dates need the model. Use both, model last.
• Exception messages are PHI carriers. f"failed on {note}" leaks; scrub exception text and error-tracker payloads, not just logger.info calls.
• Fail closed, never open. A scrub error must redact the content, never emit the unscrubbed original.
• No raw PHI even in DEBUG. "It's only debug logs" is how breaches happen; the guard applies at every level.

Standards & references

• HIPAA Security Rule — audit controls & log protection (45 CFR 164.312): https://www.hhs.gov/hipaa/for-professionals/security/laws-regulations/
• OWASP Logging Cheat Sheet (what not to log): https://cheatsheetseries.owasp.org/cheatsheets/Logging_Cheat_Sheet.html
• OpenTelemetry span processors (attribute redaction on export): https://opentelemetry.io/docs/specs/otel/trace/sdk/#span-processor
• Python logging.Filter API: https://docs.python.org/3/library/logging.html#filter-objects
• OpenMed PHI detector: openmed.extract_pii (openmed/core/pii.py).