Spaces:
Sleeping
Sleeping
File size: 8,963 Bytes
60d4850 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 | """
NER-Based PHI Detection using Microsoft Presidio
Extends the existing regex-based PHI detection in compliance.py with
ML-powered Named Entity Recognition for detecting:
- Patient names (not detectable by regex alone)
- Addresses and geographic locations
- Ages over 89 (HIPAA Safe Harbor)
- Contextual identifiers that regex patterns miss
Architecture:
Presidio (primary, ML-based) -> regex fallback (existing patterns)
The detector is loaded lazily to avoid startup overhead when not needed.
"""
from __future__ import annotations
import logging
from dataclasses import dataclass, field
from typing import List, Optional
from app.config import settings
logger = logging.getLogger(__name__)
# Presidio entity types mapped to HIPAA Safe Harbor categories
_HIPAA_ENTITY_TYPES = [
"PERSON", # Names
"PHONE_NUMBER", # Phone numbers
"EMAIL_ADDRESS", # Email addresses
"US_SSN", # Social Security Numbers
"LOCATION", # Addresses, geographic locations
"DATE_TIME", # Dates (DOB, admission, discharge)
"IP_ADDRESS", # IP addresses
"US_DRIVER_LICENSE", # Driver's license
"MEDICAL_LICENSE", # Medical license numbers
"URL", # URLs that may contain PHI
"NRP", # Nationality/religious/political group
"AGE", # Ages (>89 are identifiers under Safe Harbor)
]
# Minimum confidence score for a detection to be included
_MIN_CONFIDENCE = 0.4
@dataclass
class PHIDetection:
"""A single PHI detection result."""
entity_type: str
start: int
end: int
score: float
text_snippet: str # First 4 chars + *** for audit logging
@dataclass
class PHIScanResult:
"""Result of a comprehensive PHI scan."""
detections: list[PHIDetection] = field(default_factory=list)
is_clean: bool = True
detection_count: int = 0
entity_types_found: list[str] = field(default_factory=list)
method: str = "regex" # "presidio", "regex", or "hybrid"
@property
def phi_count(self) -> int:
return self.detection_count
class PHIDetector:
"""PHI detection service combining Presidio NER with regex fallback.
Usage:
detector = get_phi_detector()
result = detector.scan("Patient John Smith, DOB 01/15/1985")
redacted = detector.redact("Patient John Smith, DOB 01/15/1985")
"""
def __init__(self):
self._analyzer = None
self._anonymizer = None
self._presidio_available = False
self._initialized = False
def _init_presidio(self) -> bool:
"""Lazily initialize Presidio analyzer and anonymizer."""
if self._initialized:
return self._presidio_available
self._initialized = True
try:
from presidio_analyzer import AnalyzerEngine
from presidio_anonymizer import AnonymizerEngine
self._analyzer = AnalyzerEngine()
self._anonymizer = AnonymizerEngine()
self._presidio_available = True
logger.info("Presidio PHI detector initialized successfully")
except ImportError:
logger.warning(
"presidio-analyzer not installed. Falling back to regex-only PHI detection. "
"Install with: pip install presidio-analyzer presidio-anonymizer"
)
self._presidio_available = False
except Exception as e:
logger.warning(f"Presidio initialization failed: {e}. Using regex fallback.")
self._presidio_available = False
return self._presidio_available
def scan(self, text: str) -> PHIScanResult:
"""Scan text for PHI using Presidio (primary) + regex (fallback).
Args:
text: Text to scan for PHI.
Returns:
PHIScanResult with all detections.
"""
if not text:
return PHIScanResult()
detections: list[PHIDetection] = []
method = "regex"
# Try Presidio first
if self._init_presidio() and self._analyzer is not None:
method = "presidio"
try:
results = self._analyzer.analyze(
text=text,
entities=_HIPAA_ENTITY_TYPES,
language="en",
score_threshold=_MIN_CONFIDENCE,
)
for r in results:
snippet = text[r.start:r.start + 4] + "***" if r.end - r.start > 4 else "***"
detections.append(PHIDetection(
entity_type=r.entity_type,
start=r.start,
end=r.end,
score=r.score,
text_snippet=snippet,
))
except Exception as e:
logger.warning(f"Presidio scan failed, falling back to regex: {e}")
method = "regex"
# Always run regex as fallback/supplement
from app.compliance import detect_phi as regex_detect_phi
regex_detections = regex_detect_phi(text)
if regex_detections:
if method == "presidio":
method = "hybrid"
# Merge regex detections, avoiding duplicates (overlapping ranges)
presidio_ranges = {(d.start, d.end) for d in detections}
for rd in regex_detections:
rd_range = (rd["start"], rd["end"])
# Check for overlap with existing detections
overlaps = any(
not (rd_range[1] <= p[0] or rd_range[0] >= p[1])
for p in presidio_ranges
)
if not overlaps:
detections.append(PHIDetection(
entity_type=rd["pattern_type"],
start=rd["start"],
end=rd["end"],
score=1.0, # Regex matches are binary
text_snippet=rd["match"],
))
entity_types = list({d.entity_type for d in detections})
return PHIScanResult(
detections=detections,
is_clean=len(detections) == 0,
detection_count=len(detections),
entity_types_found=entity_types,
method=method,
)
def redact(self, text: str) -> str:
"""Redact all detected PHI from text.
Uses Presidio anonymizer if available, otherwise falls back to
the existing regex-based redaction in compliance.py.
Args:
text: Text containing potential PHI.
Returns:
Text with PHI replaced by redaction markers.
"""
if not text:
return text
# Try Presidio anonymizer first
if self._init_presidio() and self._analyzer is not None and self._anonymizer is not None:
try:
results = self._analyzer.analyze(
text=text,
entities=_HIPAA_ENTITY_TYPES,
language="en",
score_threshold=_MIN_CONFIDENCE,
)
if results:
anonymized = self._anonymizer.anonymize(text=text, analyzer_results=results)
redacted = anonymized.text
else:
redacted = text
# Second pass with regex for anything Presidio missed
from app.compliance import redact_phi_text
redacted = redact_phi_text(redacted)
return redacted
except Exception as e:
logger.warning(f"Presidio redaction failed, using regex: {e}")
# Regex-only fallback
from app.compliance import redact_phi_text
return redact_phi_text(text)
def redact_for_storage(self, text: str) -> tuple[str, PHIScanResult]:
"""Redact PHI and return both the clean text and a scan report.
Suitable for use before storing text in databases or vector stores.
Performs double-pass redaction for safety.
Args:
text: Text to redact.
Returns:
Tuple of (redacted_text, scan_result).
"""
redacted = self.redact(text)
# Verify the redacted text is clean
verification = self.scan(redacted)
if not verification.is_clean:
# Second pass
redacted = self.redact(redacted)
verification = self.scan(redacted)
return redacted, verification
# ---------------------------------------------------------------------------
# Singleton accessor
# ---------------------------------------------------------------------------
_detector: Optional[PHIDetector] = None
def get_phi_detector() -> PHIDetector:
"""Get or create the singleton PHI detector instance."""
global _detector
if _detector is None:
_detector = PHIDetector()
return _detector
|