""" src/nlp/ner.py ──────────────────────────────────────────────────────────────── Named Entity Recognition pipeline for clinical text. Architecture ──────────── A thin interface (BaseNERPipeline) sits in front of the actual implementation (SpacyNERPipeline). This makes the NER model swappable without touching the ETL, API, or dashboard. The scispaCy model (en_core_sci_lg) is trained on PubMed abstracts and the CRAFT corpus. It recognises biomedical entities but does not distinguish between sub-types out of the box. We apply a post-processing step that maps the raw entity labels to our five clinical categories: Raw scispaCy label → Our label ─────────────────────────────── DISEASE → DISEASE CHEMICAL → MEDICATION Any entity matching a procedure keyword list → PROCEDURE Any entity matching an anatomy term list → ANATOMY Remaining entities → SYMPTOM This mapping is deliberately simple and can be refined once real labelled data is available. Installation ──────────── pip install spacy scispacy pip install https://s3-us-west-2.amazonaws.com/ai2-s2-scispacy/ \ releases/v0.5.3/en_core_sci_lg-0.5.3.tar.gz ──────────────────────────────────────────────────────────────── """ from __future__ import annotations import re from abc import ABC, abstractmethod from dataclasses import dataclass from src.utils.config import ModelConfig from src.utils.logger import get_logger logger = get_logger(__name__) # ── Output dataclass ────────────────────────────────────────────── @dataclass class Entity: """A single named entity extracted from clinical text. Attributes: text : The extracted text as it appears in the note. label : Normalised entity type — one of DISEASE, MEDICATION, PROCEDURE, SYMPTOM, ANATOMY. start : Start character offset in the source text. end : End character offset in the source text. confidence : Model confidence score (0.0–1.0). None when the model does not provide scores. note_id : Optional database ID of the parent note. Populated when processing stored notes. """ text: str label: str start: int end: int confidence: float | None = None note_id: int | None = None def to_dict(self) -> dict: """Serialise to a plain dictionary for JSON responses. Returns: Dict with all fields; confidence is rounded to 3 d.p. if present. """ return { "text": self.text, "label": self.label, "start": self.start, "end": self.end, "confidence": round(self.confidence, 3) if self.confidence else None, "note_id": self.note_id, } # ── Label mapping ───────────────────────────────────────────────── # Fallback for raw labels that are unambiguous STRUCTURAL categories — # trusted directly if a model ever emits them, since en_core_sci_lg only # emits the generic "ENTITY" label and bc5cdr only emits DISEASE/CHEMICAL # in practice (both handled by dedicated logic in _normalise_label). # Deliberately excludes DISEASE, GENE, and SPECIES: those carry the same # false-positive risk that motivated gating DISEASE behind the ICD-10 # chapter lookup rather than trusting a raw label outright (see # _normalise_label's docstring for the full rationale). _SAFE_STRUCTURAL_LABEL_MAP: dict[str, str] = { "ANATOMY": "ANATOMY", "ORGAN": "ANATOMY", "CELL": "ANATOMY", "PROCEDURE": "PROCEDURE", } # Standalone fragments that are only meaningful as part of a longer phrase. # e.g. bc5cdr tags "mellitus" from "type 2 diabetes mellitus" as a separate span. _FRAGMENT_TERMS: frozenset[str] = frozenset({ "mellitus", "magna", "vera", "simplex", "complex", "syndrome", "disease", "disorder", "condition", }) # A span starting with one of these describes the ABSENCE of a finding # ("denies obesity", "no evidence of fracture") — reporting it as DISEASE # would claim the patient HAS the condition they explicitly don't. _NEGATION_PREFIXES: tuple[str, ...] = ( "denies", "denied", "no evidence of", "no history of", "without", "negative for", "ruled out", "r/o", "not consistent with", "absence of", ) # A span starting with one of these is a leaked section-header / reporting # fragment from an imperfect NER span boundary ("complaint of allergies"), # not the entity itself. _GENERIC_REPORT_PREFIXES: tuple[str, ...] = ( "complaint of", "complaints of", "complains of", "history of", "c/o", ) # Terms that are never clinically useful entities (section headers, # demographic words, generic document words, dosing qualifiers, # directional / generic anatomical qualifiers). _SKIP_TERMS: frozenset[str] = frozenset({ # demographics / section headers "patient", "patients", "male", "female", "man", "woman", "history", "impression", "assessment", "plan", "findings", "medications", "medication", "allergies", "allergy", "subjective", "objective", "note", "notes", "discharge", "summary", "report", # dosing qualifiers "once", "twice", "daily", "weekly", "twice daily", "once daily", "three times", "four times", "per day", "per week", "tab", "tabs", "tablet", "tablets", "capsule", "capsules", "mg", "mcg", "units", # generic / directional words that look like entities but aren't "area", "region", "site", "location", "level", "side", "right", "left", "right side", "left side", "bilateral", "proximal", "distal", "anterior", "posterior", "superior", "inferior", "medial", "lateral", # generic clinical document words "procedure", "procedures", "technique", "approach", "identified", "noted", "seen", "given", "placed", "performed", "stable condition", "stable", "normal", "negative", "positive", "anon", "anonymous", # vague clinical descriptors that appear as isolated entities "attention", "complication", "complications", "elevated", "day", "days", "week", "weeks", "month", "months", "local", "general", "incidental", "significant", "increased", "decreased", "moderate", "mild", "severe", "mild to moderate", "moderate to severe", "port", "head", "operative site", "operative", "postoperative", "intraoperative", "perioperative", # quality/appearance descriptors — not entities on their own "purulent", "serous", "serosanguinous", "sanguineous", "hemorrhagic", "haemorrhagic", "fibrinous", # negative findings — absence of a symptom, not a reportable entity "afebrile", "asymptomatic", "nontender", "non-tender", # bare adjectives / incomplete fragments (real usage always pairs # these with a noun, e.g. "undescended testis", "allergic reaction", # "diabetic neuropathy", "pitting edema", "H. pylori" with "H." split # off as a false sentence-end abbreviation) "allergic", "undescended", "diabetic", "pitting", "pylori", # bare type-qualifier split off from "type 1/2 diabetes mellitus" # etc. by NER span splitting — meaningless without the condition # name it modifies, and dangerous left standalone: bc5cdr/embedding # matching can map it to an unrelated condition ("type 2" -> # "Obesity, class 2") rather than recognising it as a fragment "type 1", "type 2", "type i", "type ii", # generic verbs / state words picked up as spurious entities — # found via gold-standard sample review (2026-06-26): these carry # no clinical meaning on their own ("the airway would improve", # "before she leaves", "any problem with her going home") "improve", "improves", "improved", "episodes", "infiltrated", "outside", "seated", "admission", "leaves", "problem", "problems", "instantaneous", "integrity", "unimproved", "insertion", "warm", "clear", "healthy-appearing", "years", "smoke", "smoking", # section-header fragments leaking through as bare entities "operation", "operations", "systems", # measurement units (already have mg/mcg/units; mmHg was missing) "mmhg", # equipment / supply names, not clinical entities "cuff", "webril", "hand reamer", # lab-test / panel names and social-history substances that bc5cdr # also tags CHEMICAL, which would otherwise be trusted as MEDICATION # at step 4 — these are not medications being administered "alcohol", "tobacco", "nicotine", "cholesterol", "glucose", "ana", "creatinine", "hemoglobin", "sodium", "potassium", "calcium", "triglycerides", "bilirubin", # anatomy abbreviation bc5cdr sometimes mistags as CHEMICAL "mca", # more generic verbs/fragments/connectors found in second review pass # (2026-06-26) — meaningless as standalone entities "this", "that", "insidious", "evaluate", "team", "study", "decision", "distribution", "patient lives", "prescription", "consistent", "consistent with", "caring", "dear doctor", # lab test name (not a symptom or medication) "inr", # exam/test names that don't fit DISEASE/SYMPTOM/PROCEDURE cleanly "romberg", # medical supplies/materials, not clinical entities "stockinette", "sterile saline", "stitch", # social-history item, not a clinical finding "tattoos", "tattoo", # section-header fragment "past surgical history", # dosing-frequency abbreviations with periods (the word-boundary # regex in text_utils.py's abbreviation expander only matches # period-free forms like "qd"; "q.d" with periods slips through) "q.d", "b.i.d", "t.i.d", "q.i.d", "q.d.", "b.i.d.", "t.i.d.", "q.i.d.", "p.o", "p.o.", # role/department abbreviations, not clinical entities "pcp", "appointment", "emergency department", # found via dashboard review (2026-06-27): ECG lead labels, lab/ # biomarker names, dosing/administration terms, and medical # specialty names — none are symptoms or diseases on their own "v1-v4", "v1", "v2", "v3", "v4", "v5", "v6", "troponin", "loading dose", "cardiology", "neurology", "oncology", "urology", "radiology", "pathology", "dermatology", "psychiatry", "gastroenterology", "endocrinology", "pulmonology", "nephrology", "rheumatology", "neurosurgery", # found via dashboard review (2026-06-28): narrative/relational # phrases, clinical scales and vital-sign labels (not the finding # itself), treatment categories, and location/unit references — # none are symptoms or diseases on their own "family", "arrival", "family informed", "gcs", "blood pressure", "anticoagulation", "neurocritical care unit", "critical prognosis", }) # Keywords that signal a procedure mention. _PROCEDURE_KEYWORDS: frozenset[str] = frozenset({ # surgical actions "surgery", "surgical", "operation", "procedure", "resection", "excision", "biopsy", "osteotomy", "incision", "dissection", "closure", "suture", "ligat", "anastomosis", "hemostasis", "electrocautery", "cauterization", "angioplasty", "stenting", "intubation", "ventilation", "dialysis", "transfusion", "infusion", "transplant", "repair", "reconstruction", "amputation", "debridement", "drainage", "aspiration", "lavage", "catheterization", "catheterisation", "appendectomy", "cholecystectomy", "hysterectomy", "mastectomy", "colostomy", "tracheostomy", "orchiectomy", # anaesthesia / sedation "anesthesia", "anaesthesia", "anesthetic", "anaesthetic", "sedation", # devices placed during procedures "stent", "catheter", "drain", "cannula", # imaging / diagnostics — specific compound terms rather than a # bare "ct " substring, which would collide with "infarct " (ends # in "...ct ") and incorrectly flag it as a procedure "ct scan", "ct head", "ct chest", "ct abdomen", "ct pelvis", "ct brain", "ct angiogram", "mri", "x-ray", "ultrasound", "echocardiogram", "electrocardiogram", "ekg", "ecg", "eeg", # neurosurgical procedures "craniectomy", "craniotomy", # interventional cardiology / common abbreviations "pci", "cabg", "ptca", "ercp", "cath", "endoscopy", "colonoscopy", "bronchoscopy", "laparoscopy", }) # Disease/diagnosis abbreviations and phrases the model often misses. _DISEASE_TERMS: frozenset[str] = frozenset({ "stemi", "nstemi", "acs", "mi", "chf", "hf", "copd", "ckd", "esrd", "afib", "af", "vt", "vf", "dvt", "pe", "uti", "cad", "dm", "htn", "gerd", "ibd", "ibs", "ms", "als", "ra", "sle", "hiv", "st elevation", "st depression", "st changes", }) # Common drug-name suffixes — catch most small-molecule medications. _MEDICATION_SUFFIXES: tuple[str, ...] = ( "metformin", "insulin", "statin", "vastatin", # atorvastatin, simvastatin … "pril", # lisinopril, enalapril … "sartan", # losartan, valsartan … "olol", # metoprolol, atenolol … "dipine", # amlodipine, nifedipine … "azole", # fluconazole, omeprazole … "mycin", "cillin", "cycline",# antibiotics "afil", # sildenafil, tadalafil … "mab", "umab", "ximab", # monoclonal antibodies "tide", "tide", # peptide drugs "aspirin", "warfarin", "heparin", "clopidogrel", "prednisone", "prednisolone", "dexamethasone", "morphine", "codeine", "oxycodone", "fentanyl", "amoxicillin", "azithromycin", "ciprofloxacin", "furosemide", "spironolactone", "hydrochlorothiazide", "albuterol", "salbutamol", "budesonide", "levothyroxine", "thyroxine", ) # Medication-form / dosage-form words. A mention combining one of these # with anything else ("cough syrup", "eye drops") is describing a # medication being taken, regardless of what label the NER model assigned — # bc5cdr occasionally mistags an OTC remedy as DISEASE in context. _MEDICATION_FORM_WORDS: tuple[str, ...] = ( # NOTE: deliberately no "capsule" — it's genuinely ambiguous in # clinical text (medication form vs. anatomical structure: joint # capsule, lens capsule, renal capsule), and the false-positive risk # outweighs the narrow benefit here. "syrup", "lozenge", "ointment", "cream", "lotion", "suspension", "drops", "inhaler", ) # Anatomical term signals _ANATOMY_KEYWORDS: frozenset[str] = frozenset({ "heart", "lung", "liver", "kidney", "brain", "spine", "abdomen", "abdominal", "thorax", "pelvis", "femur", "tibia", "fibula", "conjunctiva", "conjunctival", "cornea", "corneal", "retina", "retinal", "artery", "vein", "aorta", "ventricle", "atrium", "cortex", "cerebellum", "cerebrum", "frontal", "parietal", "temporal", "occipital", "trachea", "bronchus", "alveoli", "esophagus", "oesophagus", "stomach", "duodenum", "colon", "rectum", "bladder", "ureter", "urethra", "prostate", "uterus", "ovary", "testis", "arm", "leg", "chest", "shoulder", "knee", "hip", "ankle", "wrist", "elbow", # additional limb / joint / extremity terms (same "X + pain" trap — # ICD-10 codes these under musculoskeletal M-codes, not R-codes) "back", "neck", "foot", "feet", "hand", "hands", "finger", "fingers", "toe", "toes", "groin", "flank", "joint", "joints", "limb", "limbs", "forearm", "calf", "thigh", "heel", "sole", "palm", "molar", "quadriceps", "metatarsal", "metatarsophalangeal", # additional surface / soft-tissue structures "skin", "scrotum", "penis", "vulva", "perineum", "muscle", "muscles", "tissue", "tissues", "fascia", "nerve", "nerves", "tendon", "ligament", "cartilage", "bone", "bones", "marrow", "vessel", "vessels", "lymph", "lymph node", "lymph nodes", "gland", "glands", "thyroid", "parathyroid", "adrenal", "pancreas", "spleen", "gallbladder", "bile duct", "appendix", "diaphragm", "peritoneum", "pleura", "pericardium", "meninges", "wound", "wounds", # surgical / hernia anatomy terms that bc5cdr mis-tags as DISEASE "cord", "ring", "canal", "oblique", "inguinal", "vas", "deferens", "spermatic", "epididymis", "sac", "pouch", "fossa", "sulcus", # head / neck / oral / ENT structures (same mis-tagging pattern) "palate", "throat", "tooth", "teeth", "gum", "gums", "tongue", "tonsil", "tonsils", "adenoid", "adenoids", "uvula", "pharynx", "larynx", "epiglottis", "sinus", "sinuses", "nasal", "nostril", "nostrils", "jaw", "mandible", "maxilla", "ear", "eardrum", }) # ── ICD-10 chapter classifier ───────────────────────────────────── class _ICD10Classifier: """Classify entities as DISEASE or SYMPTOM via ICD-10 chapter lookup. Lazily loads ``data/raw/icd10_codes.csv``, then uses rapidfuzz to fuzzy-match entity text against the 74k ICD-10 descriptions. The matched code's first character determines the label: R... → SYMPTOM (Chapter 18: Signs, symptoms, abnormal findings) V–Z → None (External causes / admin codes — defer to fallback) A–Q, S–T → DISEASE Every result is cached so each unique entity string is only looked up once per process lifetime. """ THRESHOLD = 85 # minimum WRatio score to accept a match def __init__(self) -> None: # dict[first_letter] -> ([descriptions], [codes]) self._index: dict[str, tuple[list[str], list[str]]] = {} self._cache: dict[str, str | None] = {} self._loaded = False def _load(self) -> None: if self._loaded: return self._loaded = True try: import pandas as pd from src.utils.config import Paths df = pd.read_csv(Paths.icd10_csv, dtype=str).dropna( subset=["description", "code"] ) descs = df["description"].str.lower().tolist() codes = df["code"].tolist() # Group by first letter of description for fast pre-filtering. # Reduces each fuzzy search from 74k to ~3-5k candidates. for desc, code in zip(descs, codes, strict=False): key = desc[0] if desc else "_" bucket = self._index.setdefault(key, ([], [])) bucket[0].append(desc) bucket[1].append(code) logger.debug( "ICD-10 classifier ready: %d codes across %d buckets", len(descs), len(self._index), ) except Exception as exc: logger.warning("ICD-10 classifier unavailable: %s", exc) def classify(self, entity_text: str) -> str | None: """Return DISEASE, SYMPTOM, or None (no confident match). None means the caller should fall back to keyword / model-label rules. """ if entity_text in self._cache: return self._cache[entity_text] self._load() label = None text_lower = entity_text.lower().strip() key = text_lower[0] if text_lower else "_" bucket = self._index.get(key) if bucket: from rapidfuzz import fuzz, process descs, codes = bucket # Get multiple candidates rather than just the top one — WRatio # frequently ties several descriptions at the same score for a # short query (e.g. "cough" scores 90 against "cough variant # asthma", "cough syncope", AND "cough, unspecified" alike). # extractOne returns whichever ties first in iteration order, # which is not necessarily the right one. Break ties by # preferring the shortest description — closest to an exact # match, least likely to be an unrelated compound condition # that merely contains the query as a substring. candidates = process.extract( text_lower, descs, scorer=fuzz.WRatio, score_cutoff=self.THRESHOLD, limit=10, ) if candidates: _desc, _score, idx = max( candidates, key=lambda c: (c[1], -len(c[0])) ) first = (codes[idx] or " ")[0].upper() if first == "R": label = "SYMPTOM" elif first not in "VWXYZ": label = "DISEASE" self._cache[entity_text] = label return label _icd10_clf = _ICD10Classifier() _SYMPTOM_FAST: frozenset[str] = frozenset({ "pain", "pains", "ache", "aches", "aching", "tenderness", "swelling", "bleeding", "blood loss", "discharge", "nausea", "vomiting", "diarrhea", "diarrhoea", "fatigue", "weakness", "numbness", "tingling", "dizziness", "shortness of breath", "dyspnea", "dyspnoea", "fever", "chills", "sweating", "palpitations", "syncope", "seizure", "confusion", "dysuria", "hematuria", "haematuria", "incontinence", "urinary frequency", "urinary retention", "nocturia", "retention", "chest pain", "chest tightness", "chest discomfort", "edema", "oedema", "ascites", "jaundice", "cyanosis", "tachycardia", "bradycardia", "hypotension", # weight / appetite signs (ICD-10 R63.x — stored as "Abnormal weight loss" # so first-letter bucket lookup misses the R-code; catch here explicitly) "weight loss", "weight gain", "anorexia", "appetite loss", "malaise", "lethargy", "myalgia", "arthralgia", "rash", "erythema", "pruritus", "urticaria", "haemoptysis", "hemoptysis", "epistaxis", "polyuria", "oliguria", "anuria", "constipation", "bruising", "bruise", "bruised", "ecchymosis", "diplopia", "blurred vision", "tinnitus", "vertigo", "distention", "distension", "bloating", "diaphoresis", }) # Pathology-indicating suffix words/morphemes. An anatomy keyword combined # with one of these is a compound DISEASE NAME ("liver disease", # "hyperthyroidism"), not a body-part reference — the anatomy word is the # disease's root, not what the entity is actually about. Without this # check, the anatomy-substring safety net (step 10) would intercept these # before bc5cdr's own DISEASE tag ever gets a chance at ICD-10 arbitration, # since "thyroid" is a literal substring of "hyperthyroidism" and "liver" # of "liver disease". _DISEASE_NAME_HINTS: tuple[str, ...] = ( "disease", "failure", "itis", "osis", "oma", "pathy", "megaly", "emia", "uria", "algia", ) def _looks_like_disease_name(text_lower: str) -> bool: """Heuristic: does this anatomy-containing text look like a disease name? Checks for an explicit pathology suffix word ("liver disease") or the hyper-/hypo- endocrine prefix pattern ("hyperthyroidism"). """ if any(hint in text_lower for hint in _DISEASE_NAME_HINTS): return True return bool(text_lower.startswith(("hyper", "hypo"))) _SYMPTOM_QUALIFIERS: tuple[str, ...] = ( "pain", "ache", "aching", "tenderness", "swelling", "discomfort", "numbness", "weakness", "fatigue", "nausea", # found via gold-standard sample review (2026-06-26): anatomy term + # one of these is a symptom/finding report, not an anatomy mention # ("bladder incontinence", "abdominal injury", "sore throat", # "vascular abnormalities") "incontinence", "injury", "injuries", "sore", "abnormal", "abnormality", ) # Matches a run of 2+ consecutive punctuation marks — a reliable signal # that a NER span has crossed a sentence/section boundary and appended a # stray fragment (e.g. "gastroesophageal reflux disease.,past" — the # ".,past" leaked from the next sentence, not part of the entity). A # single period/comma is left alone since that can appear legitimately # within a real phrase or abbreviation. _SPAN_ARTIFACT_PATTERN = re.compile(r"[.,;:]{2,}") def _clean_entity_text(text: str) -> str: """Strip a trailing punctuation+fragment span-boundary artifact. Args: text: Raw entity text as extracted by spaCy (``ent.text``). Returns: Cleaned text with any trailing merge artifact removed. """ match = _SPAN_ARTIFACT_PATTERN.search(text) if match: text = text[:match.start()] return text.strip(" .,;:") def _normalise_label(raw_label: str, entity_text: str) -> str | None: """Map a raw NER label to one of our five clinical categories. Architecture ──────────── Two model types produce spans: • bc5cdr → raw labels DISEASE or CHEMICAL (high precision, domain-specific) • en_core_sci_lg → raw label ENTITY (broad coverage, no sub-type) DISEASE classification is restricted to bc5cdr DISEASE spans. en_core_sci_lg ENTITY spans can only become PROCEDURE / ANATOMY / MEDICATION / SYMPTOM — never DISEASE — because WRatio fuzzy-matching ICD-10 against any arbitrary biomedical token produces too many false positives ("antibiotics" → "Adverse effect of antibiotics" T36.x → DISEASE). Priority order (same for both model types up to step 8): 1. Negation / span-artifact filter — "denies X", "complaint of X", spans containing a stray location word like "room" 2. Skip filter — headers, demographics, severity qualifiers 3. Fragment filter — "mellitus", "syndrome" … 4. CHEMICAL / DRUG → MEDICATION 5. Disease abbreviations — STEMI, DVT, COPD … (too short for ICD-10) 6. Procedure keywords — incision, stent, anesthesia, MRI … 7. Medication suffixes / known drug class names 8. Anatomy exact match 9. Anatomy substring match ("cord structures", "external oblique" …) 10. Symptom fast-path — known R-code terms 11. bc5cdr DISEASE only → ICD-10 chapter lookup (DISEASE vs SYMPTOM) If ICD-10 has no opinion, keep as DISEASE (trust bc5cdr). 12. Default: SYMPTOM Returns: One of ``"DISEASE"``, ``"MEDICATION"``, ``"PROCEDURE"``, ``"ANATOMY"``, ``"SYMPTOM"``, or ``None`` (discard entity). """ text_lower = entity_text.lower().strip() label_up = raw_label.upper() # 1. Negation / span-artifact filter. A span beginning with a negation # word ("denies obesity") describes the ABSENCE of a finding, not a # reportable entity — classifying it as DISEASE would claim the # patient HAS the condition they explicitly don't. A span beginning # with a generic reporting phrase ("complaint of allergies") is a # leaked section-header fragment from imperfect span boundaries. # "room" catches NER span-boundary errors that merge an unrelated # location word with an adjacent clinical term (e.g. "emergency # room" + "cancer" -> "room cancer") — not a real entity either way. if any(text_lower.startswith(p) for p in _NEGATION_PREFIXES): return None if any(text_lower.startswith(p) for p in _GENERIC_REPORT_PREFIXES): return None if "room" in text_lower: return None # 2. Skip filter if text_lower in _SKIP_TERMS: return None # 3. Fragment filter if text_lower in _FRAGMENT_TERMS: return None # 4. bc5cdr CHEMICAL / DRUG — high-precision medication label if label_up in ("CHEMICAL", "DRUG"): return "MEDICATION" # 5. Disease abbreviations (single tokens too short for ICD-10 fuzzy match) if text_lower in _DISEASE_TERMS: return "DISEASE" # 6. Symptom fast-path (known R-code surface forms) — EXACT match, # checked before any substring-based rule. An exact match is an # unambiguous signal and must win over accidental substring # collisions in the keyword lists below (e.g. "distention" contains # "stent" — a PROCEDURE device keyword — so without this ordering # it would wrongly resolve to PROCEDURE before ever being checked # against the symptom list). if text_lower in _SYMPTOM_FAST: return "SYMPTOM" # 7. Procedure keywords if any(kw in text_lower for kw in _PROCEDURE_KEYWORDS): return "PROCEDURE" # 8. Medication suffix / known drug class, or a medication-form word # ("cough syrup", "eye drops") — these override even a bc5cdr # DISEASE tag, since bc5cdr occasionally mistags an OTC remedy # mention as a disease in context ("offered him a cough syrup"). if any(text_lower.endswith(sfx) or sfx in text_lower for sfx in _MEDICATION_SUFFIXES): return "MEDICATION" if any(form in text_lower for form in _MEDICATION_FORM_WORDS): return "MEDICATION" # 9. Anatomy — exact single-word match if text_lower in _ANATOMY_KEYWORDS: return "ANATOMY" # 10. Anatomy — substring match for multi-word phrases ("cord structures"). # If a symptom qualifier is also present ("knee pain", "shoulder # ache"), this is a symptom report, not an anatomy mention — resolve # to SYMPTOM directly rather than falling through to ICD-10, which # codes joint/limb pain under musculoskeletal M-codes (M25.5x), not # R-codes, so the chapter heuristic would otherwise misclassify it # as DISEASE. if any(kw in text_lower for kw in _ANATOMY_KEYWORDS): if any(q in text_lower for q in _SYMPTOM_QUALIFIERS): return "SYMPTOM" if not _looks_like_disease_name(text_lower): return "ANATOMY" # else: this is a compound disease name built on an anatomy root # ("liver disease", "hyperthyroidism") — fall through to ICD-10 # arbitration instead of forcing ANATOMY. # 11. ICD-10 lookup — ONLY for bc5cdr DISEASE spans. # Determines whether a confirmed biomedical disease entity is # a true DISEASE or a SYMPTOM/sign (ICD-10 Chapter 18, R-codes). # en_core_sci_lg ENTITY spans skip this step entirely. if label_up == "DISEASE": icd10_label = _icd10_clf.classify(entity_text) if icd10_label is not None: return icd10_label return "DISEASE" # bc5cdr confident; ICD-10 had no strong match # 12. Trust an unambiguous structural raw label directly, if the # model provided one (e.g. an alternate/future NER model emitting # ANATOMY / ORGAN / PROCEDURE / CELL directly, rather than the # generic ENTITY label en_core_sci_lg uses). Deliberately excludes # GENE / SPECIES / DISEASE from this fallback — those carry the # same false-positive risk that motivated restricting DISEASE to # the ICD-10-gated path above, so they must not bypass it here. mapped = _SAFE_STRUCTURAL_LABEL_MAP.get(label_up) if mapped: return mapped # 13. Default for ENTITY spans: SYMPTOM return "SYMPTOM" # ── Base interface ──────────────────────────────────────────────── class BaseNERPipeline(ABC): """Abstract interface for NER pipelines. Concrete implementations (SpacyNERPipeline, and any future HuggingFace NER pipeline) must implement ``extract()``. All code that uses NER should type-hint against this base class, not the concrete implementation. """ @abstractmethod def extract(self, text: str) -> list[Entity]: """Extract named entities from a single text string. Args: text: Cleaned clinical note text. Returns: List of :class:`Entity` objects sorted by start offset. """ ... def extract_batch(self, texts: list[str]) -> list[list[Entity]]: """Extract entities from a list of texts. Default implementation calls extract() in a loop. Concrete subclasses may override with a more efficient batched implementation. Args: texts: List of clinical note strings. Returns: List of entity lists, one per input text. """ return [self.extract(t) for t in texts] # ── scispaCy implementation ─────────────────────────────────────── class SpacyNERPipeline(BaseNERPipeline): """NER pipeline backed by a scispaCy biomedical model. Loads the model once at construction time and reuses it for all subsequent calls. The model is expensive to load (~2s) but fast to run (~10ms per note). Args: model_name: Name of the spaCy/scispaCy model to load. Defaults to the value in ModelConfig. Example:: pipeline = SpacyNERPipeline() entities = pipeline.extract("Patient has hypertension and diabetes.") for ent in entities: print(ent.text, ent.label) # → hypertension DISEASE # → diabetes DISEASE """ def __init__(self, model_name: str = ModelConfig.ner_model) -> None: self._model_name = model_name self._nlp = None # lazy-loaded on first use def _load_model(self) -> None: """Load the spaCy model into memory. Called lazily on the first extract() call so that importing this module does not trigger a slow model load at startup. Raises: OSError: If the model is not installed. The error message includes the installation command. """ try: import spacy logger.info("Loading NER model: %s", self._model_name) # extract()/extract_batch() only ever read doc.ents, so the # tagger/parser/lemmatizer/attribute_ruler components are pure # overhead here -- excluding them is verified to produce # byte-identical entity output (tok2vec + ner are the only # components ner actually depends on) while trimming memory, # which matters on constrained deployment hosts. self._nlp = spacy.load( self._model_name, exclude=["tagger", "attribute_ruler", "lemmatizer", "parser"], ) logger.info("NER model loaded ✓") except OSError as exc: raise OSError( f"spaCy model '{self._model_name}' is not installed.\n" "Install it with:\n" f" pip install https://s3-us-west-2.amazonaws.com/" f"ai2-s2-scispacy/releases/v0.5.3/" f"{self._model_name}-0.5.3.tar.gz" ) from exc @property def model_name(self) -> str: """Return the spaCy model name.""" return self._model_name @property def nlp(self): """Return the loaded spaCy model, loading it if necessary.""" if self._nlp is None: self._load_model() return self._nlp def extract(self, text: str) -> list[Entity]: """Extract named entities from a clinical note. Args: text: Cleaned clinical note text. Run :func:`src.utils.text_utils.clean_clinical_text` on raw input before passing it here. Returns: List of :class:`Entity` objects sorted by start offset. Empty list if text is empty or no entities are found. """ if not text or not text.strip(): return [] doc = self.nlp(text) entities = [] for ent in doc.ents: cleaned_text = _clean_entity_text(ent.text) # Skip very short tokens — usually OCR artefacts or initials if len(cleaned_text) < 3: continue label = _normalise_label(ent.label_, cleaned_text) if label is None: continue entities.append(Entity( text = cleaned_text, label = label, start = ent.start_char, end = ent.start_char + len(cleaned_text), # scispaCy does not expose per-entity scores natively; # we leave confidence as None rather than fabricating a value confidence = None, )) return sorted(entities, key=lambda e: e.start) def extract_batch(self, texts: list[str], batch_size: int = 32) -> list[list[Entity]]: """Extract entities from multiple texts using spaCy's pipe(). More efficient than calling extract() in a loop because spaCy processes documents in parallel where possible. Args: texts: List of clinical note strings. batch_size: Number of texts to process per batch. Returns: List of entity lists, one per input text. """ if not texts: return [] results = [] # spaCy's nlp.pipe is the idiomatic way to batch-process for doc in self.nlp.pipe(texts, batch_size=batch_size): entities = [] for ent in doc.ents: cleaned_text = _clean_entity_text(ent.text) if len(cleaned_text) < 3: continue label = _normalise_label(ent.label_, cleaned_text) if label is None: continue entities.append(Entity( text = cleaned_text, label = label, start = ent.start_char, end = ent.start_char + len(cleaned_text), confidence = None, )) results.append(sorted(entities, key=lambda e: e.start)) return results # ── Hybrid pipeline ─────────────────────────────────────────────── class HybridNERPipeline(BaseNERPipeline): """Combines two scispaCy models for best coverage and accuracy. ``fine_model`` (default: en_ner_bc5cdr_md) runs first and produces accurate DISEASE and MEDICATION labels from a model trained specifically on biomedical text. ``broad_model`` (default: en_core_sci_lg) runs second and contributes any spans that do not overlap with the fine model's output — typically PROCEDURE, ANATOMY, and SYMPTOM entities that bc5cdr was not trained to detect. Overlap rule: if a broad-model span shares any characters with a fine-model span, the broad span is dropped (fine wins). Args: fine_model: Model name for precise DISEASE/MEDICATION detection. broad_model: Model name for broad entity coverage. """ def __init__( self, fine_model: str = "en_ner_bc5cdr_md", broad_model: str = "en_core_sci_lg", ) -> None: self._fine = SpacyNERPipeline(fine_model) self._broad = SpacyNERPipeline(broad_model) @property def model_name(self) -> str: """Return a descriptive name for the hybrid pipeline.""" return f"hybrid({self._fine.model_name} + {self._broad.model_name})" @staticmethod def _overlaps(a: Entity, b: Entity) -> bool: return max(a.start, b.start) < min(a.end, b.end) def extract(self, text: str) -> list[Entity]: fine_ents = self._fine.extract(text) broad_ents = self._broad.extract(text) merged = list(fine_ents) merged.extend(broad for broad in broad_ents if not any(self._overlaps(broad, f) for f in fine_ents)) return sorted(merged, key=lambda e: e.start) def extract_batch(self, texts: list[str], batch_size: int = 32) -> list[list[Entity]]: fine_batches = self._fine.extract_batch(texts, batch_size) broad_batches = self._broad.extract_batch(texts, batch_size) results = [] for fine_ents, broad_ents in zip(fine_batches, broad_batches, strict=False): merged = list(fine_ents) merged.extend(broad for broad in broad_ents if not any(self._overlaps(broad, f) for f in fine_ents)) results.append(sorted(merged, key=lambda e: e.start)) return results # ── Factory function ────────────────────────────────────────────── def build_ner_pipeline( model_name: str | None = None, ) -> BaseNERPipeline: """Construct and return the configured NER pipeline. Pass ``model_name="hybrid"`` (or set ``NER_MODEL=hybrid`` in ``.env``) to use :class:`HybridNERPipeline`, which combines ``en_ner_bc5cdr_md`` (accurate DISEASE/MEDICATION) with ``en_core_sci_lg`` (broad coverage for PROCEDURE/ANATOMY/SYMPTOM). Any other value is treated as a single spaCy model name and passed to :class:`SpacyNERPipeline`. Args: model_name: Override the default model from config. Use ``"hybrid"`` for the two-model pipeline. Returns: A ready-to-use :class:`BaseNERPipeline` instance. """ name = model_name or ModelConfig.ner_model if name == "hybrid": logger.debug("Building hybrid NER pipeline (bc5cdr + en_core_sci_lg)") return HybridNERPipeline() logger.debug("Building NER pipeline with model: %s", name) return SpacyNERPipeline(model_name=name)