Editing AI for Drug Adverse Effect Detection (section)

== <span style="color: #FFFFFF;">Applying</span> ==
'''NLP extraction of adverse drug events from clinical notes:'''
<syntaxhighlight lang="python">
from transformers import pipeline, AutoTokenizer, AutoModelForTokenClassification
import re

# BioBERT fine-tuned for ADE detection (NER on clinical text)
model_name = "allenai/biomed_roberta_base"  # or "d4data/biomedical-ner-all"
ner_pipeline = pipeline("ner", model="allenai/biomed_roberta_base",
                        aggregation_strategy="simple")

# ADE-specific NER model (trained on ADE corpus)
ade_model = AutoModelForTokenClassification.from_pretrained(
    "ncats/DrugProt-bioBERT-base-cased"  # Drug + protein NER
)

def extract_drug_ade_pairs(clinical_note: str) -> list[dict]:
    """Extract (drug, adverse_effect) pairs from clinical text."""
    entities = ner_pipeline(clinical_note)
    drugs, adverse_effects = [], []
    for ent in entities:
        if ent['entity_group'] in ['DRUG', 'CHEMICAL']:
            drugs.append({'text': ent['word'], 'start': ent['start'], 'end': ent['end']})
        elif ent['entity_group'] in ['DISEASE', 'SYMPTOM', 'ADE']:
            adverse_effects.append({'text': ent['word'], 'start': ent['start']})
    # Simple proximity-based relation extraction: drug + ADE within same sentence
    pairs = []
    sentences = clinical_note.split('.')
    for sent in sentences:
        sent_drugs = [d for d in drugs if d['text'].lower() in sent.lower()]
        sent_aes = [a for a in adverse_effects if a['text'].lower() in sent.lower()]
        for drug in sent_drugs:
            for ae in sent_aes:
                pairs.append({
                    'drug': drug['text'],
                    'adverse_effect': ae['text'],
                    'sentence': sent.strip()
                })
    return pairs

# Disproportionality analysis on FAERS data
import pandas as pd
import numpy as np

def reporting_odds_ratio(faers_df: pd.DataFrame, drug: str, event: str) -> dict:
    """Compute ROR for drug-event pair in spontaneous reports."""
    a = len(faers_df[(faers_df['drug'] == drug) & (faers_df['event'] == event)])
    b = len(faers_df[(faers_df['drug'] == drug) & (faers_df['event'] != event)])
    c = len(faers_df[(faers_df['drug'] != drug) & (faers_df['event'] == event)])
    d = len(faers_df[(faers_df['drug'] != drug) & (faers_df['event'] != event)])
    if b == 0 or c == 0: return {'ror': None, 'signal': False}
    ror = (a * d) / (b * c)
    # 95% CI on log scale
    se = np.sqrt(1/a + 1/b + 1/c + 1/d)
    ci_low = np.exp(np.log(ror) - 1.96 * se)
    ci_high = np.exp(np.log(ror) + 1.96 * se)
    signal = ror > 2.0 and ci_low > 1.0 and a >= 3  # Standard signal threshold
    return {'ror': ror, 'ci': (ci_low, ci_high), 'n': a, 'signal': signal}
</syntaxhighlight>

; Drug safety AI tools
: '''FAERS analysis''' → OpenVigil (web), FDA Sentinel System, WHO VigiBase / VigiAccess
: '''EHR pharmacovigilance''' → CPRD Aurum + SCCS, Optum ClinicalAI, FDA MedWatch
: '''NLP ADE extraction''' → MedEx, cTAKES, clinicalBERT NER models
: '''DDI prediction''' → DDIMDL, SSI-DDI (graph learning), DrugBank + ML
: '''Social media''' → Social Media Mining for Health (SMM4H), MedWatcher
</div>

<div style="background-color: #8B4500; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">