SafeCombo AI: Deep Learning for Pharmacovigilance & Drug Safety 💊🤖

SafeCombo AI is an end-to-end machine learning pipeline designed to predict severe adverse drug reactions (ADRs) caused by polypharmacy (drug-drug interactions). By mining the FDA's Adverse Event Reporting System (FAERS), this project cleans, standardizes, and models real-world clinical data to build a hierarchical safety dashboard for healthcare professionals.

🌟 Value Proposition & Key Features

Raw FDA data is notoriously noisy, filled with typos, varied dosages, and overlapping symptoms. This project transforms that chaos into a clinical decision-support tool through:

• Deep NLP & Semantic Entity Resolution: Uses rapidfuzz and the official FDA National Drug Code (NDC) database to standardize over 26,000 messy drug text entries into 15,000+ clean, generic pharmaceutical entities.
• Clinical Phenotyping: Maps thousands of vague, self-reported symptoms (e.g., "headache", "tiredness") into 21 hyper-specific clinical risks (e.g., GI Bleeding, CNS Depression, Acute Renal Toxicity) based on MedDRA standards.
• Surgical Data Purification: Implements strict noise-reduction filters to eliminate fatal polypharmacy outliers (>6 drugs) and systemic failure "shotgun" reactions, improving the signal-to-noise ratio.
• Robust Deep Learning Architecture: A PyTorch neural network built to resist overfitting on sparse text data, utilizing heavy L2 regularization, dropout (50%), and weighted loss (BCEWithLogitsLoss) to accurately learn rare but fatal clinical events without hallucinating.
• Dynamic Thresholding: Automatically calculates the optimal classification threshold (F1-Score maximization) for each specific disease class, replacing naive 0.5 thresholds with statistically sound cutoffs.

🛠️ Technical Pipeline

1. Data Ingestion: Automated bulk downloading of JSON records from the openFDA API.
2. Data Engineering: PySpark/Pandas regex cleaning to strip dosages, formulations, and administrative noise.
3. Fuzzy Matching: Resolution of brand names and typos against the FDA NDC standard.
4. Target Binarization: Multi-label encoding of the 21 clinical phenotypes.
5. Vectorization: TF-IDF transformation of the purified drug regimens.
6. Modeling & Inference: PyTorch DNN training, followed by a hierarchical reporting engine.

📊 Example Inference Output

The model generates a clean, hierarchical clinical report. Example for Warfarin + Ibuprofen + Aspirin:

🩺 Purified Safety Profile: WARFARIN + IBUPROFEN + ASPIRIN
=================================================================
🧠 Nervous System & Psych
----------------------------------------
  ↳ Risk: Vestibular/CNS           🟠 MOD | 56.1%  ███████████░░░░░░░░░

🩸 Bleeding & GI Disorders
----------------------------------------
  ↳ Risk: GI Bleeding              🔴 HIGH | 77.7%  ███████████████░░░░░
  ↳ Risk: Systemic Bleeding        🟡 LOW | 35.2%  ███████░░░░░░░░░░░░░
=================================================================