ADEGuard — AI That Uncovers Hidden Vaccine Side Effects

🏆 Category

Healthcare | AI for Social Good | NLP

---

❓ Problem

Millions of Adverse Drug Event (ADE) reports are collected every year — especially post-COVID-19 — but:

• 70%+ of critical details are buried in free-text narratives
• Traditional safety systems struggle to interpret noisy text at scale
• Hidden side effects delay response, impacting patient safety

Healthcare needs a faster way to convert unstructured patient reports into actionable clinical signals.

---

💡 Solution: ADEGuard

ADEGuard is an AI-powered system that:

✔ Extracts ADEs & Drugs from raw text ✔ Clusters similar adverse events using semantic similarity ✔ Classifies severity & risk factors in real-time ✔ Surfaces safety insights to clinicians with explainability

🚑 Short story: AI that reads what humans report — but others overlook.

---

🧠 What It Does

Feature	Description
ADE & Drug NER	BioBERT identifies critical medical entities
Age- & Severity-Aware Risk	Metadata + weak labels refine risk classification
Meaningful Clustering	HDBSCAN + SentenceBERT create event groupings
Explainability	SHAP highlights why predictions were made
Safety Dashboard	Real-time Streamlit app visualizes trends & alerts

---

🛠️ How We Built It

Component	Tech
Language Model	BioBERT (fine-tuned)
Embeddings	SentenceBERT
Clustering	HDBSCAN
Classification	Hybrid rule-based + weak supervision
Visualization	Streamlit + Python
Dataset	VAERS + manually annotated samples

Training Objective: We combine gold-standard annotations with weak labels to improve scalability:

[ \text{Loss} = \alpha L_\text{gold} + (1 - \alpha)L_\text{weak} ]

This improved generalization while reducing annotation load.

---

🚧 Challenges

• Extreme noise in patient-written reports
• Entity boundary errors in clinical text
• Avoiding false alarms while detecting clinically relevant signals
• Balancing speed vs. accuracy in clustering large datasets

---

🎯 Accomplishments We’re Proud Of

🏅 Detected ADE patterns not visible in structured fields 🏅 Delivered model decisions with explainable reasoning 🏅 End-to-end working prototype viable for safety monitoring teams

---

📈 What’s Next

• Clinical validation with pharmacovigilance partners
• Expand coverage to multiple vaccines & medications
• Model serving for real-time hospital integration
• Add LLM-based causal inference to strengthen severity attribution

---

🙋‍♂️ Team

Developer / Researcher: Karthikeyan Ravichandran — MS Data Science, Healthcare AI

---

Built With

• fastapi
• generativeai
• python
• streamlit