AuraMed

Inspiration

In modern healthcare, over 40% of elderly and chronic care patients are prescribed five or more medications simultaneously—a state known as polypharmacy. Standard clinical decision support systems (CDSS) rely on static lookup tables that only check pairwise drug interactions (Drug A + Drug B). They fail completely when predicting multi-drug cascades (Drug A + B + C + D) and ignore vital patient biological factors, such as liver enzyme clearance rates or renal function markers. This blind spot leads to severe, preventable adverse drug reactions (ADRs) that account for millions of emergency hospital admissions worldwide.

We were inspired to build AuraMed AI to move away from rigid, rule-based systems and transform pharmacological analysis into a dynamic, intelligent graph structure.

What It Does

AuraMed AI is an explainable machine learning engine that predicts complex, multi-drug adverse interactions and biological pathway collisions before a prescription is finalized.

Knowledge Graph Representation: Models drugs, biological target proteins, metabolic enzymes (e.g., CYP450 family), and physiological pathways as a connected web rather than flat lists. Polypharmacy Cascade Prediction: Evaluates multi-drug combinations simultaneously to detect non-linear interactions and metabolic overload. Explainable AI (XAI): Uses Integrated Gradients (via Captum) to provide clear, human-understandable biological reasons for every flagged risk, showing clinicians exactly why a combination is dangerous rather than acting as a black box.

How We Built It

AuraMed AI is built entirely in Python as a backend API engine:

Data Ingestion & Graph Construction: Extracted biomedical data from open repositories (ChEMBL, BioSNAP, DrugBank Open Data) using Pandas and NetworkX to construct a dynamic drug-target-pathway graph.
Deep Learning Model: Implemented a Graph Convolutional Network (GCN) using PyTorch Geometric (PyG) to learn node embeddings and compute similarity scores across unseen multi-drug pathways.
Interpretability Engine: Integrated Captum to compute feature attributions, mapping predictions back to specific enzymes and target proteins.
Microservice API:Wrapped the deep learning pipeline in FastAPI with Pydantic data validation, serving real-time JSON risk predictions to any front-end or electronic health record (EHR) system.

Challenges We Faced

Data Heterogeneity & Sparsity: Integrating diverse biological datasets with varying chemical formats required extensive data normalization and graph schema mapping. Graph Computational Overhead: Scaling Graph Neural Networks to process multi-node subgraphs in real-time under low-latency constraints required optimizing edge sampling and model batching. Model Interpretability: Translating complex mathematical tensor weights into clinically meaningful biological explanations that medical professionals can trust.

What We Learned

How Graph Neural Networks excel at learning high-dimensional structural representations compared to standard tabular machine learning models.
The critical necessity of Explainable AI (XAI) in healthcare software—clinicians require actionable biological proof before modifying patient prescriptions.
How to structure microservices using FastAPI and PyTorch Geometric for scalable, production-grade inference pipelines.