MedContext Precision Oncology

Inspiration

Precision-oncology reports contain more than molecular names. A variant can be present, absent, uncertain, historical, current, or even refer to a family member.

I built MedContext Precision Oncology to convert that unstructured clinical language into structured, contextualized facts that can be consumed by downstream AI systems.

What it does

MedContext detects biomedical entities, parses their molecular structure, extracts the relevant context, and combines symbolic clinical cues with frozen BioLORD-2023 embeddings.

Four lightweight XGBoost classifiers then predict:

assertion;

experiencer;

temporality;

certainty.

The result is compiled into either a detailed Full Audit Packet or a smaller Compact AI Packet for RAG systems, agents, APIs, or analytics.

How we built it

Precision-oncology reports contain more than molecular names. A variant can be present, absent, uncertain, historical, current, or even refer to a family member.

I built MedContext Precision Oncology to convert that unstructured clinical language into structured, contextualized facts that can be consumed by downstream AI systems.

Challenges we ran into

The main challenge was preserving clinical context without making the system unnecessarily complex. In particular, conflicting findings required separate entity mentions and labels, while context windows had to remain small enough to avoid unrelated information.

The project reinforced that clinical NLP is not only about detecting entities, but about understanding what is being said about them.

Accomplishments that we're proud of

The final pipeline works end-to-end, from raw report text to structured clinical facts and downloadable JSON outputs.

On the held-out synthetic benchmark, MedContext achieved:

Fact extraction F1: 1.000

Exact four-target context match: 0.963

Mean inference latency: ~94 ms/report

I am especially proud that the system correctly separates conflicting mentions of the same molecular entity into different contextual facts.

What we learned

The final pipeline works end-to-end, from raw report text to structured clinical facts and downloadable JSON outputs.

On the held-out synthetic benchmark, MedContext achieved:

Fact extraction F1: 1.000

Exact four-target context match: 0.963

Mean inference latency: ~94 ms/report

I am especially proud that the system correctly separates conflicting mentions of the same molecular entity into different contextual facts.

What I learned

This project reinforced that clinical NLP is not only about recognizing biomedical entities, but about understanding what is being said about them.

I also learned that deterministic clinical rules and frozen biomedical embeddings can complement each other effectively without requiring expensive Transformer fine-tuning.

What's next for MedContext Precision Oncology

Next steps include:

evaluation on de-identified real-world reports;

broader oncology and EHR terminology;

FHIR-compatible outputs;

direct integration with RAG systems and clinical AI agents;

external validation and confidence calibration.

MedContext is a research and hackathon prototype and is not intended for clinical decision-making.

Built With

• biolord
• streamlit
• transformers
• xgboost