AyuSynapse

Ayusynapse — Multimodal AI for reliable, auditable trial eligibility and accelerated enrollment.

Comment

about AyuSynapse

Inspiration

Neurology and oncology clinical trials often struggle with slow recruitment and misaligned eligibility, delaying life-saving research and treatments. We were inspired by the gap between patients with urgent needs and the trials designed to help them. The idea for Ayusynapse was born from this challenge: using explainable, multimodal AI to make clinical trial matching faster, more reliable, and ethically auditable.

What it does

Ayusynapse ingests heterogeneous clinical data — structured (lab reports, biomarkers) and unstructured (clinical notes, imaging metadata) — and applies multimodal AI models to:

  • Parse eligibility criteria from trial protocols.
  • Match patients with trials in neurology and oncology.
  • Provide an auditable trail of reasoning for each match.

In short: We accelerate trial enrollment while ensuring transparency and trust.

How we built it

  • Data pipeline: Pre-processed mock EMRs, trial criteria, and biomedical literature.
  • NLP + LLMs: Used transformer-based models for semantic parsing of trial documents.
  • Multimodal integration: Combined structured (tabular biomarkers, lab values) with unstructured (textual notes) inputs using fusion architectures.
  • Explainability: Added interpretable layers (e.g., SHAP values, attention visualization) so that physicians can see why a patient qualifies.
  • Prototype UI: Developed a clean interface for clinicians to query and view eligibility results.

Mathematically, we modeled patient–trial matching as a similarity optimization:

[ \text{Match Score}(p, t) = \alpha \cdot \text{sim}{\text{text}}(E_p, C_t) + \beta \cdot \text{sim}{\text{biomarker}}(B_p, R_t) ]

where:

  • (E_p) = patient’s clinical notes
  • (C_t) = trial criteria text
  • (B_p) = patient biomarkers
  • (R_t) = required trial biomarkers
  • (\alpha, \beta) are tunable weights

Challenges we ran into

  • Heterogeneous data formats — EMRs vary widely in structure, requiring heavy preprocessing.
  • Ambiguous trial criteria — Some eligibility conditions are written vaguely, demanding semantic disambiguation.
  • Balancing explainability with performance — Making the AI interpretable without sacrificing accuracy was tough.
  • Time constraints — Building a multimodal system in a hackathon setting meant rapid iteration.

Accomplishments that we're proud of

  • Successfully integrated multimodal inputs into one unified matching pipeline.
  • Built an explainable AI prototype that clinicians could trust.
  • Designed a professional-grade branding and interface (logo, UI, workflow) to make the project pitch-ready.
  • Demonstrated that patient-trial matching can be both fast and transparent.

What we learned

  • How to handle biomedical text and eligibility parsing at scale.
  • The importance of interpretability in AI for healthcare.
  • Collaboration across clinical, technical, and design perspectives strengthens outcomes.
  • That AI in clinical research must prioritize ethics, fairness, and auditability.

What's next for Ayusynapse

  • Expand beyond neurology & oncology to other therapeutic areas.
  • Validate the system with real-world trial datasets.
  • Partner with hospitals and CROs for pilot deployment.
  • Add imaging data integration (radiomics, MRI features) to enhance eligibility matching.
  • Work toward regulatory-grade compliance (HIPAA/GDPR alignment, FDA explainability standards).

Built With

Share this project:

Updates