Zero Knowledge Medicine

Privacy Preserving Fixed Coordinate Temporal Architecture

Piyush Chaudhari

Rich Data Poor Data

• Healthcare data is touted as the new currency

• It is easy to deposit in a bank, but there is little to no yield on it

• Clinical Informatics Decision Making is currently built on averages and trained on dataset with limited scope

• Reasons

  • HIPAA and Privacy Data Sharing Restrictions
  • Even when shared, data is paranormally distributed
  • Common Data Models have common representations but variable transformation workflows
  • Intensive, but not interoperable resource investment

The Fix - Fixed Coordinate Temporal Data

• We will look at genomic data structure for inspiration, not instruction
• Each patient will be represented as a point in a shared coordinate space similar to a genome \(Health Coordinate = z(t) \in R^d\)

where \(z\) are the physiological coordinates, \(t\) is time since birth and \(R^d\) represents d-Dimension real space

• The health history of each patients becomes a sequence from birth -> present -> future

• Deterministic Coordinates similar to genomic data (Mutation at chromosome a, postion b means the same for everyone which helps direct comparision across individuals and ML at scale)

The Fixed Coordinate Temporal Data Model

A comprehensive represention of Fixed Coordinate Temporal Data Model - Click the image for details

Fixed Coordinate Temporal Data Model - Base Layer

Fixed Coordinate Temporal Data Model - Example Phenotype Layer

Fixed Coordinate Temporal Data Model - Example Subphenotype Layer

Fixed Coordinate Temporal Data Model - Splicing Example

Fixed Coordinate Temporal Data Model - Applications

What about Privacy?

• Data is touted as the new currency, but the most important universal currency is trust

• The proposed model while making representation and analysis easier, makes deidentification exponentially tougher

• As the length of the sequence increases, the odds of another sequence with ~100% similarity decreases

• HIPAA and Privacy laws doesn’t allow hospitals to share patient data in raw or even transformed format

Zero Knowledge Proofs

General ZKP workflow - Click on the image for details

Zero Knowledge Proofs (ZKPs) can prove - for a given input and model, an output is produced, without revealing any of the above

Step 1 - Hospital A publishes one commitment K — a cryptographic fingerprint of its model

Step 2 - Hospital B can run that commitment against any of its own patients, any number of times, without Hospital A ever being involved again

Example: For Patient 1, the arithmetic circuit finds the feature vector satisfies the AKI decision boundary → proof is TRUE, and the onset window is disclosed. Image generated using Claude

Example: For Patient 2, the same circuit finds the vector falls outside the boundary → proof is FALSE, and nothing further is revealed — not even how far below the threshold the patient is. Image generated using Claude

Moving Forward

• Phase 1 - Transformation architecture PoC for fixed coordinate temporal data model to and from common CDMs like i2b2, OMOP, Epic Caboodle
  • The source code for open-source data models will be made available as an open-source resource
• Phase 2 - Test Phase A on sample AKI patients at UF Health and validate it over larger datasets, like Epic COSMOS
• Generate prototype ZK STARK infrastructure using temporal data model for AKI risk prediction, similar patient matching and digital twin trajectory on a small sample of patients

Current Limiting Factors

• Fixed Coordinate Temporal Data Model
  • Requires clean source datasets
  • Variability in methods is lower but not \(0\)
• Zero Knowledge Proofs
  • Expensive and slow for complex models (e.g. Neural Networks), may need quantization
  • Latency, especially if needed real-time
  • ZKPs are based on Math Intelligence, so not black boxes, but interpretation is more difficult than standard workflows
  • Model Drift as data changes

Why Now?

• Verification resources can be federated (similar to cryptocurrency blockchains) since no sensitive data is required in this step
• Medicine is finally digital and massive with multi-omics domains
• ZKP will enable ML models to be trained with larger, more diverse data points
• ZKP, along with fixed coordinate data model can achieve “actual” Clinical Decision Support System

Instead of can we afford to do it, the question should be can we afford not to do it?

Thank You!

This presentation was made using Quarto. Quarto enables you to weave together content and executable code into a finished document. Click here learn more about Quarto.