Odyssey

Inspiration

In May 2018, in Kozhikode, Kerala, an acute encephalitis case passed through multiple healthcare providers before anyone recognized the underlying pathogen. By the time Nipah virus was officially confirmed, hospital-acquired transmission had occurred, resulting in a fatality rate near 89% (16 deaths out of 18 confirmed cases). Kerala has faced this acute threat repeatedly across 2018, 2019, 2021, and 2023. In each retrospective analysis, public health reports noted the same structural bottleneck: early outbreak detection relies entirely on identifying unusual clusters of atypical symptoms, but our healthcare infrastructure provides no real-time mechanism to detect that pattern across isolated institutions.

At the exact same time, rare disease patients face a heartbreaking 5-to-7-year "diagnostic odyssey." The knowledge to solve an obscure condition often exists—it is just locked in an electronic health record at a hospital three towns or three countries away.

Under privacy laws like India's Digital Personal Data Protection (DPDP) Act, HIPAA, and GDPR, hospitals cannot simply pool or transmit raw clinical charts. This creates the Data Silo Deadlock. Seeing recent lab breakthroughs like MedLatentDx (June 2026) prove that latent representations could allow private cross-institutional learning, we realized: a rare disease diagnostic search tool and an early-warning outbreak radar are the exact same infrastructure, viewed through two different lenses.

What it does

Odyssey is a privacy-preserving signal network that enables healthcare systems to collaboratively identify rare diseases and detect regional disease outbreaks without ever transmitting identifiable patient records or clinical narratives.

For the Clinician (Case Search): A doctor confronted with a confusing, undiagnosed patient can type a natural-language case summary. Instead of returning a black-box AI diagnosis, Odyssey queries a distributed registry of mathematical signatures to return: "Two other hospitals in the network observed a structurally similar symptom pattern—here are the anonymized match scores, relevant phenotypes, and institutional contact points."
For Public Health & Epidemiologists (The Surveillance Heatmap): By plotting stored vector signatures along temporal and geospatial coordinates, the same registry functions as a population-level outbreak radar. If three separate clinics in a 40 km radius log an anomalous spike in atypical acute encephalitic fevers over 72 hours, Odyssey clusters the mathematical signatures on a map—surfacing an outbreak cluster before it even has a clinical name.

How we built it

We built Odyssey as a zero-data-leakage pipeline where Protected Health Information (PHI) never crosses institutional perimeters.

+-------------------------------------------------------------------------------------+
| HOSPITAL LOCAL PERIMETER                                                            |
|                                                                                   |
| Free-Text Intake ──► [Phenotype Extractor] ──► HPO Term List: {HP:0001250}          |
|                                   │                                               |
|                                   ▼                                               |
|                    [Contextual LLM Judge]                                           |
|                 (Assesses Local Novelty)                                            |
|                                   │                                               |
|                              Unique? (Yes)                                          |
|                                   ▼                                               |
|                    [Sentence Encoder]                                                |
|       (384-dim vector z, HPO codes, confidence score)                              |
+----------------------------------------┬--------------------------------------------+
                                         ▼
+-------------------------------------------------------------------------------------+
| SHARED SIGNAL REGISTRY                                                              |
|                                                                                   |
| Dual-Lens Engine: Fast Cosine Vector Index + Geospatial Temporal Mapping            |
|                                                                                   |
| [ Lens A: Case Search ]      [ Lens B: Heatmap Radar ]                             |
| Clinical Pattern Retrieval   Regional Geospatial Clustering                         |
+-------------------------------------------------------------------------------------+

Phenotype Extraction (Clinical NLP): A local extraction service parses raw clinical text and binds findings strictly to the Human Phenotype Ontology (HPO) (e.g., "episodic burning in extremities" $\rightarrow$ HP:0012534). Standardizing to an ontology eliminates clinical jargon variance and establishes an invariant vocabulary across different hospital nodes.
Context-Aware Uniqueness Evaluation: Rather than using naive numerical similarity cutoffs that corrupt multi-agent calibration, each hospital runs a local evaluation step. An agent checks incoming cases against the facility's recent history to confirm genuine anomaly before emitting signals.
Latent Representation Layer: The structured HPO terms are projected into a 384-dimensional dense vector space:

$$\mathbf{z} = \text{Encoder}(\text{HPO}_1, \text{HPO}_2, \dots, \text{HPO}_k) \in \mathbb{R}^{384}$$

Patient names, raw narratives, and identifiers remain strictly within local hospital storage. The only data transmitted over the wire is:

$$\mathcal{P} = \big{\text{hospital_id}, \, \text{anon_case_ref}, \, {\text{HPO codes}}, \, \mathbf{z}, \, \text{confidence_score}\big}$$

Dual-Lens Vector Registry: The shared network stores vectors in a high-speed cosine-indexed vector repository:

$$S_C(\mathbf{z}{\text{query}}, \mathbf{z}{\text{stored}}) = \frac{\mathbf{z}{\text{query}} \cdot \mathbf{z}{\text{stored}}}{\Vert{}\mathbf{z}{\text{query}}\Vert{} \Vert{}\mathbf{z}{\text{stored}}\Vert{}}$$

Lens A: Computes nearest-neighbor similarity against stored signatures to facilitate cross-institutional peer discovery.
Lens B: Aggregates regional vector density across time to visualize early outbreak hot spots.

Challenges we ran into

The Multi-Agent Calibration Trap (AgentRx Precedent): Recent literature (AgentRx, May 2026) revealed that naive multi-agent voting architectures frequently underperform single models because overconfident nodes drown out nuanced signals. We ran into this early on: arbitrary cosine similarity cutoffs produced excessive false positives on common seasonal fevers. We solved this by pushing the qualitative evaluation upstream, requiring each node to reason over its own historical distribution before considering a case for network broadcast.
The Embedding Inversion Realization: A common misconception is that vector embeddings are inherently privacy-preserving. In reality, modern transferable inversion attacks can partially reconstruct clinical narratives from unmasked vectors. While cutting-edge research like MedLatentDx suggests full Key-Value attention cache distillation, implementing custom cross-attention KV distillation was out of reach for a rapid hackathon sprint. We balanced feasibility and honesty by transmitting normalized fixed-length vectors, cleanly documenting the mathematical abstraction layer, and earmarking calibrated Differential Privacy ($\epsilon$-DP) noise as our primary architectural next step.
Multi-Vocabulary Normalization: Clinicians rarely use the same phrasing for atypical symptoms. Early tests using raw sentence embeddings directly on free text yielded low cross-hospital similarity. Enforcing the HPO ontology mapping as an intermediary layer proved crucial to bridging semantic discrepancies.

Accomplishments that we're proud of

Unifying Two Huge Problems into One Architecture: Proving that solving the 7-year rare disease diagnostic journey and the 7-day outbreak detection window does not require two separate architectures—just one privacy-safe registry with two lenses.
A Working, End-to-End System: Going beyond academic paper artifacts by building an end-to-end demo with partitioned hospital nodes, structured HPO parsing, and real-time geospatial/vector visualization.
Zero PHI Leakage by Design: Creating an architecture that hospital compliance and legal teams under DPDP Act / HIPAA frameworks can actually evaluate, because identifiable clinical prose never crosses institutional firewalls.

What we learned

Move the Math, Not the Data: Healthcare doesn't need reckless centralization to build powerful predictive systems. Decentralized computing using abstract mathematical representations preserves sovereignty while enabling vital collective intelligence.
Retrieval Beats Black-Box Diagnosis: Doctors do not want an AI black-box telling them "This patient has Fabry Disease (91%)". They want evidence retrieval: "Hospital B saw an identical phenotype profile last year; here is their contact." Framing the system around peer-to-peer evidence discovery builds immediate clinical trust.
Standard Ontologies are Non-Negotiable: Machine learning models in healthcare require structured anchors. Relying on HPO codes over raw text transformed noisy embeddings into actionable signals.

What's next for Odyssey

Calibrated Differential Privacy ($\epsilon$-DP): Formally implementing and benchmarking Gaussian Differential Privacy noise injection directly onto the embedding vectors to provide quantifiable mathematical guarantees against reconstruction attacks.
True Latent KV Cache Distillation: Upgrading from fixed-length sentence embeddings to native LLM hidden-state and Key-Value (KV) cache block distillation with cross-family model alignment adapters (as demonstrated in MedLatentDx).
Multi-Modal Diagnostic Signals: Extending phenotype profiles beyond textual symptoms to include latent embeddings from genomic variant profiles (VCF data) and dysmorphology imaging.
Downstream Clinical Trial Matching: Connecting matched rare disease cases directly to active orphan drug trials on ClinicalTrials.gov and ICMR registries, turning late-stage diagnoses into immediate treatment pathways.