VitalGuard

The Challenge

Clinical decision systems suffer from high latency and disjointed telemetry. Vitals are recorded, but standard systems use hardcoded threshold alarms (e.g. "Heart Rate > 120 = Alert") without understanding the patient's baseline, leading to severe alarm fatigue.

We needed a system that could stream continuous vital data, contextualize it against the patient's unique history, and catch anomalies in real-time before they became critical.

Architecture

VitalGuard was built as an event-driven microservice architecture to handle high-frequency streams:

Ingestion: Node.js WebSockets handling continuous streams of vital telemetry.
Processing: FastAPI layer routing time-series windows to an LLM-powered Agent.
Decision Engine: Agentic workflows evaluating vitals against historical baselines.
Frontend: React application for live telemetry visualization and clinician alerts.

Tradeoffs

Agentic Inference vs Deterministic Rules: We traded deterministic rule execution for LLM-driven inference to catch subtle, multi-variable anomalies (e.g., slowly dropping O2 coupled with spiking HR). This increased processing latency slightly but dramatically reduced false positives.
Polling vs WebSockets: Opted for WebSockets for bidirectional real-time communication, trading infrastructure simplicity for the low latency required by clinical environments.

Failures & Iterations

Inference Bottleneck: Initially, feeding every continuous data point to the LLM choked the system and resulted in massive API limits and latency.
- Resolution: Implemented a localized sliding window approach in the Node layer. We only triggered the Agent when a local standard deviation threshold was crossed, using deterministic math as a first-pass filter before invoking the expensive AI context.

Outcome

Won 1st Place at HackaNova 5.0 and DSH Hacks V1.
Demonstrated scalable healthcare telemetry processing capable of handling 500+ simulated events per second with sub-2s anomaly detection.

Lessons Learned