Real-Time Patient Deterioration Prediction System

• 

  Dashboard

• 

  Patient Detail with AI Assessment

Inspiration

84% of in-hospital cardiac arrests show documented warning signs within 8 hours of the event; yet traditional intermittent monitoring often misses this critical window. We wanted to transform this 8-hour warning window into actionable clinical intelligence using real-time streaming.

What it does

A real-time clinical decision support system that predicts patient deterioration using streaming vital signs and laboratory data. Clinicians get a priority pool of high-risk patients filtered by risk level and AI confidence, with Gemini-powered assessments that explain why each patient is flagged, reducing alert fatigue while maintaining sensitivity.

Architecture

How we built it

• BigQuery hosts MIMIC-IV data, enabling SQL-based extraction of vitals and labs for 500 ICU patients
• Confluent Kafka ingests vital signs and lab results in real-time, joined by patient stay ID
• Automated MEWS scoring calculates risk on every incoming data point
• Vertex AI Gemini 2.5 Flash provides zero-shot clinical reasoning with confidence-aware assessments
• React dashboard with risk filtering, trend charts, and real-time alert feed
• Cloud Run + GCE for deployment on Google Cloud

Challenges we ran into

• Stream joining: Synchronizing vitals (continuous) with labs (intermittent) required temporal windowing with a 4-hour staleness threshold
• Alert fatigue: Traditional early warning scores have PPV as low as 4-7%. We implemented trend-based confirmation requiring 2+ consecutive elevated readings
• AI confidence calibration: Teaching Gemini to express appropriate uncertainty when vital signs are incomplete or labs are stale

Accomplishments that we're proud of

• End-to-end streaming pipeline processing 266K vital records and 25K lab results in real-time
• Clinically validated MEWS scoring with SpO2 integration
• Confidence-aware AI that explains its reasoning, not just the risk score
• Filtering system that turns 500 patients into a focused priority pool

What we learned

• Zero-shot LLMs can provide clinically useful interpretations without training data
• The real value isn't just prediction, but reducing cognitive load for clinicians
• Streaming architecture enables intervention during the "Golden Hour" that batch processing misses

What's next

• MIMIC-CXR integration: Chest X-ray analysis for respiratory deterioration
• Nursing notes: Unstructured text analysis for subtle clinical cues
• Federated learning: Multi-hospital deployment without sharing patient data

Built With

• css
• flask
• gcp
• google-bigquery
• python
• react
• typescript