Inspiration Night shifts in hospitals are where small changes turn into major emergencies. Staffing is thinner, cognitive load is higher, and clinicians must constantly decide: “Who is most concerning right now?” The problem isn’t lack of data. Hospitals are flooded with vitals, labs, and notes. The real issue is prioritization and communication under pressure. We built NOX to act as a Night Operations Copilot — helping teams rapidly identify deterioration trends and generate escalation-ready communication in seconds.
What it does NOX is a ward-level prioritization and communication engine. It: Ranks all patients by deterioration risk Detects recent change patterns across vitals and labs Surfaces interpretable drivers behind each ranking Visualizes ward risk with a heatmap Shows a timeline of deterioration for any patient Generates one-click: Night Rounds Brief SBAR escalation packet Provides conversational access through Ask NOX NOX is explicitly positioned as: A decision-support tool — not a diagnosis system. It accelerates prioritization and standardizes escalation communication during night operations.
How we built it NOX is built using: Python Gradio (Hugging Face Spaces deployment) NumPy / Pandas Matplotlib Structured JSON ward schema Architecture: Ward JSON ingestion Feature extraction from vitals, labs, and notes Trend detection (recent window deltas) Risk scoring engine Driver attribution + confidence scoring UI rendering: Ranked ward board Heatmap visualization Focus view timeline Communication artifact generation Everything runs in a lightweight, reproducible Space environment.
Challenges we ran into Balancing meaningful prioritization without overclaiming clinical authority Ensuring outputs were interpretable and not black-box Designing a workflow that matches real hospital communication (SBAR, rounds brief) Handling strict Gradio runtime changes under deadline pressure We prioritized stability and demo clarity over complexity.
Accomplishments that we're proud of Built a fully functioning ward-level prioritization engine under time pressure Designed outputs aligned with real clinical workflows Maintained explainability (drivers + evidence sources) Created a complete demo loop: Upload → Rank → Explain → Escalate Deployed live on Hugging Face Spaces Most importantly, we built something that feels usable — not theoretical.
What we learned In healthcare tools, interpretability matters more than raw complexity Judges respond strongly to workflow realism A clear demo flow is more powerful than a complex architecture Communication standardization (SBAR) is as important as detection
What's next for NOX Next steps: Integration with FHIR-compatible EHR exports Role-aware dashboards (nurse vs resident vs charge nurse) Calibration and uncertainty refinement Audit logging and explainability dashboards Evaluation on de-identified datasets Prospective simulation testing
Long term vision: NOX becomes a real-time prioritization layer running continuously in hospital night operations.
Built With
- gradio
- huggingface
- python
Log in or sign up for Devpost to join the conversation.