Nursie

Nursie: The Invisible Safety Net — Dignity-First AI Surveillance for the Next Generation of Care.

Comment

Building an “AI Nurse” for Long-Term Care: mmWave Fall Detection + Sensor Fusion + Voice Support

Falls among older adults are a major public-health issue in Canada. They’re a leading cause of injury-related hospitalizations and deaths for people 65+, and the trend is moving in the wrong direction. National surveillance data shows fall-related hospitalizations for older adults rising 47% between 2008 and 2019 (from 49,152 to 72,392). :contentReference[oaicite:0]{index=0}
On the mortality side, the burden is also climbing (e.g., 5,581 older adults died due to a fall in 2019, and deaths due to falls increased strongly in recent years). :contentReference[oaicite:1]{index=1}

In Québec long-term care settings (CHSLD), the problem is often amplified by two realities:

  • Falls can happen unwitnessed (especially at night).
  • Staffing shortages can delay assistance.

Those delays matter. When someone remains on the floor for a prolonged period (“long lie”), the complications can be severe—hypothermia, dehydration, tissue damage, rhabdomyolysis, pressure injuries, kidney failure, and more. :contentReference[oaicite:2]{index=2}

This post explains our prototype system: an AI nurse assistant that combines privacy-preserving room sensing, wearable vital signs, and a voice interface, all connected through a local on-site network.

TL;DR

We’re building an on-premises assistant for long-term care that:

  • Detects falls using mmWave radar + a floor vibration mat
  • Tracks context with door open + presence/occupancy
  • Monitors vital signs via a smartwatch
  • Sends immediate alerts (phone call + dashboard) to staff on emergencies
  • Builds a resident profile over time in MongoDB to flag unusual behavior (e.g., leaving the apartment at night)
  • Provides a voice interface for residents (questions, appointment reminders, medication prompts)

Design Principles

1) Fast, reliable detection > fancy features

If a fall is detected late, the system fails its core mission. Our priority is high confidence + fast alerting.

2) Sensor redundancy to reduce false positives

No single sensor is perfect:

  • Radar can confuse some edge cases (e.g., abrupt sitting, blankets, furniture occlusion).
  • Vibration sensors can trigger on dropped objects.
  • Wearables can be off-wrist or missing.

So we fuse signals to improve confidence.

3) Privacy by architecture

We keep everything inside the facility’s LAN:

  • No “always-streaming camera”
  • No raw data leaving the network by default
  • Local processing wherever possible

System Overview

At a high level:

  • *Room module Ipad *: the “AI nurse” endpoint in the resident room for voice interaction.

  • mmWave radar: continuous position/activity sensing + fall detection

  • Floor vibration mat (prototype currently cardboard): detects “impact/vibration signature” when something hits the floor

  • ESP32 sensor nodes running FreeRTOS:

    • ESP32 #1: vibration mat sensor → ROS 2 messages
    • ESP32 #2: door magnetic sensor + PIR presence sensor → ROS 2 messages

  • Wearable (smartwatch): heart rate monitoring (vital signs)

  • Local network: our own router flashed with OpenWRT, hosting a private LAN for all devices

Backend + Dashboard

  • Emergency alert workflow: call nurse + dashboard alert
  • Resident behavior profile + anomaly detection stored in MongoDB
  • AI workflow (Gemini + ElevenLabs + webhooks)
    • ElevenLabs posts send_event webhooks to the backend over HTTPS
    • Gemini triages transcripts into structured priority / intent / summary
    • The backend persists triage + events in MongoDB
    • Triggers escalations when critical

Challenges we ran into

  • Integrating mmWave radar was difficult due to environmental noise; distinguishing a human heartbeat from a moving curtain required careful signal-processing tuning.
  • We had to solve “Connection Errors” in our live telemetry stream so the dashboard stays updated even when local Wi-Fi fluctuates.

Accomplishments that we're proud of

  • A system that is compliant by design with Québec’s privacy framework (Law 25 / Act respecting the protection of personal information in the private sector).[^5]
  • A local AI agent that can differentiate between a “soft landing” on a bed and a critical fall on the floor.
  • Bridging the gap between hardware (LiDAR, Radar) and a professional-grade nurse dashboard that prioritizes alerts.

What we learned

  • The importance of Edge Computing—processing data where it’s collected to maximize privacy and speed.
  • As mechanical and software engineering students, we learned that the best medical tech is “invisible”: it provides the most protection when it requires the least intervention from the user.
  • We gained deeper insight into the ethical implications of AI in geriatric care.

What's next for Nursie

  • Refine our Apple Watch integration as a secondary “gold-standard” validator for high-risk patients.
  • Explore predictive models that analyze gait patterns to estimate fall risk before the first fall occurs.
  • Scale the hardware into a sleek, medical-grade casing that can be deployed across CHSLDs throughout Québec.

Built With

Share this project:

Updates