## Inspiration

Modern healthcare environments—especially hospitals, nursing homes, and home-care settings—struggle with continuous patient monitoring. Falls, seizures, wandering, and inactivity often go unnoticed until it is too late, leading to severe injuries or emergencies.

Caregivers must monitor multiple patients and camera feeds at once, which is unrealistic, exhausting, and prone to human error. We wanted to build a system that acts as an AI assistant for patient safety, available 24/7, scalable, affordable, and proven through real-world testing.

This inspired us to create IPBMS — Intelligent Patient Behavior Monitoring System, combining Computer Vision, AI, real-time streaming, and blockchain-backed integrity to bring hospital-grade monitoring anywhere.

## What it does

IPBMS provides end-to-end AI-powered patient monitoring:

Real-time detection of falls, seizures, inactivity, leaving the bed, and abnormal behaviors.
Edge AI processing on RTSP cameras or smart devices.
Automated snapshot hashing + on-chain verification using a Polkadot parachain.
Instant alerts for caregivers via mobile app and dashboard.
Role-based workflows: caregiver-first response, delayed customer alerts, acknowledgement, escalation, and resolution.
Medical context enrichment: image captions, recommended actions, and patient profiles.
Multi-channel notifications through FCM, WebSocket, SMS/Email (future).

It is deployable from a single home-care room to large multi-floor healthcare facilities.

## How we built it

We architected IPBMS as a unified, multi-layer ecosystem:

1. VisionEdge AI (Python)

RTSP camera ingest (IMOU, HikVision…)
Keyframe extraction
Fall detection (YOLOv8-Pose, MoveNet, PoseNet)
Seizure detection (VSViG transformer model)
BLIP captioning + NLLB translation
SHA-256 snapshot hashing
Polkadot on-chain verification module
Snapshot upload via API/MinIO

2. Backend (NestJS + PostgreSQL)

Event pipeline for all AI detections
RBAC (doctor, caregiver, admin…)
Subscription & billing
Polkadot proof syncing (txHash storage)
WebSocket + Supabase Realtime streaming
Audit logs and system health monitoring

3. Admin Dashboard (React + ShadcnUI)

Realtime event monitoring
Blockchain-verified event checker
Provider management, billing, analytics

4. Mobile App (Flutter)

Realtime caregiver alerts
30-second delayed customer alerts
Event viewer + medical info
Verified-on-chain badge for integrity

5. Blockchain Verification Layer (Polkadot)

On-chain storage of image hashes + metadata
Privacy-safe integrity proofs
Cross-chain ready
Low-fee, high-throughput substrate chain

Everything is containerized and orchestrated for cloud/on-prem deployment.

## Challenges we ran into

Realtime AI performance on edge devices with limited compute.
Ensuring high accuracy for fall & seizure detection in difficult lighting or occluded environments.
Reducing false positives, especially when patients sit up, lie down, or move naturally.
Synchronizing multi-layer pipelines (AI → backend → mobile).
Guaranteeing sub-second alerts even with network fluctuations.
Integrating blockchain without violating medical privacy.
Designing caregiver-first workflows that avoid alert fatigue.
Storage optimization for thousands of snapshots per day.

These challenges shaped the reliability and resilience of our final architecture.

## Accomplishments that we're proud of

Built a real-time AI monitoring system with end-to-end sub-second latency.
Successfully integrated Polkadot on-chain verification for medical incident integrity.
Achieved 87%+ fall detection accuracy and 72%+ seizure detection accuracy in real testing.
Designed a caregiver-first alert workflow that dramatically reduces false positives.
Created a full multi-platform ecosystem (Edge → Backend → Dashboard → Mobile).
Optimized the system to run even on Orange Pi 5+ and mini PCs.
Built a scalable enterprise-grade backend with RBAC, billing, and monitoring.
Verified the system in simulated real-world hospital environments.

## What we learned

Real-world data is messy, and AI models require constant tuning and threshold adjustments.
Healthcare workflows must prioritize caregivers, not just patients.
Blockchain can add trust and forensic integrity without storing personal data.
Multi-camera setups require smart frame selection to stay efficient.
Human-centered UX is essential for emergency systems — alerts must be clear, actionable, and verified.
Edge AI deployment is very different from cloud AI — optimization is everything.
Collaboration between AI, backend, mobile, and blockchain teams is crucial for a complex system like this.