A closed-loop sensor mat with edge AI inference and autonomous micro-pneumatic actuation that prevents bedsores before they form. No nurse intervention required.
Each metric is sampled at 10 Hz, windowed over 5 minutes, and fed into the on-device CNN inference engine running on the FPGA.
Four hardware and software layers work together: sensors capture raw data, the FPGA fuses and infers, cloud aggregates and trains, and the dashboard alerts.
Every component was selected for a specific reason, not availability. Medical-grade sensing, deterministic real-time control, and privacy-first cloud.
| Component | Role | Why this part |
|---|---|---|
| 32x32 cap. array | Pressure mapping | 1024-cell resolution gives per-zone granularity impossible with commercial FSR mats. Detects localized pressure spikes before clinical thresholds. |
| MLX90640 | Skin temperature IR | Non-contact 32x24 IR array via I2C. Detects skin temperature differential caused by reduced perfusion before visual symptoms appear. |
| MAX30102 (x8) | Tissue oxygenation | Reflectance PPG nodes at high-risk anatomical zones. SpO2 drop at contact point is the earliest measurable biological signal of ischemia. |
| Lattice CrossLink-NX | CNN inference engine | Low-power FPGA with deterministic sub-millisecond inference latency. Runs quantized 8-bit CNN with no OS jitter. Parallel sensor stream processing. |
| STM32H7 | System controller | 480 MHz Cortex-M7 with dual-core, hardware FPU, 1 MB RAM. Runs FreeRTOS with PID control, safety interlocks, and telemetry serialization. |
| ESP32-S3 | Wireless gateway | Dual-mode Wi-Fi + BLE 5.2. Handles MQTT and BLE simultaneously without burdening the real-time controller. Proven hospital Wi-Fi compatibility. |
| Flower (flwr) | Federated learning | Privacy-preserving model training. Gradient aggregation across hospital beds without sharing raw patient sensor data. HIPAA and DPDP Act aligned. |
| FastAPI + InfluxDB | Cloud backend | Async Python backend with native WebSocket support. InfluxDB handles 10 Hz time-series telemetry from hundreds of beds efficiently. |
Compared to manual repositioning protocols and existing commercial smart mattresses.
Four distinct user groups interact with different layers of the system.
The current system is a research prototype. These are the next engineering priorities.