🦾 Stroke Rehab System

An end-to-end wearable IoT and Edge-AI platform for real-time stroke rehabilitation scoring.

This project is a complete hardware-to-software pipeline designed to track, analyze, and gamify physical therapy for stroke patients. By combining custom dual-IMU wearables with a mobile-optimized 1D Convolutional Neural Network (CNN) running locally on iOS via Zetic, the system provides zero-latency movement quality scores and voice feedback.

🏗 High-Level Architecture

The system is divided into three core pillars:

1. Embedded Hardware: Captures raw 6-DOF human movement data at 80Hz.
2. AI Scoring: Processes the time-series data to evaluate movement quality.
3. Swift App: Manages the BLE connection, runs the Edge AI, and guides the patient.

---

⚙️ 1. Embedded Hardware (Data Generation)

The wearable component is built from the ground up for low latency and high data throughput, acting as the foundation for the AI model.

• Microcontroller: ESP32-C3 powered by a 4.8V LiPo battery.
• Dual Sensors: 2x MPU6050 (Bicep and Wrist). Both sensors run on a shared hardware I2C bus utilizing custom addressing (0x68 for Bicep, 0x69 via AD0 pin for Wrist) to prevent collisions.
• Diagnostics: An integrated I2C OLED screen (0x3C) provides real-time device status and debugging.
• Custom BLE Pipeline: * The ESP32 unpacks and filters accelerometer and gyroscope axes before transmission.
  • An expanded MTU size blasts a tightly packed 76-byte payload over Bluetooth Low Energy (BLE) at 80Hz.
  • Payload includes: timestamp, raw accel/gyro, and bicep_PRY, wrist_PRY channels for both sensors.

---

🧠 2. AI & Machine Learning (Zetic Edge Inference)

The AI pipeline evaluates movement quality by comparing patient data against the JU-IMU dataset (Stroke vs. Healthy subjects). The model is specifically optimized for small-dataset generalization and on-device export.

Preprocessing & Signal Alignment

• Side-Aware Selection: Dynamically swaps sensor channels based on whether the patient has Left or Right hemiparesis, ensuring the affected limb always maps to the same input channels (resulting in a 12-channel input).
• Global Normalization: Uses a global $(x - \mu) / \sigma$ over all training time-steps rather than per-sample Z-scores. This is crucial as it preserves the cross-patient magnitude differences that distinguish weak stroke movements from healthy ones.
• Interpolation: Variable length movements are linearly interpolated to exactly 128 timesteps.

1D CNN Architecture

The model utilizes a "shrink time, grow features" pattern suitable for continuous sensor streams:

• Input: (Batch, 12, 128)
• Conv Blocks: Three stride-2 blocks with progressively decreasing kernel sizes (7 → 5 → 3) to capture wide temporal context early and refine local features later.
• Mobile Optimizations: Uses BatchNorm (highly stable on mobile vs. GroupNorm) and a fixed AvgPool1d(kernel=16) to avoid compatibility issues with mobile converters.
• Output: Generates a 0–100 movement quality score based on softmax confidence probabilities.

Export to Zetic

The trained PyTorch model is packaged via torch.export into a deterministic .pt2 graph, along with the saved global $\mu$ and $\sigma$ constants. This ensures identical preprocessing and ultra-fast local inference on the iOS client.

---

📱 3. Swift App (Frontend & Orchestration)

The iOS application serves as the command center for the patient, processing the hardware data and surfacing the AI insights.

• Real-Time Monitor: Uses CoreBluetooth to subscribe to the ESP32's custom characteristic, caching the 80Hz stream into 128-timestep sliding windows.
• Local Inference via Zetic: The Swift app holds the trained .pt2 AI model. It passes the cached data window into the Zetic runtime, executing the CNN entirely on-device (no cloud computing delay).
• Exercise Guide & Gamification: Translates the AI's 0–100 movement quality score into a visual performance dashboard.
• Audio Feedback: Integrates ElevenLabs TTS to provide encouraging, real-time voice guidance to the patient based on their workout selection and current performance score.

📚 References & Resources

• JU-IMU Dataset & AI Research: Oh, Y., Choi, S.-A., Shin, Y., Jeong, Y., Lim, J., & Kim, S. (2024). Investigating Activity Recognition for Hemiparetic Stroke Patients Using Wearable Sensors: A Deep Learning Approach with Data Augmentation. Sensors, 24(1), 210. DOI: 10.3390/s24010210
• Dataset Repository: youngminoh7/JU-IMU on GitHub
• Hardware Datasheet: MPU-6050 Product Specification (PDF)

Made with ❤️ for LA Hacks 2026 by Mj, Ethan, Scott, Ian