IMU Insole for Diabetic Neuropathy Gait Science Fair

ISEF Category: Biomedical Engineering

Subcategory: Biomechanics  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Diabetic peripheral neuropathy quietly damages foot nerves until small wounds turn into amputations. The earliest sign is a subtle change in how someone walks. A $20 IMU sensor in an insole picks up that change. A small transformer model deployed on-device flags it before anyone notices a limp.

What Is It?

An IMU (inertial measurement unit) combines accelerometer and gyroscope sensors. Cheap modules like the MPU-6050 fit under a heel and log motion at hundreds of samples per second.

Plantar tilt and heel-strike timing are early-warning features of diabetic neuropathy. PhysioNet hosts public diabetic-foot gait datasets you can train on.

A transformer is a deep-learning model designed for sequences. A small variant runs on a microcontroller, which is essential for daily wear. The whole pipeline turns a $20 sensor into a wearable screening tool.

Why This Is a Good Topic

Wearable diabetic-foot monitoring is an active research area and the materials are inexpensive. You will learn sensor calibration, sequence modeling, edge deployment, and validation against public clinical data.

Research Questions

• How does IMU sample rate change classifier balanced accuracy?
• What is the effect of insole stiffness on signal-to-noise ratio?
• Does the transformer outperform an LSTM on PhysioNet diabetic-foot data?
• To what extent does subject demographics affect calibration?
• Which feature window length best detects early neuropathy?
• How does on-device quantization shift accuracy?
• What is the effect of sampling frequency on battery life?

Basic Materials

• MPU-6050 or MPU-9250 IMU module.
• ESP32 development board.
• LiPo battery and TP4056 charge module.
• Foam insoles to embed sensors.
• Soldering iron and jumper wires.
• Laptop with PlatformIO or Arduino IDE.

Advanced Materials

• Medical-grade insole pressure mats.
• Calibrated treadmill for gait collection.
• Clinical mentor for diabetic-cohort recruitment.
• High-precision IMU like the BNO086.

Software & Tools

• PyTorch: Trains the transformer model.
• TensorFlow Lite Micro: Deploys the model on ESP32.
• PhysioNet datasets: Provides labeled training data.
• Python (scikit-learn): Computes reliability diagrams.

Experiment Steps

1. Lock the insole hardware and document calibration.
2. Decide the feature window length and target labels before training.
3. Build a strict subject-wise data split.
4. Train the transformer with cross-validation and report calibrated probabilities.
5. Quantize and deploy the model on-device.
6. Compare on-device accuracy and latency to laptop inference.

Common Pitfalls

• Mounting IMUs differently between subjects, which shifts axes.
• Mixing subjects across train and test sets.
• Reporting accuracy on imbalanced classes without balanced metrics.
• Skipping reliability diagrams and trusting raw probabilities.
• Forgetting to measure power draw at the chosen sample rate.

What Makes This Competitive

Move past simple classification. A competitive project runs subject-wise data splits, reports calibrated probabilities with reliability diagrams, and compares the transformer to an LSTM baseline. On-device latency and battery-life measurements turn the demo into a real wearable.

Project Variations

• Add a thermal sensor for ulcer-risk warning.
• Replace transformer with a tiny CNN and compare battery life.
• Test the insole during stair climbing for added challenge.

Learn More

• PhysioNet diabetic foot datasets: Open access with documentation.
• PubMed: Search diabetic peripheral neuropathy gait reviews.
• TensorFlow Lite Micro guides: Free deployment tutorials.
• NIH PubMed Central: Open-access papers on wearable diabetic monitoring.
• MIT OpenCourseWare: Course 6.S191 Introduction to Deep Learning.