A wearable rehabilitation sleeve that reads muscle activation in real time and tells you exactly when you're doing it right, and when you're overdoing it.
Over 100,000 ACL surgeries are performed in the US each year. Within days of surgery, quadriceps atrophy begins — and a phenomenon called Arthrogenic Muscle Inhibition (AMI) means the nervous system actively suppresses the very muscles patients need to train.
The result: patients can't tell whether they're activating the right muscles at all. Without real-time feedback, at-home rehab becomes guesswork, and adherence drops sharply without a clinician present.
ThreshHold is a wearable thigh sleeve with an embedded EMG sensor. It monitors electrical muscle activity in real time and communicates through two LED signals: green for correct activation, red for fatigue or overexertion.
An adaptive algorithm using Welford's online method calculates a personalised activation threshold each session, so the system adjusts to each patient's baseline rather than relying on fixed values. A Python GUI logs session data and generates an AI physiotherapy summary to track progress over time.
My core responsibility was the activation detection algorithm. I evaluated several approaches — including fixed-threshold, peak-detection, and moving-average methods — and ran comparative tests to identify which performed most reliably on noisy, low-amplitude EMG signals. The Welford online z-score method worked the best. It adapts to each patient's baseline without storing all prior samples, making it both accurate and lightweight enough to run on the Arduino.
On the hardware side, I handled the complete circuit wiring: connecting the MyoWare 2.0 EMG sensor, Arduino Uno, HC-05 Bluetooth module, and LED outputs, including power regulation and signal routing. I also configured the HC-05 for serial communication so data streams wirelessly from the sleeve to the laptop during a session.
The sleeve operates in a simple state machine. Each session opens with a calibration window, moves into active coaching when the patient starts exercising, and signals completion with a slow red pulse when the session ends.
ThreshHold Python GUI calibration and live session view
ThreshHold Python GUI end-of-session stats and AI coaching
ThreshHold housing Rev 1 CAD render
ThreshHold housing Rev 2 CAD render
Circuit: MyoWare 2.0, Arduino Uno, HC-05, LEDs
The hardest part of this project was working with the EMG signal itself. EMG data from the MyoWare 2.0 is noisy, highly person-dependent, and sensitive to electrode placement. Testing multiple detection algorithms against real muscle data made me realize that a method that looks clean in theory often falls apart on biological signals. Fixed thresholds failed immediately across different users. The z-score approach worked because it stops assuming what a "normal" signal looks like and just asks: is this activation meaningfully different from this person's own baseline?
Working on the wiring also surfaced something I hadn't expected: most of the hardware debugging time was spent on the Bluetooth link, not the sensor. The HC-05 has strict AT-command timing requirements, and mismatched COM ports between the display device and the Arduino firmware caused silent failures that looked like sensor problems. That experience shifted how I approach hardware integration. It made me isolate and validate each communication layer independently before combining them.
The AI-assisted design process was genuinely different from any ideation method I'd used before. Seeing 30 human ideas clustered in a t-SNE map made the gaps in our thinking visible in a way that a list or a Pugh chart never could. The concepts the LLM generated to fill those gaps weren't always good, but they broke us out of the implicit constraints we'd all been working within. The final design carries ideas from both human and AI rounds.
If I were to continue this project, I'd prioritise two things: a more rigorous cross-user validation study to stress-test the adaptive threshold across different muscle compositions and injury states, and a cleaner physical form factor. The current neoprene sleeve works, but long-term adherence in a real rehab context would require something a patient actually wants to put on every day. This is a design challenge that I want to work on.