Developing an AI-Powered IoT Rehabilitation Wearable for Physical Therapy

In physical medicine and rehabilitation (PM&R), one of the biggest challenges we observed during real deployments is the "recovery gap" - the time between clinical sessions when patients are expected to perform exercises on their own.

The recovery gap is the time that a patient has to spend between clinical sessions, when they are responsible for performing exercises on their own.

In many cases, progress is impeded not by lack of motivation or willingness to perform exercises, but rather from performing exercises incorrectly and not having feedback about their performance or knowledge about whether their execution is correct.

This project was created for a practical purpose: To close that gap with a clinical-grade IoT rehabilitative wearable, which provides real-time patient feedback, and provides clinicians with objective, continuous data for tracking recovery through the use of embedded systems, biomechanics, and artificial intelligence.

Our approach combined embedded systems, biomechanics, and AI to create a solution that works reliably outside controlled clinical environments.

Most consumer-based fitness devices define and measure a user's fitness by counting step and calorie usage, whereas rehabilitation devices require significant precision in measuring the quality of movement.

Working with physiotherapists during the early prototyping phase, we learned directly from them what is truly important in recovering from an injury: Range of Motion (ROM), joint stability, and muscle engagement.

The system was designed to provide the following features:

• Real time feedback for correction of exercise
• Tracking recovery over a long period of time (e.g. improvement in knee flexion from ACL surgery)
• Detection of compensatory movements, which are a major cause of delayed or recurrent injuries.

The primary insight from the clinical partners was that "doing the exercise" is not sufficient - the success of recovery from an injury depends upon doing the task correctly and with quality. This understanding is what forms the basis for our system design.

Real-world usability considerations required balancing factors such as accuracy, comfort and battery life when designing the system. We created a multi-node sensor system comprised of multiple small wearable units positioned throughout the affected limb (e.g., thigh and shin).

Sensing Engine

The Bosch BNO055 was selected as the sensing technology to use in the solution. The BNO055 has an integrated 9-axis absolute orientation sensor, plus an on-chip sensor fusion co-processor that directly outputs quaternion and Euler angle data.

This significantly reduced the computational load on the main processor and improved the ability to respond in near real-time to changes in user motion.

Processing Unit

After evaluating several MCUs, we chose the Nordic nRF52840 Microcontroller (Arm Cortex-M4F) for its low power usage, BLE 5.3 support, and ability to continually send data without sacrificing battery life.

EMG Integration (Optional)

For advanced clinical users, we've included the option to add dry contact surface EMG (sEMG) sensors to track muscle activation.

This will help medical professionals and clinicians determine how well the patient is moving, as well as their recruitment and fatigue patterns of muscle groups, which are all significant factors when designing a rehabilitation plan.

We also had to solve practical challenges such as sensor placement variability, skin contact consistency, and noise due to movement, all of which required iterative hardware testing.

While raw sensor data may seem useful by itself, without transformation into useful insight, it's not valuable at all. A large portion of this project involved developing algorithms that fit biomechanical principles used by physiotherapists to allow derived form and movement patterns to be stored, archived and compared.

Quaternion-Based Motion Tracking

To avoid issues like gimbal lock during multi-axis joint movement, we processed orientation data in quaternion space. We chose to process each joint's orientation data in quaternion space. By using them as bases for calculating each joint's sequential orientation, we are able to accurately track the user's joint movement through each complex shoulder and/or hip exercise. This ensured accurate tracking even during complex shoulder or hip exercises.

Dynamic Range of Motion (ROM)

A monitored system calculates the user care of each of their joints via the peak angle of the user's joint, in real-time, and allows for comparison between:

• Patient's baseline
• Clinician defined target

And allows patients to immediately known whether they are achieving the desired movement.

Compensatory Movement Detection

Utilizing multiple sensor nodes, we were able to develop a logical sequence to identify incorrect movement patterns.

For example, during a squat, if the torso leans excessively forward instead of proper knee flexion, the system identifies it and triggers a haptic feedback alert.

The most challenging aspect of this functionality was to determine where the proper threshold of accuracy and sensitivity is used - If too many alerts occur, the user becomes frustrated.

From experience, we know that hardware alone doesn't solve the problem-it's the ecosystem that drives adoption. We developed a dual-interface system:

Patient Mobile App

• Real-time feedback during exercises
• Visual guidance using simple UI and 3D representations
• ""Perfect Rep" tracking to encourage correct form
• Gamified experience to improve engagement

Clinician Dashboard

• Compliance tracking (whether exercises are completed)
• Quality analysis (correct vs incorrect reps)
• Recovery trends (ROM improvements over time)

This allowed therapists to move from subjective assessments to data-driven decision-making.

Cloud Backend

We used AWS IoT Core with MQTT-based communication for reliable data transmission. The backend was designed with:

• Secure data pipelines
• Scalable architecture for multiple patients
• HIPAA-aligned data handling practices

Our pilot roll-out with physiotherapy partners resulted in some tangible improvements, including:

30% Increase in Patient Compliance

The real-time feedback and reminders encouraged patients to adhere more consistently to their rehabilitation programs.

15% Faster Achievement of ROM Targets

Through instantaneous identification and correction of errors, excessive movement from inefficient repetitions was minimized.

Less Clinical Workload

Therapists were able to concentrate on those who required intervention--allowing them to effectively manage 2x the number of patients.

In addition to the quantitative measures, we obtained clinically important qualitative measures from patient feedback as a result, patients reported feeling more confident executing exercises independently.

Through the use of IoT and AI, this initiative illustrates that rehabilitation can be transitioned from an intermittent, expressive process to a constant, information-oriented process. With the combination of wearable sensing, biomechanical intelligence, and cloud-based analytics, the new model was able to evaluate, quantify, and develop every action taken.