Jenny Ji Hyun Kim

Precision Health [Wearables]

Re-Align: Posture Detecting with Novel Capacitive Sensor

[2022]

Center for Project-Based Learning

ETH Zürich

Ji Hyun Kim, Dr. Christian Vogt, Dr. Seonyeong Heo

Neutral vs. Forward Head Posture [Koseki et al., 2019]

BACKGROUND

In our increasingly sedentary society, changes in our physiology are becoming evident, especially due to the prolonged use of smartphones and extensive hours spent in front of computers. A significant consequence of this lifestyle is the development of ‘Forward Head Posture’ in many individuals. This posture alteration primarily impacts the musculoskeletal system, often leading to chronic pain that can result in frequent headaches. This change in our posture may compress our organs in unusual ways (i.e. our lungs become compressed due to our rounded shoulders). Its downstream effects may negatively affect various organ systems such as the respiratory and cardiovascular systems.

SOLUTION

To combat this issue, we have introduced a smart necklace designed to track and analyze posture patterns. This device is enhanced with TinyML, enabling it to process data efficiently on a microcontroller unit (MCU). It employs a range of sensors, including an accelerometer, gyroscope, and a novel capacitive sensor known as QVAR. Leveraging this sophisticated technology, this smart necklace provides insightful feedback and practical suggestions for improving posture, thereby helping to reduce the health risks associated with poor postural habits

CONTRIBUTIONS

I brought the concept of a posture-detecting necklace to the Project-Based Learning (PBL) laboratory. Working alongside electrical engineers in our team, we successfully developed a prototype of this innovative device. Then, I collected data from a set of participants performing activities of daily living (in good and poor posture) as well as a few physiotherapy exercises. I dedicated considerable effort to training and evaluating various traditional machine learning and deep learning models. This ensured that our device could effectively analyze and offer feedback on posture. Of all models, Random Forest achieved the highest accuracy with the fastest training time.

Training Time vs. Accuracy for Different Models and Sensor Configurations

Confusion Matrix:

RNN-LSTM, Generalized

Confusion Matrix: TCN, Individualized

Confusion Matrix:

Random Forest, Generalized

MOTIVATION

My motivation for this project stemmed from my experience with practicing yoga. My practice highlighted the importance of the skeletal system and how it’s significantly influenced by posture. In today’s society, where occupational demands and social pressures often lead to excessive phone use and performance stress, there’s a lack of focus on improving posture. Existing posture-detecting devices rely on basic algorithms that calculate deviations from a calibrated ‘zero’ position. By developing a machine learning-enabled, stylish smart-necklace, my goal was to enhance user retention and encourage a greater awareness of posture.

Made in Palo Alto with lots of Sunshine and Coffee