Table 1 Comparative analysis of related work.

Xie et al. (The present study)

An accelerometer is utilized

The peak power consumption is 54.05 mW

Eight types of upper limb rehabilitation movements are analysed

The testing accuracy achieved is 95.09%, with a final deployment accuracy of 88.01%

Lui and C. Menon.³³

One wrist-worn inertial device (MPU9250) and an Omron D6T thermal sensor

-

Twenty-four types of upper limb movements

The classification accuracy for the 24 upper limb movements reaches 93.55%

Giordano, et al.⁵²

One accelerometer

Peak power consumption is approximately 8 mA. 8 mA represents the operating current of the device during operation (the common voltage of NRF52832 is 3 V, and the corresponding power consumption is 24mW)

four different usage classes (Tool transportation, no-load, metal, and wood drilling)

achieving an

accuracy of 90.6% with a model size of approximately 30 kB

Wang et al.⁴²

The upper limb wearable rehabilitation training data collection device is composed of three inertial sensor units, MPU6050

–

There are eight types of upper limb rehabilitation exercises

The CNN-LSTM model demonstrates the best performance with an identification accuracy of 99.67%, followed by the multilayer LSTM model with an identification accuracy of 97.00%

Basterretxea et al.⁴⁵

The device utilizes accelerometers and stretch sensors to capture data

1.13 mW sensing and 11.24 mW computation power consumption

It can distinguish between various activities such as walking, sitting, standing, driving, lying down, jumping, and the transitions between these states

With a power consumption of less than 12.5 mW, the system achieves an accuracy of 97.7% in recognizing six activities and their transitions

Alessandrini et al.⁴³

photoplethysmogram (PPG) sensors and accelerometers

–

the device can differentiate between walking, ascending stairs, descending stairs, sitting, standing, and lying down

Average test time is 150.1 at the

highest test accuracy of 95.54% on

the microcontroller unit (MCU)

Xu and Yuan⁴⁶

The accelerometers are worn on the mid-waist and front of the ankle

–

The device can distinguish between nine types of activities: walking, running, ascending stairs, descending stairs, resting, standing, sitting-to-standing, standing-to-squatting, and squatting-to-standing

The average recognition rate for these nine activities is 97.26%, with an identification rate of 98.56% for the four transition movements

Choudhury et al.⁴⁷

The device employs accelerometers, microphones, and light sensors

Power consumption is approximately 43 mW

(with all sensors running

continuously)

It can differentiate between activities such as walking, running, cycling, using an elliptical machine, and using a stair machine

Highest Accuracy is 93.8% (Inference

v2.0 with temporal information

included and supervised learning)

Li et al.⁴⁴

Two IMU modules (MPU9250), each comprising a tri-axis accelerometer, a tri-axis gyroscope, and a tri-axis magnetometer

–

Five types of upper limb rehabilitation exercises: forearm pronation & supination, lumbar touch, shoulder touch, shoulder anteflexion, and shoulder extension

The overall recognition accuracy for five motions achieves 99.34%