Abstract
Researchers have developed "Magic Monster," a mobile exergame designed to transform tedious wrist rehabilitation for children with cerebral palsy (CP) into an engaging magic-themed challenge. By leveraging standard smartphone accelerometers and gyroscopes, the system utilizes a Convolutional Neural Network (CNN) to accurately classify therapeutic movements. Crucially, the system automatically adjusts difficulty to match each child's unique range of motion, ensuring both accessibility and therapeutic progress.
Introduction
For children with cerebral palsy, physical therapy is a lifelong necessity that often becomes repetitive and unmotivating. Exercises such as dorsiflexion—bending the wrist upward—are vital for maintaining motor skills but are frequently neglected due to their boring nature. This research addresses the dual challenge of increasing engagement through gamification and ensuring exercise accuracy through automated movement classification. Because every child with CP has a different baseline range of motion, a standard detection system is insufficient; an adaptive approach is required to meet the child where they are.
System Design: The Magic Monster
The researchers developed a mobile application housed within a monster-shaped plush sheath. The smartphone screen serves as the monster's face, providing immediate visual feedback as the child performs "magic tricks" via wrist movements.
This physical interface makes the device less intimidating and provides a better grip for children with limited dexterity. The system includes a five-movement calibration phase, allowing the game to adapt to the specific physical capabilities of the user rather than enforcing a rigid standard.
Data Processing and Classification
The movement recognition engine relies on data from the phone's internal accelerometer and gyroscope.
The team extracted 42 statistical features (including mean, variance, and skew) from the three axes of both sensors. Using the Minimum Redundancy Maximum Relevance (mRMR) algorithm, they identified the 21 most effective features. Notably, the standard deviation of the gyroscope's x-axis emerged as the strongest indicator for range of motion.
While Long Short-Term Memory (LSTM) networks are often the standard for time-series data, this study found that a Convolutional Neural Network (CNN) achieved superior performance. The CNN's spatial filters were better suited to capturing the specific "burst" patterns characteristic of wrist dorsiflexion.
Results and Adaptive Difficulty
The CNN model achieved an impressive accuracy of 0.970 on independent test sets and maintained a high accuracy of 0.948 during field tests with children with CP. The system demonstrated higher precision (0.929) than recall (0.846), ensuring that children are not incorrectly credited for improper movements.
To maintain motivation, the system employs a rule-based adaptive difficulty mechanism. By monitoring the last 10 movements, the game increases the required range of motion or speed if the success rate exceeds 90%, or lowers the threshold if it falls below 60%.
Conclusion and Future Work
Magic Monster demonstrates the potential for low-cost, sensor-based mobile games to support home-based rehabilitation. While the current results are promising, the study's small sample size (8 children) and reliance on healthy adult data for initial training suggest areas for improvement. Future research will focus on longitudinal studies to evaluate long-term engagement and the impact of sensor hardware variations across different smartphone models.
Summary updated