Migaco: Redefining Smart Toothbrushing through Geomagnetic Sensing and Deep Learning

TL;DR

Migaco is an innovative, low-cost system designed to help young children develop healthy toothbrushing habits. Eschewing expensive electronics, it uses a simple magnet attached to a toothbrush and a smartphone's geomagnetic sensor. By leveraging Convolutional Neural Networks (CNNs) to interpret magnetic field disturbances, the system achieves over 98% accuracy in detecting which teeth a child is brushing, all while turning the chore into an interactive game.

The Problem: Why "Smart" Toothbrushes Fail Children

Traditional "smart" toothbrushing solutions often suffer from a design mismatch for their target audience: children. Previous SOTA (State-of-the-Art) products like GUM PLAY utilize 3-axis accelerometers and Bluetooth modules.

• Durability: Small children often treat objects violently; fragile electronics break easily.
• Maintenance: Battery charging and synchronization are barriers to daily use.
• Cost: High price points prevent wide-scale adoption.

The author's research intuition was to move the "intelligence" from the toothbrush to the smartphone, turning the brush into a passive, indestructible peripheral.

Methodology: High-Precision 1D Signal Recognition

The core technical challenge of Migaco is Information Scarcity. A smartphone's geomagnetic sensor typically provides 1D orientation data (alpha rotation). Distinguishing between brushing the "front-right" vs. "front-center" of the teeth using only 1D data is noisy and complex.

1. Hardware Architecture

The setup is elegantly simple: a magnet on the brush and a smartphone fixed in front of the child's face. As the magnet moves, it disturbs the local magnetic field.

2. Deep Learning Strategy

To solve the classification problem, the author compared three models:

• MLP (Multi-Layer Perceptron): A baseline for pattern matching.
• RNN (Recurrent Neural Network/LSTM): Theoretically suited for time-series data.
• CNN (Convolutional Neural Network): Typically for images, but here used to analyze "graphical" patterns of motion.

Surprisingly, CNN outperformed RNN. The author noted that toothbrushing is a repetition of short, specific actions rather than a long-term sequence (like language), making the spatial pattern-matching of CNNs more effective for this specific task.

Experimental Results & Validation

The system was optimized through several iterations:

• Sampling Frequency: Data collected every 50ms.
• Dataset Overlap: Method 2 (heavily overlapping windows) significantly boosted accuracy compared to non-overlapping windows.
• Accuracy: The CNN model reached 98.3% accuracy for 3-zone detection and 95.5% for 4-zone detection (distinguishing upper vs. lower teeth).

The system was implemented as a Web Application (HTML5/JavaScript) communicating with a Python-based backend. In real-world tests with children aged 4-6, parents reported high engagement and improved communication.

Critical Insight & Conclusion

Migaco’s true value lies in its Inductive Bias. The author discovered that distinct brushing zones create unique "signatures" in the magnetic field not just because of position, but because of individual habits (e.g., the brush tip moves more when cleaning lower teeth).

Takeaways:

• Passive Sensing: Proves that magnets + smartphones can replace expensive IMUs for specific localized tasks.
• CNN for Time-Series: Reinforces the trend of using "Computer Vision" logic to solve signal processing problems when the signals are repetitive.
• Limitation: The system currently requires individual training because magnetic environments and brushing habits vary per person. Future work must focus on Generalization or an easier "calibration" phase.

In the landscape of "AI for Good," Migaco stands out as a practical, human-centric application that prioritizes user experience and cost-efficiency over hardware complexity.