Predictive Activity Monitoring via Wearable Sensors

Authors: Yutong Lu, Pourya Momtaz, Yiran Wang

Datathon 5, CHL5230, Dalla Lana School of Public Health

Introduction

• Investigated using wearable sensor data to predict human activities via deep learning algorithms.
• Aimed to find the best model for accurate activity prediction, crucial for healthcare monitoring and tailored interventions.

Data Engineering Process

• Utilized sequential sensor data from nine subjects, featuring activity labels and measurements from ankle and arm sensors.
• Preprocessed data by organizing subjects' data, splitting into training and test sets, and generating sequences for LSTM modeling.

Analysis

• Explored LSTM models with different architectures, learning rates, and regularization techniques.
• Found that a learning rate of 0.001 achieved the best training accuracy, while the model struggled with imbalanced data.

Findings

• Models performed well in predicting the majority class (Activity 0), indicating challenges with imbalanced data.
• Attempted downsampling and regularization but faced difficulties capturing patterns across all activity classes effectively.

Conclusion

• LSTM model showed good accuracy but struggled with imbalanced data, particularly predicting other activity classes besides the majority. Emphasized the need for robust strategies in handling imbalanced data for real-world healthcare applications.

Resources

• Presentation slides available

References

• Included references from studies on human activity recognition based on wearable sensor data.
• For more comprehensive details, please refer to the full report available in the provided GitHub repository.