TL;DR
This paper introduces Deep Ventilation, a deep learning system that predicts physical effort by analyzing ribcage movement from a non-invasive sensor. It offers a more reactive and representative measure of effort compared to traditional heart rate monitoring, potentially leading to more accurate and personalized fitness tracking.
What Problem Does This Paper Address?
Global physical activity levels are declining, leading to a surge in wearable devices for tracking health metrics. While step counters and heart rate monitors are common, they have limitations. Step counts are a crude measure, and heart rate, though better, suffers from "cardiovascular drift" during prolonged exercise and is slow to react to rapid changes in intensity, like those in High-Intensity Interval Training (HIIT). The authors propose breathing as a more reactive and representative physiological signal for tracking physical effort, aiming to bridge the gap between perceived exertion and objective measurement.
Methods
The core of Deep Ventilation is a deep learning model, specifically a Long Short-Term Memory (LSTM) network, trained to predict minute ventilation (the volume of air a person breathes per minute) directly from ribcage movement data. This movement is captured by a non-invasive Respiratory Inductance Plethysmography (RIP) sensor that measures forces from expansion and contraction of the ribcage. The raw sensor data, sampled at 10Hz, is fed into the LSTM model in sliding windows. The LSTM architecture is chosen for its ability to learn long-term dependencies in sequential data. Dropout layers are used for regularization to prevent overfitting. The final output is a single value representing minute ventilation in liters per minute (l/m). The system is implemented as a web application that communicates with a PyTorch model running on an HTTP server via a RESTful API.
Datasets
The authors collected a dataset consisting of synchronized measurements from five male subjects (aged 26±1 years) performing sub-maximal and incremental exercise tests on a cycle ergometer. The data includes:
- Timestamp (s)
- Minute ventilation (l/m) - Ground truth from a Douglas Bag system and exercise spirometer.
- Ribcage movement (mV) - Measured by the Flow RIP sensor.
- Heart rate (bpm)
This data was made available in four columns for the deep learning model. The authors note that the test user was not part of the training data, as the data was collected in 2017.
Key Insights and Ideas
- Breathing as a direct proxy for effort: The paper highlights that breathing patterns are a more immediate and reactive indicator of physical exertion than heart rate, which can lag or drift.
- Non-invasive sensing potential: Using a RIP sensor for ribcage movement offers a less intrusive way to capture breathing dynamics compared to face masks and spirometers, which can be inconvenient for many activities.
- LSTM's suitability for breathing data: The LSTM architecture proves effective in capturing the temporal dependencies within breathing signals to predict minute ventilation.
- Real-time feedback loop: The system provides instant feedback on predicted minute ventilation, closely matching the user's perception of exercise intensity.
Key Results
The Deep Ventilation model was evaluated against ground truth data from an exercise spirometer. The authors report that in 90% of the breathing data, the model exhibited a maximum deviation of 20% from the ground truth. They also note that the LSTM model's output showed lower fluctuation, contributing to a perception of stability for the user.
Limitations and Future Work
The authors acknowledge several limitations and suggest future directions:
- Data diversity: The current model is trained on only five users. Future work involves collecting more data from individuals with diverse attributes (ethnicity, age, weight, height, fitness levels, climate experience) and with variations in sensor placement and tension to improve robustness.
- Sensor fusion: To improve accuracy, they suggest incorporating a second sensor to measure abdominal breathing.
- Personalization: The model could be fine-tuned for individual users (N-of-1 trials) for even greater accuracy.
- On-device deployment: The authors aim to embed the model directly into web pages using TensorFlow.js, enabling real-time prediction on client-side devices (smartphones) without relying on a server, provided the device has GPU acceleration.
- Advanced architectures: Exploring attention-based transformer models is suggested as a way to potentially improve prediction accuracy by being independent of the input sequence.
Summary updated