Detecting and measuring respiratory events in horses during exercise with a microphone: deep learning vs. standard signal processing.

Overview

• This study investigates methods to automatically detect horse breathing sounds during exercise using microphones, comparing deep learning techniques with traditional signal processing approaches.
• The aim is to accurately capture respiratory events and dynamic respiratory rates to better understand equine health and performance during training.

Background and Motivation

• Monitoring respiration parameters in horses during exercise is important for assessing health and training effects.
• Respiratory rate and patterns provide critical data for improving equine welfare and optimizing athletic performance.
• Traditional respiratory monitoring can be challenging during high-intensity activity due to noise and movement artifacts.
• The study explores using microphones to non-invasively record respiratory sounds and assess them automatically.

Research Objectives

• To develop automatic methods for detecting cyclic respiratory events (specifically exhalations) in microphone-recorded signals from exercising Standardbred trotters.
• To extract dynamic respiratory rates in real-time or near real-time from these recordings.
• To compare the performance of deep learning-based approaches with adapted standard signal processing techniques.
• To evaluate the methods on recordings from high-intensity exercise sessions and test generalization to lower-intensity, unlabelled data.

Methods

• Data Collection:
  • Microphone signals were captured from Standardbred horses during high-intensity trotting exercise.
  • The recordings contained noisy background and complex respiratory sounds due to exercise conditions.
• Automatic Detection Approaches:
  • Deep Learning Models:
    • Temporal Convolutional Networks (TCNs) designed to capture temporal dependencies and recognize exhalation sounds in noisy signals.
    • Long Short-Term Memory (LSTM) networks as another deep learning method to model temporal sequences.
  • Signal Processing Method:
    • An adapted, more traditional approach leveraging signal characteristics to identify respiratory events.
• Evaluation Metrics:
  • F1 Score (harmonic mean of precision and recall) to assess exhalation event detection accuracy.
  • Mean Absolute Error (MAE) and Confidence Intervals (CI) to measure the accuracy of estimated respiratory rates.
  • Limits Of Agreement (LOA) to examine the range of agreement between methods and ground truth.

Results

• Deep learning models effectively detected exhalation sounds despite noisy exercise environments.
• Median F1 score for exhalation detection was 0.94 with temporal convolutional networks, indicating high precision and recall.
• On lower intensity, unlabelled exercise data where exhalation sounds are less distinct, models still showed promising detection ability.
• Comparison of models:
  • Temporal Convolutional Networks outperformed LSTMs with median F1 = 0.94 vs. 0.90.
  • MAE ± CI for respiratory rate estimation: TCN at 1.44 ± 1.04 bpm; LSTM at 3.11 ± 1.58 bpm; signal processing at 2.36 ± 1.11 bpm.
  • LOA showed TCN had tighter agreement (0.63 ± 7.06 bpm), indicating less variability in estimates.

Conclusions and Future Directions

• This is the first study to successfully automate detection of equine respiratory sounds and calculate dynamic respiratory rates during exercise using microphone data.
• Deep learning, especially temporal convolutional networks, proved superior to standard signal processing techniques.
• Future work includes:
  • Validating these models on data recorded during lower intensity exercise, where sounds are more subtle.
  • Exploring different microphone placements on horses to optimize signal quality and detection accuracy.
  • Developing practical solutions for continuous, regular respiratory monitoring to support equine health and performance management.