Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evolutionary Optimization of 1D-CNN for Non-contact Respiration Pattern Classification

Md Zobaer Islam, Sabit Ekin, John F. O'Hara, Gary Yen

TL;DR

This study tackles non-contact respiration pattern classification using incoherent infrared light by deploying a 1D-CNN whose architecture is optimized via a genetic algorithm. To manage computational demands, the authors leverage transfer learning from a pre-trained, augmented 1D-CNN and validate performance on eight breathing classes, including a faulty data category, achieving about 87% test accuracy. The approach combines a carefully designed GA with a four-gene chromosome to tune the untrained portion of the network, while leveraging a pre-trained backbone to accelerate training. The work demonstrates a practical pathway for accurate, efficient respiratory anomaly detection suitable for home and clinical monitoring, and highlights avenues for further improvement through GPU acceleration and multi-objective optimization.

Abstract

In this study, we present a deep learning-based approach for time-series respiration data classification. The dataset contains regular breathing patterns as well as various forms of abnormal breathing, obtained through non-contact incoherent light-wave sensing (LWS) technology. Given the one-dimensional (1D) nature of the data, we employed a 1D convolutional neural network (1D-CNN) for classification purposes. Genetic algorithm was employed to optimize the 1D-CNN architecture to maximize classification accuracy. Addressing the computational complexity associated with training the 1D-CNN across multiple generations, we implemented transfer learning from a pre-trained model. This approach significantly reduced the computational time required for training, thereby enhancing the efficiency of the optimization process. This study contributes valuable insights into the potential applications of deep learning methodologies for enhancing respiratory anomaly detection through precise and efficient respiration classification.

Evolutionary Optimization of 1D-CNN for Non-contact Respiration Pattern Classification

TL;DR

This study tackles non-contact respiration pattern classification using incoherent infrared light by deploying a 1D-CNN whose architecture is optimized via a genetic algorithm. To manage computational demands, the authors leverage transfer learning from a pre-trained, augmented 1D-CNN and validate performance on eight breathing classes, including a faulty data category, achieving about 87% test accuracy. The approach combines a carefully designed GA with a four-gene chromosome to tune the untrained portion of the network, while leveraging a pre-trained backbone to accelerate training. The work demonstrates a practical pathway for accurate, efficient respiratory anomaly detection suitable for home and clinical monitoring, and highlights avenues for further improvement through GPU acceleration and multi-objective optimization.

Abstract

In this study, we present a deep learning-based approach for time-series respiration data classification. The dataset contains regular breathing patterns as well as various forms of abnormal breathing, obtained through non-contact incoherent light-wave sensing (LWS) technology. Given the one-dimensional (1D) nature of the data, we employed a 1D convolutional neural network (1D-CNN) for classification purposes. Genetic algorithm was employed to optimize the 1D-CNN architecture to maximize classification accuracy. Addressing the computational complexity associated with training the 1D-CNN across multiple generations, we implemented transfer learning from a pre-trained model. This approach significantly reduced the computational time required for training, thereby enhancing the efficiency of the optimization process. This study contributes valuable insights into the potential applications of deep learning methodologies for enhancing respiratory anomaly detection through precise and efficient respiration classification.
Paper Structure (15 sections, 8 figures, 1 table)

This paper contains 15 sections, 8 figures, 1 table.

Table of Contents

  1. Introduction
  2. System Design and Implementation
  3. Experimental setup
  4. Data collection
  5. 1D-CNN Optimization
  6. Transfer Learning Application
  7. Optimization Using Genetic Algorithm
  8. Build Chromosome and Initial Population:
  9. Get Fitness Values:
  10. Perform Crossover:
  11. Perform Mutation:
  12. Perform Selection:
  13. Results and Discussion
  14. Conclusion and Future Directions
  15. Acknowledgements

Figures (8)

  • Figure 1: The hardware setup of the overall system used for data collection.
  • Figure 2: Time domain representation of sample data from each class (the distance between the source-photodetector and the robot was 1 m).
  • Figure 3: Pre-trained, trimmed and extended models.
  • Figure 4: Transfer learning framework for 1-dimensional respiration data classification.
  • Figure 5: Genetic algorithm workflow.
  • ...and 3 more figures