Morphological autoencoders for apnea detection in respiratory gating radiotherapy

https://doi.org/10.1016/j.cmpb.2020.105675Get rights and content

Highlights

  • Recognition of apnea and regular breathing patterns in respiratory gating training and a reference dataset.

  • A morphological analysis was followed through autoencoders, where supervised learning was used to evaluate input-output correlation of the autoencoder.

  • This morphological approach led to promising results in both datasets, hence avoiding feature engineering and enhancing transferability to other time series analysis.

Abstract

Background and Objective: Respiratory gating training is a common technique to increase patient proprioception, with the goal of (e.g.) minimizing the effects of organ motion during radiotherapy. In this work, we devise a system based on autoencoders for classification of regular, apnea and unconstrained breathing patterns (i.e. multiclass). Methods: Our approach is based on morphological analysis of the respiratory signals, using an autoencoder trained on regular breathing. The correlation between the input and output of the autoencoder is used to train and test several classifiers in order to select the best. Our approach is evaluated in a novel real-world respiratory gating biofeedback training dataset and on the Apnea-ECG reference dataset. Results: Accuracies of 95 ± 3.5% and 87 ± 6.6% were obtained for two different datasets, in the classification of breathing and apnea. These results suggest the viability of a generalised model to characterise the breathing patterns under study. Conclusions: Using autoencoders to learn respiratory gating training patterns allows a data-driven approach to feature extraction, by focusing only on the signal’s morphology. The proposed system is prone to be used in real-time and could potentially be transferred to other domains.

Introduction

One compelling use case of respiration-related signals is patient training for respiration-gated thoracic radiotherapy. Several radiation treatments can greatly benefit from breathing control during radiotherapy [1], [2]. By being aware and capable of repeatedly reproduce respiration cycles that are adequate to the treatment, patients can control their breathing in a way that minimizes the effects of organ motion on the process [3]. The internal motion of the upper abdomen and thoracic organs, which is provoked by respiration motion, is the cause of blurring in imaging, uneven administration of radiation dosimetry and other situations conflicting with the proper therapeutic procedure. Besides the advantage of modulating therapy to free-breathing, some types of breast and lung cancer highly benefit from breath-hold techniques. Previous studies reported the impact of deep inspiration breath-hold (DIBH) in reducing cardiac irradiation during radiotherapy, thereby reducing the risk of cardiovascular morbidity after the treatment [1], [4], [5].

A breathing cycle is mainly characterized by two phases: inhalation and exhalation. These two phases are common to all individuals, however, they are affected by our physicality, physiology and also our surrounding environment (e.g. poor air quality and pollution, which can lead to respiratory problems). Nevertheless, all these factors can contribute to a significant inter- and intra-subject variability. Therefore, it is of utmost importance to adjust thoracic radiotherapy to respiration gating [6]. This adjustment implies the patient to be aware of her/his respiratory signal and to perform a conscious control during the therapy. Respiratory Gating Biofeedback Training (RGBT) is used generally for such purpose.

An important part of RGBT is training breath-hold, i.e., consciously provoked apnea. In Fig. 1, we present three distinct breathing patterns of two subjects, each illustrating a free, a regulated (in green) and a breath-hold breathing pattern (in blue). While in free breathing (Free) different amplitudes and frequencies are possible, in regulated breathing (Sinus) the breathing pattern is more consistent. Finally, apnea starts with deep inspiration and is followed by the breath-hold. The top patient was able to sustain his breath for the entire apnea sample, however, the bottom one presents oscillations during the breath-hold time, corresponding to the recuperation of breath. The presence of these fluctuations in apnea samples increases their morphological variability, compared to the regulated breathing (Sinus) samples.

The discrimination between apnea and normal breathing is not unusual in research, specially in sleep apnea. Sleep apnea is assessed through polysomnography [7], where several physiological signals are measured to properly analyze different apnea types. In sleep apnea, many studies successfully distinguish between apnea and normal breathing [8], [9], [10], however most techniques explored to date rely on multimodal analysis, where respiratory-based signals do not play a major role. Mendonça and colleagues [11] thoroughly reviewed previous work on obstructive sleep apnea (OSA), and concluded that the most valuable single source sensor is ECG, and respiratory-based signals are only the third best choice. Despite the preference for ECG in the characterization of sleep apnea, our work will use only nasal airflow, since our purpose is to transition to the RGBT case-study.

In this work we aim to classify apneas from regular breathing samples in a non-fiducial way, targeting a real world implementation for RGBT. Contrasting with previous literature on apnea detection, we will distinguish apneas in healthy subjects, which are purposely holding their breath during RGBT. Radiation oncology would benefit not only from patient training RGBT, but also from models that could predict who could endure therapy under breath-hold techniques [12]. Although there is a significant difference between breathing patterns (e.g. as illustrated by the samples in Fig. 1), we have no clear insights on what features will perform best.

Instead of trying to find the best features that enable the distinction of these two classes, we recur to a morphological approach, grounded by the ability of an autoencoder to reproduce normal breathing. Autoencoders (AE) is a type of artificial neural networks composed by encoding layers that compress the input into a reduced representation, and decoding layers that reconstruct the output based on the encoded representation of the input [13]. This methodology can find the best-encoded representation of the signal so that the output replicates the input as faithfully as possible. Recently, studies on electrophysiological signals are using AE as a method of feature extraction [14], to perform denoising [15] or even to detect distortions [16]. Other studies have also developed anomaly detectors based on variational encoders [17] or robust deep AE [18]. The application of AE to avoid feature engineering has been used in research, such as the work of Yang and colleagues [19], where the authors used a recurrent variational AE to model respiration and Kullback Divergence to compare the output with the input, acting as an apnea detector. In our work, we go further on by combining morphological AE with supervised classification learning.

The remainder of this paper is divided in the following sections: In Section 2, the theory of the algorithms applied for supervised learning is briefly described. Then, in Section 3, the used datasets are explained, followed by Section 4, where the signal processing steps are described, as well as the system’s architecture for classification. The several experiments undertaken are presented in Section 5, alongside with their best results and a brief discussion. In the end of this paper, Section 6 summarizes the overall conclusions inferred from the experiments and the prospects for future work.

Section snippets

Background

Autoencoder is an unsupervised algorithmic structure based upon neural networks, which are combination of nodes stacked in layers. The relations between input and outputs of each node are non-linear and defined by an activation function. In order to train the network, the backpropagation method is used. This method starts with a forward pass, where the input is given to the model and, in each layer, bias is added and weights are multiplied. The output of the network is then evaluated with the

BrainAnswer RGBT

Due to the lack of RGBT datasets, BrainAnswer1 created a new corpus comprised of 45 subjects, herein designated as BA-RGBT, used in our work. The range of users comprised 33 female and 12 male with an average age of 22 ± 1.3 years old. All data was collected in accordance with the institutional ethical regulations, the General Data Protection Regulation (GDPR), and with the voluntary agreement of the participants enrolled in the RGBT process.

The data acquisition consisted

Architecture

The methodology developed in this work follows the workflow depicted in Fig. 3. Raw respiration data is filtered and normalised (preprocessing), then it is used as input to the AE, which returns an output. The output data and input data are compared by correlation. The correlation vector is then given to a classifier, which will predict if a signal belongs to the apnea or the normal class. This process follows a cross-validation with several folds, where the user only belongs to the training

Classifier evaluation and correlation vector size

The first part of our architecture (AE training) remained just like it was described in Section 4. In this section we present results for different classifiers and different number of correlation points (correlation size). To come to a decision on these two questions, we evaluate the results of both datasets, where the results for BA-RGBT are in Fig. 5 and the results for Apnea-ECG are in Fig. 6. The high results with all classifiers reveal the robustness of our approach in creating linearly

Conclusion

In this work we studied the ability of distinguishing between apnea and normal respiration patterns using a non-fiducial approach in a RGBT setting. AE were used to model each state, just by recurring to the morphological differences between the signals. Although automatic apnea detection has been extensively covered in sleep studies, the state-of-the-art is lacking comparable work on RGBT. For this reason our work explored a novel dataset (BA-RGBT), albeit also being applied to a reference

Declaration of Competing Interest

The authors declare no conflict of interest.

Acknowledgements

We would like to acknowledge Professor Telmo Pereira, from the Physiology Department of the Polytechnic Institute, Coimbra Health School, for his insight in the physiological explanations of the results.

This work has been partially funded by the Xinhua Net Future Media Convergence Institute under project S-0003-LX-18, by the Ministry of Economy and Competitiveness of the Spanish Government, co-founded by the ERDF (PhysComp project, TIN2017-85409-P), and by FCT - Fundação para a Ciência e

References (29)

  • C. Avcı et al.

    Sleep apnea classification based on respiration signals by using ensemble methods

    Biomed Mater Eng

    (2015)
  • F. Mendonca et al.

    A review of obstructive sleep apnea detection approaches

    IEEE J Biomed Health Inform

    (2018)
  • M. Feng et al.

    Machine learning in radiation oncology: opportunities, requirements, and needs

    Front Oncol

    (2018)
  • P. Baldi

    Autoencoders, unsupervised learning, and deep architectures

    Proceedings of ICML workshop on unsupervised and transfer learning

    (2012)
  • Cited by (2)

    • A semi-supervised autoencoder framework for joint generation and classification of breathing

      2021, Computer Methods and Programs in Biomedicine
      Citation Excerpt :

      Likewise, very few computer-aided diagnostic tools have been presented for physical breathing signals. Abreu et al. [22] present an autoencoder framework that discriminates between apnea and regular breathing, focusing on gating radiotherapy treatments. We investigate whether it is possible to combine classification and generation of breathing signals within a single model.

    View full text