DABiG: Breath pattern classification using the hybrid deep learning with optimal feature selection

Abstract

Background

A person's breathing pattern can be a reflection of their emotional and physical well-being because it shows the frequency, intensity, and rhythm of their breathing.

Objective

This research article presents a comprehensive approach to breathe pattern classification utilizing gyroscope and accelerometer readings obtained from individuals using two distinct sensors. The study encompasses the acquisition of six diverse breathing patterns, with a focus on data pre-processing through Min-Max normalization.

Methods

To select essential features from the normalized data, an innovative optimization algorithm, Adaptive Chimp Optimization (AdCO), is introduced. AdCO integrates an adaptive weighting strategy into the conventional Chimp optimization algorithm, enhancing convergence rates and enabling global optimal feature selection. Furthermore, the article introduces the application of the selected features in breath pattern classification using a hybrid deep learning mechanism, DABiG. DABiG leverages the Bidirectional Gated Recurrent Unit (BiGRU), a neural network architecture capable of processing sequential data bi-directionally.

Results

Spatial and temporal attention mechanisms are incorporated into DABiG to enhance its ability to focus on relevant spatial regions and time steps within the breath pattern data.

Conclusion

Spatial attention assigns weights to spatial regions, while temporal attention assigns weights to time steps, improving feature extraction and classification accuracy.

Keywords

military breath pattern analysis hybrid deep learning adaptive chimp optimization double attention min-max normalization

1 Introduction

Breathing is one of the most essential physiological functions for all living things. An irregular breathing pattern may be the first sign that anything is amiss in any psychological, physiological, or mechanical problem.¹ Therefore, respiration needs to be considered in all physical therapy evaluations. Breathing techniques have changed dramatically with the ages, philosophies, and civilisations. Since the turn of the century, Western medicine has acknowledged the significance of breathing for health, and more recent research has carefully reviewed the role of the breath in both wellness and illness.^2,3 In favour of the general role that physical therapy may play in improving breathing patterns through symptom reduction or elimination and wellbeing promotion.⁴ Thus far, there hasn't been much discussion of aberrant breathing patterns or breathing re-education in the physiotherapy literature.⁵

A population of Army Basic Combat Training (BCT) troops showed an increased prevalence of febrile Airway, Respiratory Rate, and Inspiratory Effort (ARI) when living in newer, more energy-efficient buildings.⁶ Other studies have generally looked at the interior settings of office buildings, with varied degrees of effectiveness. According to recent studies, the environment has a critical role in the spread of respiratory illnesses within a community of military trainees and social isolation.⁷ While only a tiny fraction of ARIs in the US progress to significant consequences, the majority cause acute morbidity, which results in missed workdays. The typical classifications for ARIs include upper and lower respiratory infections.⁸ Most upper respiratory infections are caused by viruses that resemble mild illnesses. Particularly in elderly adults, immunocompromised individuals, and young children in developing nations, lower respiratory infections such as pneumonia can have far more serious outcomes.⁹ According to the World Health Organisation, lower respiratory infections rank third among all causes of death globally each year, accounting for 4.18 million deaths or 7.1% of all fatalities. Young children, the elderly, and people with compromised immune systems are more likely to contract these illnesses in developing nations. Respiratory infections account for 25–30% of infectious disease hospitalisations in U.S. military populations and are the primary source of illness that does not require hospitalisation.¹⁰

Using data on chest displacement, several tests have been conducted to precisely calculate the respiration rate. Patients with respiratory diseases cannot be used to represent abnormal breathing patterns using respiration rate.¹¹ Machine learning tools that recognise breathing patterns are essential for diagnosing respiratory issues. The incorporation of machine learning will be extremely beneficial for automating a more sophisticated and intelligent system. Thus, efforts were made to integrate artificial intelligence with radar imaging to develop a system that can recognise and classify patterns of breathing abnormalities.¹² Mel-Frequency Cepstral Coefficients (MFCC) feature extraction, XGBoost classifier, and FMCW radar are used to recognise non-contact breathing patterns in an interior setting. Furthermore, XGBoost recognises, reduces, and captures large amounts of breathing with the help of MFCC extraction of features.¹³

This study thus introduces a deep learning framework for the categorisation of breathing patterns. The optimal feature selection method is used in the suggested AdCO + DABiG-based breath pattern classification to select the best features and increase the breath pattern's classification accuracy. Furthermore, the BiGRU's inclusion of a dual attention mechanism helps to capture temporal and spatial characteristics as well as long-term dependent features without experiencing vanishing gradient problems. The major contributions of the research are:

AdCO Algorithm for Feature Selection: To choose the optimal best features using the Adaptive Chimp optimization (AdCO) algorithm, wherein the adaptive weighting strategy is incorporated into the conventional Chimp optimization algorithm for enhancing the convergence rate in choosing the optimal best solution.

Proposed DABiG for breath pattern Classification: To classify the breath patterns using the hybrid deep learning technique named DenseGRU is designed by hybridizing the DenseNet-121 and Gated Recurrent Unit (GRU) for enhancing the classification accuracy.

The organization of the research is: The existing breath pattern analysis methods along with the problem statement is presented in Section 2 and the proposed methodology is detailed in Section 3. Finally, Section 4 presents the experimental outcome and Section 5 concludes the research.

2 Related works

Some of the existing breath pattern analysis approaches are reviewed here. Respiration Variability Spectrogram (RVS) Sequences-based pattern analysis was introduced by¹⁴ for recognizing psychological stress levels through breathing patterns. The developed model strengths encompassed automated feature extraction, enabling robust learning from limited data, and an innate ability to capture the nuanced dynamics of breathing using the deep learning model.Here, the inclusion of an augmentation strategy effectively expanded the model's ability to learn from limited datasets that promoted better generalization to a diverse spectrum of breathing patterns. The signal transformation enabled the model to encapsulate the intricate and dynamic nature of breathing patterns more comprehensively. A machine learning-based breath pattern analysis was designed by¹⁵ using the Respiratory Rate, Temperature, and Heart Rate. The primary objective was to predict the onset of moderate-to-severe respiratory failure using readily accessible vital signs. Here, the adoption of a random forest approach for modeling offered robustness and the ability to capture complex relationships within the data. While the introduced method offered advantages such as a holistic understanding of the medical problem and the use of readily accessible data, it also came with challenges related to data quality, feature selection, and the balance between prediction accuracy and timeliness.

The introduction of an innovative wearable system based on Arduino technology was designed by¹⁶ for breath pattern monitoring, particularly for sleep apnea diagnosis and continuous surveillance, held significant promise. The system autonomously identifies and categorizes various respiratory patterns like instances of breathlessness, encompassing exhalation and inhalation. Arduino's flexibility and adaptability in capturing patient information made it a suitable choice for this specialized medical application. Still, the computational power and data processing capabilities have impacted the system's overall performance. XGBoost-based respiratory was implemented by¹⁷ for respiratory disorder detection byclassifying breathing patterns. In this, the Mel-frequency cepstral coefficient (MFCC) for extracting relevant features from the breathing signal. The potential applications in respiratory disorder detection held promise, but they also entailed challenges related to computational resources and real-world implementation.

2.1 Motivation

Using image processing techniques based on artificial intelligence and machine learning, breathing disorder diagnosis devices have been created in response to current advances in wearable sensors and devices for monitoring respiration. To recognize the condition using ECG readings, for instance, deep learning algorithms have been applied. Medical practitioners have been able to establish criteria for sleep apnea diagnosis by the real-time estimate of visual equipment patterns from images of a patient's stomach and chest. However, present devices have high pricing, difficult detecting procedures, poor accuracy, and difficulty with usage. As a result, it is critical to fulfill the demand for low-cost, automated devices that can deliver simple, measurements in real time while guaranteeing high rates of specificity, sensitivity, and accuracy.Thus, a deep learning-based military breathing pattern analysis is introduced to enhance the classification accuracy with minimal computation complexity.

3 Proposed methodology

The primary objective of military breathing pattern analysis is to gather information about a soldier's respiratory behavior to ensure their health, performance, and readiness. Monitoring and analyzing breathing patterns in the military is crucial for several reasons, including evaluating the physical and psychological well-being of soldiers, identifying signs of stress or respiratory distress, and optimizing performance in high-stress situations. Thus, the military pattern analysis is introduced in this research for enhancing the classification accuracy with minimal computation complexity. Initially, the breath patterns are gathered from individuals using the wearable sensors. Then, the data normalization is devised to make the data in standard form to make the computations simpler. Then, the optimal best features are acquired using the proposed adaptive Chimp optimization (AdCO) algorithm, which is designed by incorporating the adaptive weighting strategy within the conventional Chimp optimization algorithm for choosing the most informative features that enhance the classification accuracy. Finally, the breathing patterns are classified using the hybrid deep learning method named Dual attention Bidirectional Gated Recurrent Unit (DABiG). Here, the dual attention mechanism is integrated into the conventional BiGRU to enhance classification accuracy. The workflow diagram is presented in Figure 1.

Figure 1.

Workflow of AdCO + DABiG-based breath pattern classification.

3.1 Data acquisition

The data is gathered using the gyroscope and accelerometer readings are acquired from the individuals using two various sensors. Six various breathing patterns are collected from the individuals and are utilized for the classification task.¹⁸ The six various breath patterns areNormal breathing, Apnea, coughing, Muller, Sighing, and Yawning. These breath patterns were captured based on the physician's instruction for 5 min and 5 repetitions.

3.2 Data normalization

Data standardization is essential for making the computations simpler, which is accomplished through the data normalization technique. Min-Max normalization is utilized for performing the data normalization and is stated as:

\begin{matrix} S t_{B} = \frac{B - B_{l}}{B_{h} - B_{l}} \end{matrix}

(1)

The data used for processing is shown as B, the normalized data is shown as

S t_{B}

, the lowest value is shown as

B_{l}

, and the highest value is shown as

B_{h}

. Then, the essential features are chosen from the normalized data using the proposed optimization algorithm.

3.3 Optimal feature selection

For the selection of optimal best features from the normalized data, the proposed Adaptive Chimp Optimization (AdCO) is utilized. The AdCO is designed by integrating the adaptive weighting strategy within the conventional Chimp optimization algorithm for enhancing the convergence rate. The chimp is one among the Great apes like chimpanzees, bonobos, gorillas, and orangutans that are native to Africa. Brain-to-Body Ratio (BBR) level is higher for chimpanzees and dolphins; still, BBR is relatively high for chimpanzees. BBR is a measure of brain size relative to body size. The mammal with a higher level of BBR has smart behavior; hence, Chimp is considered for solving the optimization issue by considering its hunting behavior. Still, the Chimp optimization algorithm has the challenge of local optimal trapping, which is solved by incorporating the adaptive weighting strategy with the conventional algorithm. Hence, the proposed AdCO algorithm assistsin providing the global best optimal solution in choosing the essential features that enhance the classification accuracy. Besides, the balanced exploration and exploitation rate of the algorithm efficiently solves the optimization issue.

3.3.1 Mathematical modeling

While designing the optimization algorithm, four various categories of chimps are considered: attacker, chaser, barrier, and driver.

Attacker: Attackers are skilled in predicting the prey's potential escape routes and take actions to block or influence those routes. Attackers use their ability to anticipate the prey's movements and strategically position themselves to cut off escape paths. Thus, the escaping probability of the target gets reduced and the capturing is made simpler.

Barrier: Barriers position themselves in a tree to create an obstruction or barrier that blocks the path of the prey by creating a physical obstacle.

Chaser: The Chaser is a chimp that interacts with a target and the one who follows or observes the target to capture it.

Driver: The Driver is a role responsible for coordinating and guiding the optimization process. It plays a role similar to a leader or a coordinator within the chimp group. Drivers might be responsible for selecting the best solutions or directing the search towards promising areas of the search space.

Initialization: The population of the chimps (search agents) and the maximal number of iterations are initialized.

Fitness Estimation: Estimating fitness allows algorithms to operate more efficiently by avoiding unnecessary evaluations and speeding up the convergence of optimization. It is estimated as:

\begin{matrix} F i t = \frac{A c c u r a c y + S e n s i t i v i t y + S p e c i f i c i t y}{3} \end{matrix}

(2)

where, the fitness is denoted as

F i t

Target Chasing: the target is captured in both the exploitation and exploration phases using various attacking strategies. The chasing and driving behavior of the search agent is mathematically designed as:

\begin{matrix} a = | e C_{t a r} (g) - f C_{c h i m p} (g) | \end{matrix}

(3)

\begin{matrix} C_{C h i m p} (g + 1) = C_{t a r} (g) - r . a \end{matrix}

(4)

where, the present iteration is defined as g, the location of the target is denoted as

C_{t a r}

, and the location of the search agent is denoted as

C_{c h i m p}

. Then, the coefficient vectors are denoted as

r, f, a n d e

and is defined as:

\begin{matrix} r = 2 k m_{1} - k \end{matrix}

(5)

\begin{matrix} e = 2 m_{2} \end{matrix}

(6)

\begin{matrix} f = c h a o t i c_v a l u e \end{matrix}

(7)

where, the value of k decreases from 2.5 to 0 in both the randomization and local search phases. And the random numbers has the range of

[0, 1]

is denoted as

m_{1} a n d m_{2}

Exploitation Phase: While exploring the search space, except for the attacker search agent all the remaining three types of search agents participate in capturing the target. In contrast, the attacker search agent participates in attacking the target, which is surrounded by the chaser, driver, and barrier. The solution accomplished by the search agents in this phase is designed as:

\begin{matrix} C (g + 1) = \frac{C_{M} + C_{N} + C_{O} + C_{P}}{4} \end{matrix}

(8)

where, the solution of carrier is defined as

P_{O}

, the solution of the driver is noted as

C_{P}

, the solution of the barrier is referred as

C_{N}

, the solution of the attacker is denoted as

C_{M}

, and the solution updation is represented as

C (g + 1)

. Here, the position updated by the individual chimp is expressed as,

\begin{matrix} C_{M} = C_{1} - r_{1} (a_{M}) \end{matrix}

(9)

\begin{matrix} C_{N} = C_{2} - r_{2} (a_{N}) \end{matrix}

(10)

\begin{matrix} C_{O} = C_{3} - r_{3} (a_{O}) \end{matrix}

(11)

\begin{matrix} C_{P} = C_{4} - r_{4} (a_{P}) \end{matrix}

(12)

where,

a_{M}

a_{N}

a_{O}

, and

a_{P}

relates to the distances between the target and the attacker, barrier, carrier, and driver chimps. The coefficient

r_{1}, r_{2}, r_{3}, a n d r_{4}

varies between

[0, 1]

that makes the contenders to seize the prey

C_{1}, C_{2}, C_{3}, a n d C_{4}

and is used to describe the attacker's, the barrier's, the carrier's, and the driver's best solutions. Here, the adaptive weighting strategy is incorporated into the chimp algorithm for enhancing the convergence rate without trapping the solution at local optima. The equation for the adaptive weighting strategy is stated as:

\begin{matrix} L = (1 - g / g_{max})^{1 - \tan (π \times (r a n d - 0.5) \times b / g_{max})} \end{matrix}

(13)

where, the current ad higher iteration is denoted as

g a n d g_{max}

, L and refers the adaptive weight. The control factor is defined as b, when the position of the chimp is not updated then b is added with the position. Else, b is divided by 2 while updating the position of the Chimp search agent. Thus, the role of the control factor is to fluctuate the chimp while trapping at the local optimal solution. Then, the adaptive weight included chimp for enhancing the exploration capability is formulated as,

\begin{matrix} C (g + 1)_{A d C O} = L \times C (g + 1)_{C h i m p} \end{matrix}

(14)

Exploration Phase: The exploration process involves diverging the behavior of search agents to seek the optimal solution (target) and then converging to attack it. For mathematical modeling, the divergence value is represented as r with a random value that is in the range of 1 and −1. The higher value makes the search agents explore more areas in the search space and the lower value makes the algorithm to converge. Besides, the factor e also affects the exploration capability of the AdCO algorithm. Here, the vector e has a value range between 0 and 2 that assists the algorithm in providing random weights for the target's location in the determination of distance based on Equation (5). The vector e enhances the stochastic behavior of the algorithm thus, it assistsin preventing the local optimal trapping during the final phase of iteration.

Chaotic-based solution updation: Chaotic behavior refers to behavior that appears random and unpredictable but follows deterministic rules. Thus, the chaotic behavior injects randomness, which can expedite convergence by guiding the search toward promising areas of the high-dimensional space. The formulation for the chaotic behavior-based solution updation is expressed as:

\begin{aligned} C_{c h i m p} (g + 1) = {\begin{cases} C_{t a r} (g) - r e & i f ε < 0.5 \\ c h a o t i c_v a l u e & i f ε > 0.5 \end{cases} \end{aligned}

(15)

where, the random number

ε

has the value range of 0 and 1.

Re-evaluation of fitness: The fitness is re-evaluated to check the feasibility of the solution obtained by the algorithm.

Termination: The acquisition of a better solution for choosing the optimal best features or the attainment of maximal iteration stops the termination of the algorithm.

3.4 Breath pattern classification using DABiG

Using the selected features, the breath pattern classification is devised using the hybrid deep learning mechanism. The BiGRU is a type of recurrent neural network (RNN) that processes sequential data bi-directionally. It consists of two GRU layers, one processing the data in the forward direction and the other in the backward direction. GRU stands for Gated Recurrent Unit, which is a variant of the more traditional long short-term memory (LSTM) cell. GRUs are known for their ability to capture long-range dependencies in sequential data while being computationally efficient. Each GRU cell maintains a hidden state that encodes information from previous time steps and uses this information to update its current state. Besides, spatial attention focuses on the information from different spatial locations (e.g., sensors on the chest, abdomen, and back) where data is collected. The spatial attention mechanism computes attention weights for each spatial location, indicating which sensors are most relevant for the classification task. These weights are often learned during training. The spatial attention weights are used to combine the spatial information into a single, weighted representation. Temporal attention is applied to the sequential data to capture the dynamic nature of breathing patterns over time. For each time step, the temporal attention mechanism computes a weight or attention score, indicating the importance of that time step's information relative to the others in the sequence. The temporal attention weights create a weighted representation of the sequence, emphasizing the most relevant time steps for classification. The architecture of the DABiG for breath pattern analysis is depicted in Figure 2.

Figure 2.

Architecture of the DABiG for breath pattern analysis.

Initially, the candidate state ${\tilde{p}}_{T}$ is estimated by considering the outcome of the reset gate $u_{T}$ . The estimation of the candidate state ${\tilde{p}}_{T}$ is written as:

\begin{matrix} {\tilde{p}}_{T} = \tanh (Q_{p} [u_{T} \times p_{T - 1}, D_{T}]) \end{matrix}

(16)

Here,

u_{T}

refers to the reset gate and is stated as:

\begin{matrix} u_{T} = S i g (Q_{u} [p_{T - 1}, D_{T τ}]) \end{matrix}

(17)

By using the reset gate, inappropriate details are removed and then, the update gate

l_{T}

controls the operation as:

\begin{matrix} l_{T} = S i g (Q_{l} [p_{T - 1}, D_{T}]) \end{matrix}

(18)

Finally, the outcome of the hidden state

p_{T}

is expressed as:

\begin{matrix} p_{T} = (1 - l_{T}) \times p_{T - 1} + l_{T} \times {\tilde{p}}_{T} \end{matrix}

(19)

where, the weight is defined as Q,

D_{T}

refers to the input and

S i g

refers to the sigmoid function. GRU networks are known for their ability to capture long-term dependencies in sequential data. However, they can sometimes struggle with certain patterns that require understanding both past and future contexts. BiGRU mitigates this issue by utilizing information from both directions, making it better suited for capturing long-range dependencies in breath patterns. The outcome of the backward and forward hidden state outcome is outlined as:

\begin{matrix} \vec{p_{T}} = G R U (D_{T}, {\vec{p}}_{T - 1}) \end{matrix}

(20)

\begin{matrix} \overset{\leftarrow}{p_{T}} = G R U (D_{T}, {\overset{\leftarrow}{p}}_{T - 1}) \end{matrix}

(21)

\begin{matrix} p_{T} = A_{T} p_{T}^{\to} + B_{T} p_{T}^{\leftarrow} + f_{T} \end{matrix}

(22)

where, weights concerning the forward and backward hidden layer is indicated as

A_{T} a n d B_{T}

respectively. The state of the hidden layer in both the forward and backward processing is notated as

p_{T}^{\to} a n d p_{T}^{\leftarrow}

respectively. At the time T, the state of the hidden layer is indicated as

f_{T}

. Then, the weighting mechanism is devised using the double attention mechanism named temporal and spatial attention for enhancing the breath pattern classification accuracy.

Temporal Attention Module: A temporal attention mechanism assigns weights to each time step within a breath pattern sequence. These weights represent the relative importance of each part of the sequence in making the final classification decision. After calculating the attention weights, the model computes a weighted sum of the input sequence, where each time step is multiplied by its corresponding attention weight. This weighted sum effectively highlights the most relevant characteristics of the breath pattern sequence. It is expressed as:

\begin{matrix} G_{H}^{'} = s o f t max (W_{G} X_{H} + Z_{H}) \end{matrix}

(23)

\begin{matrix} Y_{H}^{″} = G_{H}^{'} Θ X_{H} = (G_{1, H} X_{1, H}, G_{2, H} X_{2, H}, \dots \dots . G_{T, H} X_{T, H}) \end{matrix}

(24)

where,

G_{H}^{'} = (G_{1, H}, G_{2, H}, \dots . . G_{T, H})

refers to the temporal attention layer's weights,

W_{G}

and

Z_{H}

denotes the weight and bias and

Y_{H}^{″}

is the output of the temporal attention and the dot product is indicated as

Θ

Spatial Attention Module: Spatial attention mechanisms assign importance weights to different spatial features within the input data. These weights represent the relative significance of each feature for the classification task. Features with higher weights are considered more important in making the final classification decision. After computing the attention weights, the model typically performs a weighted sum or a weighted aggregation of the input data based on these weights. This aggregation process highlights the most relevant spatial features for the breath pattern classification. It is expressed as:

\begin{matrix} R_{H}^{'} = s o f t max (W_{R} X_{H} + Z_{H}) \end{matrix}

(25)

\begin{matrix} Y_{H}^{″} = R_{H}^{'} Θ X_{H} = (R_{1, H} X_{1, H}, R_{2, H} X_{2, H}, \dots \dots . R_{T, H} X_{T, H}) \end{matrix}

(26)

where,

R_{H}^{'} = (R_{1, H}, R_{2, H}, \dots . . R_{T, H})

refers to the spatial attention layer's weights,

W_{R}

and

Z_{H}

denotes the weight and bias and

Y_{H}^{″}

is the output of the spatial attention.

Softmax Classification: Finally, the softmax classification is devised for classifying the five various breath patterns.

4 Result and discussion

The implementation is employed in the PYTHON programming language and the assessment is made based on various assessment measures like accuracy, precision, recall, specificity, and F-score. In this, the conventional breath pattern analysis methods like those Proposed without FS, 1D-CNN,¹⁸ Deep Breath,¹⁶ and Ensemble learning¹⁷ are evaluated to depict the superiority of the proposed method. For the analysis, the maximal iteration of 100, the population size of 50, epochs of 100, and softmax activation are considered, which is chosen based on a trial and error approach.

4.1 Analysis by varying training data

The analysis by varying the training percentage is presented in Figure 3 for various assessment measures and the detailed analysis is depicted in Table 1. Let the accuracy evaluated by the AdCO + DABiG is 96.05% with 80% of training data, which is 95.94% using the DABiG (Proposed without AdCO-based feature selection). Thus, the role of feature selection is crucial in the proposed breath pattern classification model.

Figure 3.

Analysis by varying training data: (a) accuracy, (b) F-Score, (c) precision, (d) recall and (e) specificity.

Table 1.

Analysis by varying training data.

Training percentage/methods	Ensemble learning	Deep breath	1D-CNN	Proposed without FS	Proposed
Recall
50	88.24	91.65	92.065	92.87	93.87
60	89.72	92.45	93.045	93.3	94.63
70	90.4	94.34	95.034	94.87	95.87
80	91.87	94.37	95.37	95.94	96.85
90	92.3	95.12	95.42	96.89	97.5
Accuracy
50	89.24	90.65	90.65	91.87	92.087
60	90.72	91.45	91.45	92.3	92.63
70	91.4	93.34	93.34	93.87	94.087
80	92.87	94.37	94.37	95.94	96.05
90	93.3	95.12	95.12	96.89	97.035
Precision
50	87.24	89.65	91.95	92.87	93.07
60	89.72	90.45	92.75	93.3	94.03
70	90.4	91.34	93.84	94.87	95.07
80	91.87	92.37	94.37	95.94	96.05
90	92.3	93.12	95.12	96.09	97.105
F-Score
50	87.24	89.65	90.65	92.87	93.87
60	88.95	90.45	91.45	93.3	94.063
70	89.4	91.34	93.34	94.87	95.087
80	91.87	92.37	93.57	94.94	95.85
90	93.3	93.12	94.12	95.89	96.65
Specificity
50	88.94	89.65	90.65	92.87	93.57
60	89.92	90.45	91.645	93.73	94.13
70	90.24	91.34	93.634	94.87	95.287
80	91.87	92.37	93.657	94.794	95.585
90	93.3	93.12	94.612	95.89	96.455

The AdCO algorithm significantly contributed by enhancing the feature selection process, ensuring the selection of globally optimal features for breath pattern classification. By dynamically adjusting the weights assigned to different features, AdCO improved convergence rates and optimized the feature subset, resulting in improved classification performance. Furthermore, the DABiG model, which incorporates a BiGRU along with spatial and temporal attention mechanisms, effectively captured the complex patterns within the breath data, leading to enhanced classification accuracy. The combination of AdCO and DABiG allowed for the extraction of informative features and the robust classification of breath patterns, resulting in high performance.

4.2 Analysis by varying K-fold

The analysis by varying the K-Fold is presented in Figure 4 for various assessment measures and the detailed analysis is depicted in Table 2. In this, by dividing the dataset into k subsets, the model is trained and tested k times, with each subset used once as a testing set and the remaining k-1 subsets used for training. This process ensures that the model's performance is accurately assessed across different subsets of the data, reducing the risk of overfitting and providing more reliable performance metrics. Let the precision evaluated by the AdCO + DABiG is 97.37% with K-Fold value 10, which is 1.46%, 6.53%, 6.24%, and 7.71% superior compared to Proposed without FS, 1D-CNN, Deep Breath, and Ensemble learning methods.

Figure 4.

Analysis by varying K-Fold: (a) accuracy, (b) F-Score, (c) precision, (d) recall and (e) specificity.

Table 2.

Analysis by varying K-Fold.

Methods/ K-Fold	2	4	6	8	10
Accuracy
Proposed	97.12	94.98	97.88	95.43	97.46
Proposed without FS	94.1	93.89	94.62	92.73	91.98
1D-CNN	91.42	92.05	94.32	91.75	89.29
Deep Breath	89.42	90.05	90.32	87.75	89.29
Ensemble learning	88.12	85.94	86.76	89.82	87.67
Precision
Proposed	97.42	96.57	97.24	97.91	97.37
Proposed without FS	93.93	94.14	95.67	95.63	95.95
1D-CNN	92.73	91.89	92.95	92.15	91.01
Deep Breath	91.42	92.05	92.32	89.75	91.29
Ensemble learning	90.78	88.94	91	90.12	89.86
Recall
Proposed	96.62	95.05	97.59	96.198	98.52
Proposed without FS	94.58	94.75	94.86	94.64	94.09
1D-CNN	90.12	90.83	91.92	90.21	91.01
Deep Breath	90.12	90.83	91.92	90.21	91.01
Ensemble learning	90.22	90.74	91.05	90.41	90.13
F-Score
Proposed	96.12	94.95	97.89	96.98	96.62
Proposed without FS	94.58	94.75	94.86	94.64	94.09
1D-CNN	92.12	92.83	92.92	92.21	92.01
Deep Breath	90.12	90.83	91.92	90.21	91.01
Ensemble learning	90.22	90.34	90.05	90.41	90.13
Specificity
Proposed	95.12	94.95	96.89	96.98	97.62
Proposed without FS	94.58	94.75	93.86	94.64	94.09
1D-CNN	92.12	92.83	92.92	92.21	92.01
Deep Breath	90.12	90.83	91.92	90.21	91.01
Ensemble learning	89.22	90.34	90.05	90.41	89.13

Here, the analysis indicates the superiority of the proposed method using the novel feature selection and hybrid classifier.

4.3 Convergence analysis

The convergence analysis of the AdCO algorithm and the existing chimp optimization algorithm is portrayed in Figure 5.In this case, the inclusion of an adaptive weighting strategy in traditional chimpanzee optimization assists AdCO in enhancing the convergence rate more rapidly. In addition, the consideration of accuracy, sensitivity, and specificity as the fitness for choosing the optimal best features, assist in enhancing the classification accuracy of the proposed model.

Figure 5.

Convergence Analysis.

4.4 Ablation study

The ablation study of the proposed breath classification method is presented in Table 3, wherein the accuracy analysis is detailed.The analysis indicates the superiority of the proposed method with optimal feature selection and the hybrid deep learning model for breath pattern classification.

Table 3.
Ablation study in terms of accuracy.

Method Accuracy (%)

Proposed (AdCO + DABiG) 97.46

Proposed without AdCO (DABiG) 91.98

AdCO + BiGRU 94.32

BiGRU 89.46

Method	Accuracy (%)
Proposed (AdCO + DABiG)	97.46
Proposed without AdCO (DABiG)	91.98
AdCO + BiGRU	94.32
BiGRU	89.46

4.5 Comparative discussion

The comparative analysis of the proposed breath classification method and the existing methods is portrayed in Table 4.

Table 4.
Comparative analysis.

Reference Accuracy Precision Recall F1-Score Specificity

¹⁸ – 89 97 87 –

¹⁷ – 86.8 87.2 86.6 –

¹³ 96.5 96.4 96.3 96.4 –

Proposed 97.46 97.37 98.52 96.62 97.62

Reference	Accuracy	Precision	Recall	F1-Score	Specificity
¹⁸	–	89	97	87	–
¹⁷	–	86.8	87.2	86.6	–
¹³	96.5	96.4	96.3	96.4	–
Proposed	97.46	97.37	98.52	96.62	97.62

The proposed method acquired superior outcomes in terms of all assessment measures. The utilization of an adaptive weighting strategy in the Chimp Optimization (AdCO) for feature selection, coupled with spatial, temporal, and BiGRU-based breath pattern classification, offers several significant benefits. Initially, the adaptive weighting strategy enhances the traditional Chimpanzee Optimization algorithm by dynamically adjusting the weights assigned to different features during the optimization process. This adaptability allows for a more efficient and effective feature selection process, as it ensures that the most relevant features are given higher importance, leading to improved classification accuracy and model performance. Then, the incorporation of spatial, temporal, and BiGRU-based breath pattern classification techniques further enhances the accuracy and robustness of the classification model. By considering the spatial and temporal dependencies within the breath pattern data, the classification model becomes more adept at capturing complex patterns and variations in the data, thereby improving its ability to accurately classify different breath patterns. Hence, the combination of AdCO for feature selection and spatial, temporal, and BiGRU-based breath pattern classification offers a powerful and effective approach for analyzing and classifying breath patterns.

5 Conclusion

In this research, we have presented a holistic approach to breathing pattern classification by leveraging gyroscope and accelerometer data from individuals. The study begins with data acquisition from two different sensors and includes the collection of six diverse breathing patterns for analysis. Data pre-processing through Min-Max normalization ensures that computations are conducted efficiently. Feature selection is a critical step, and the AdCO algorithm is introduced, a novel optimization technique inspired by the hunting behavior of chimpanzees. AdCO enhances convergence rates and enables the selection of globally optimal features for breath pattern classification. To carry out breath pattern classification, the DABiG is introduced, which incorporates a BiGRU and spatial attention along with temporal attention mechanisms. The performance of the proposed AdCO + DABiG based on Accuracy, Precision, Recall, F-Score, and Specificity evaluated is 97.46%, 97.37%, 98.52%, 96.62%, and 97.62% respectively.

While the proposed approach for breath pattern classification utilizing gyroscope and accelerometer data presents promising results, some several limitations and challenges need to be addressed. Firstly, the effectiveness of the model heavily relies on the quality and quantity of the input data. While the study collects data from two different sensors and includes six diverse breathing patterns for analysis, the variability and representativeness of the data may still be limited. Moreover, while the DABiG model, incorporating BiGRU and spatial and temporal attention mechanisms, shows high accuracy in breath pattern classification, its computational complexity may limit its scalability to large datasets. Therefore, future research should focus on addressing these limitations by collecting more diverse and representative data, optimizing the AdCO algorithm and DABiG model for improved scalability and efficiency, and exploring additional evaluation metrics to assess the model's performance comprehensively. Additionally, the integration of real-time monitoring capabilities and the development of user-friendly interfaces could enhance the practicality and usability of the proposed model for various healthcare applications, such as respiratory disease diagnosis and monitoring.

Footnotes

Ethics approval statement

Not Applicable.

Patient consent statement

Not Applicable.

Permission to reproduce material from other sources

Not Applicable.

Clinical Trial registration

Not Applicable.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability

Data will be made available upon reasonable request.

References

Orlova

Gorobtsov

Sychev

, et al. Method for determining the dominant type of human breathing using motion capture and machine learning. Algorithms 2023; 16: 249.

Mysliwiec

Brock

Matsangas

, et al. 0734 Gender differences in US military personnel with insomnia, obstructive sleep apnea & comorbid insomnia and obstructive sleep apnea. Sleep 2023; 46(Supplement_1): A323–A32A.

Shaw

Harrell

. Integrating physiological monitoring systems in military aviation: a brief narrative review of its importance, opportunities, and risks. Ergonomics 2023; 66: 2242–2254.

Park

Mah

Nguyen

, et al. Modification of a conventional deep learning model to classify simulated breathing patterns: a step toward real-time monitoring of patients with respiratory infectious diseases. Sensors 2023; 23: 5592.

Gupta

Srinivas

Soumya

, et al. Automatic contact-less monitoring of breathing rate and heart rate utilizing the fusion of mmWave radar and camera steering system. IEEE Sensors J 2022; 22: 22179–22191.

CliftonSmith

Rowley

. Breathing pattern disorders and physiotherapy: inspiration for our profession. Phys Ther Rev 2011; 16: 75–86.

Bax

Robbiani

Zannin

, et al. An experimental apparatus for E-nose breath analysis in respiratory failure patients. Diagnostics 2022; 12: 776.

Zamora-Mendoza

Díaz de León-Martínez

Rodríguez-Aguilar

, et al. Chemometric analysis of the global pattern of volatile organic compounds in the exhaled breath of patients with COVID-19, post-COVID and healthy subjects. Proof of concept for post-COVID assessment. Talanta 2022; 236: 122832.

Hylton

Long

Francis

, et al. Real-world use of the breathing pattern assessment tool in assessment of breathlessness post-COVID-19. Clin Med 2022; 22: 376–379.

10.

Yun

Park

Jeong

, et al. A mask-shaped respiration sensor using triboelectricity and a machine learning approach toward smart sleep monitoring systems. Polymers (Basel) 2022; 14: 3549.

11.

Hong

J-W

Kim

S-H

Han

G-T

. Detection of multiple respiration patterns based on 1D SNN from continuous human breathing signals and the range classification method for each respiration pattern. Sensors 2023; 23: 5275.

12.

Singh

Fleury

Gupta

, et al. Involuntary breathing movement pattern recognition and classification via force-based sensors. Biomechanics 2022; 2: 525–537.

13.

Wang

Zhou

, et al. Unobtrusive and automatic classification of multiple people’s abnormal respiratory patterns in real time using deep neural network and depth camera. IEEE Internet Things J 2020; 7: 8559–8571.

14.

Cho

Bianchi-Berthouze

Julier

. DeepBreath: Deep Learning of Breathing Patterns for Automatic Stress Recognition using Low-Cost Thermal Imaging in Unconstrained Settings. 2017.

15.

Liu

Xie

, et al. Early prediction of moderate-to-severe condition of inhalation-induced acute respiratory distress syndrome via interpretable machine learning. BMC Pulm Med 2022; 22: 193.

16.

Tarim

Erimez

Degirmenci

, et al. A wearable device integrated with deep learning-based algorithms for the analysis of breath patterns. Adv Intell Syst 2023; 5: 2300174.

17.

Purnomo

Lin

D-B

Adiprabowo

, et al. Non-Contact monitoring and classification of breathing pattern for the supervision of people infected by COVID-19. Sensors 2021; 21: 3172.

18.

McClure

Erdreich

Bates

JHT

, et al. Classification and detection of breathing patterns with wearable sensors and deep learning. Sensors 2020; 20: 6481.

DABiG: Breath pattern classification using the hybrid deep learning with optimal feature selection

Abstract

Background

Objective

Methods

Results

Conclusion

Keywords

1 Introduction

2 Related works

2.1 Motivation

3 Proposed methodology

3.2 Data normalization

3.3.1 Mathematical modeling

4.1 Analysis by varying training data

Table 3. Ablation study in terms of accuracy. Method Accuracy (%) Proposed (AdCO + DABiG) 97.46 Proposed without AdCO (DABiG) 91.98 AdCO + BiGRU 94.32 BiGRU 89.46

Table 4. Comparative analysis. Reference Accuracy Precision Recall F1-Score Specificity 18 – 89 97 87 – 17 – 86.8 87.2 86.6 – 13 96.5 96.4 96.3 96.4 – Proposed 97.46 97.37 98.52 96.62 97.62

Footnotes

Ethics approval statement

Patient consent statement

Permission to reproduce material from other sources

Clinical Trial registration

Funding

Declaration of conflicting interests

Data availability

References

Table 3.
Ablation study in terms of accuracy.

Method Accuracy (%)

Proposed (AdCO + DABiG) 97.46

Proposed without AdCO (DABiG) 91.98

AdCO + BiGRU 94.32

BiGRU 89.46

Table 4.
Comparative analysis.

Reference Accuracy Precision Recall F1-Score Specificity

¹⁸ – 89 97 87 –

¹⁷ – 86.8 87.2 86.6 –

¹³ 96.5 96.4 96.3 96.4 –

Proposed 97.46 97.37 98.52 96.62 97.62