Harmonizing Respiratory Sound Insights: Unleashing the Potential of Machine Learning Classifiers Through Hyperparameter Elegance

Introduction to Respiratory Sound Analysis and Edge Computing: Respiratory diseases are a significant global health burden that requires rapid and precise diagnosis. Conventional auscultation methods are subjective and variable [6]. Automated respiratory sound analysis via machine learning is a promising approach toward more objective and efficient assessments. The growing availability of mobile and edge computing devices enables point-of-care diagnostics, underscoring the need for computationally efficient solutions.

Deep Learning for Classifying Respiratory Sound: Deep learning models, especially Convolutional Neural Networks, have shown remarkable performance in respiratory sound classification [7-11]. However, their computational demands can hinder deployment on edge devices. Hybrid CNN-RNN architectures [12, 13] and other deep learning methods [14, 15] have been explored, but computational efficiency remains a critical challenge.

Efficient Feature Extraction for Edge Devices: Consequently, the ratio between accuracy and computational cost relies on effective feature extraction. Features from the spectrum domain, including Mel-Frequency Cepstral Coefficients [16, 17], are employed. Feature selection is essential for optimising performance on resource-limited devices. Alternative methods for extracting relevant information from respiratory sounds include wavelet transforms and various signal-processing techniques [18-22]. It is crucial to investigate lower complexity features for implementation in edge devices.

Machine Learning on Edge Devices: The simplification of respiratory sound classifications task on the edge devices by Efficient feature extraction and less computationally demanding machine learning models. Such a practical point-of-care diagnostics solution will likely enhance access to timely and accurate respiratory status measurements. A recent study [23] the feasibility of abnormal sound detection with deep learning and a vest-coat stethoscope, a practical application. Such systems demonstrate evidence of the integration of wearables and edge computing for respiratory health monitoring. It is important to conduct further research to optimize machine learning models to make them work at lower energy levels or on devices with limited resources.

Public Datasets and Challenges: The ICBHI, which aims to automatically detect adventitious respiratory sounds on a large, standardized dataset, and the emergence of public respiratory sound datasets [24] have propelled research in this area considerably. These datasets are helpful for detailed evaluations and comparisons between methods, leading to collaborative innovations.

Addressing Challenges: Noise Reduction and Data Augmentation Specifically, various noise reduction methods like spectral subtraction, as well as wavelet denoising, and data augmentation methods like adding noise, time stretching, and pitch shifting, are essential to achieve the robustness and generalizability of respiratory sound classification models [25]. Finding ways to overcome these challenges has the potential to improve the performance of ML models, especially in-practice clinical scenarios.

This research focuses on developing a state-of-the-art machine learning model of good performance and low computation for automatic classification of respiratory sounds based on the ICBHI 2017 dataset trained on good responses of the ICBHI 2017 dataset. The class imbalance is also tackled through data augmentation and hyperparameter tuning, benefiting its real-world clinical usability in place of traditional analysis methods. In particular, the aims are to:

1. Build and Validate a Strong Model: This will help build a strong machine-learning model that can perform well, even in noise or other real-world differences, ensuring reliability and generalizability in diverse datasets and disease settings.
2. Address Class Imbalance: Address the challenge of class imbalance using data augmentation methods and other balancing approaches so that poorly represented sound categories can be accurately classified.
3. Reduce Computational Complexity: Create a computationally efficient model that can be used for real-time processing and deployment on devices with limited resources, enabling quicker processing and broader access in clinical environments.
4. Validate the Model Using the ICBHI 2017 Dataset: As mentioned in step no. 3, the second task is to validate the model using the ICBHI 2017 dataset. ICBHI 2017 is a widely used dataset in biomedical signal processing, making it easier to compare the model performance against existing literature and measure accuracy/efficacy objectively.

The proposed method for respiratory sound classification outlines a structured methodology for classifying ICBHI lung sound data into normal and abnormal categories. The process begins with normalisation of the raw lung sound recordings to ensure uniformity across the dataset.

1.png

Figure 1. Workflow of the proposed respiratory sound classification method

To improve the dataset's variability and resilience, audio augmentation techniques such as time masking, time shifting, and time stretching are applied. The augmented audio is then transformed into spectrograms, enabling visual representation of the frequency and temporal features of the sounds. Following this, 80 feature extraction techniques are utilised to extract meaningful characteristics from the spectrograms.

A feature selection process is then employed to identify the most relevant features by evaluating their effectiveness across various machine learning (ML) models. These optimised features are subsequently used as input for ML classifiers, which categorise the lung sounds as either normal or abnormal, facilitating effective diagnosis and analysis. The proposed method for respiratory sound classification is illustrated in Figure 1.

The audio database ICBHI 2017 was employed in the current study with 920 audio signals and 6898 respiratory cycles [34]. The samples were collected in a way that the COPD class has more samples as compared to the Asthma and LRTI classes which have a paucity of samples. To this end, data augmentation was used.

Your "Normalizing and Augmentation" section provides essential details about your preprocessing steps. Here are a few suggestions for improvement:

• Clarity and Terminology: While "time masking," "time shifting," and "time stretching" are generally understood, providing a brief explanation or referencing a standard library (like librosa in Python) could enhance clarity. For instance, you could mention that time masking is analogous to masking in image processing, where random sections are zeroed out. Similarly, specifying the range or distribution for time shifting and the stretching factor for time stretching would be beneficial. This level of detail ensures reproducibility and allows others to understand the exact transformations applied.
• Rationale: Briefly explaining why these specific augmentation techniques were chosen would strengthen the section. For example, you could mention that time masking helps the model learn to focus on relevant features even with missing information, while time stretching and shifting increase the model's robustness to variations in speech rate and timing.
• Normalization Details: You mention splitting signals into 4-second breaths, which is a form of normalization. However, you could also specify whether any amplitude normalization was performed. Common techniques include peak normalization, RMS normalization, or simply scaling the audio to a specific range. Including these details would make your preprocessing steps more comprehensive.
• Quantifying Augmentation: Instead of stating that "the number of augmentations vary with the class ratio," provide specific numbers or ratios. For instance, you could say, "The minority class was augmented to achieve a 1:1 ratio with the majority class." This adds precision and makes your methodology more transparent.

4.1 Normalizing and augmentation

The respiratory sound signals were segmented into individual breaths, each 4 seconds long, with a sampling rate of 16000 Hz. To address class imbalance and improve model robustness, we applied several data augmentation techniques using the librosa library in Python:

• Time Masking: Random sections of the audio were silenced, simulating missing information and encouraging the model to learn more robust features. The duration of masked sections was randomly chosen between 0.1 and 0.2 seconds.
• Time Shifting: Sections of the audio were randomly shifted in time by up to ±50 ms, increasing the model's tolerance to variations in speech rate.
• Time Stretching: The duration of the audio was altered without changing the pitch, using a random stretching factor between 0.8 and 1.2. This augmentation further enhances the model's robustness to variations in speech timing.

These augmentations were applied to the under-represented classes to achieve a balanced class distribution of approximately 1:1 with the majority class in ICBHI 2017 dataset. Additionally, the amplitude of each breath segment was normalized using peak normalization to ensure consistent input levels to the model.

Table 2 represents frequencies of the records in each group. Table 3 shows the number of records which have been chosen for augmentation and the number of times each of them has been augmented.

Time Masking: Random sections of the audio were silenced, simulating missing information and encouraging the model to learn more robust features. The duration of masked sections was randomly chosen between 0.1 and 0.2 seconds.

Time Shifting: Sections of the audio were randomly shifted in time by up to ±50 ms, increasing the model's tolerance to variations in speech rate.

Table 2. Number of records in each class

Table 3. Number time each record is augmented

Time Stretching: The duration of the audio was altered without changing the pitch, using a random stretching factor between 0.8 and 1.2. This augmentation further enhances the model's robustness to variations in speech timing

These augmentations were applied to the under-represented classes to achieve a balanced class distribution of approximately 1:1 with the majority class in ICBHI 2017 dataset. Additionally, the amplitude of each breath segment was normalized using peak normalization to ensure consistent input levels to the model.

Table 2 represents frequencies of the records in each group. Table 3 shows the number of records which have been chosen for augmentation and the number of times each of them has been augmented.

4.1.1 Time masking

Randomly masked audio segments between 100-200 ms using the librosa library to simulate missing information. Is the act of allowing a specific portion of the signal in auditory to the listener at a specific time. In this approach parts of the time are set to zero randomly; this in turn causes the time structure in the audio signal to be lost thus the model relies on other properties of sound.

$Y(t)=x(t) * m(t)$     (1)

In Eq. (1), Y(t) denotes the masked signal and x(t) and m(t) is an original audio signal. where m(t) equals to 1 for the cases when segment is preserved and equals to 0 when segment is masked.

4.1.2 Time shifting

Is the process of shifting an audio signal in time or in other words, moving it ahead or behind in time but not changing its content. This manipulation is done by moving the signal along the time axis in some way.

$y(t)=x(t-\Delta t)$     (2)

where, in the above Eq. (2), y(t) is the shifted signal, x(t) is the original signal and Δt is the time shift. If Δt is positive, then the signal is delayed and if Δt is negative then the signal is advanced.

4.1.3 Time stretching

Is the method of altering the temporal characteristics of an audio signal without altering the frequency content of the signal. In mathematical terms time stretching can be expressed by modifying the Short-Time Fourier Transform (STFT) of the sound signal.

$\operatorname{STFT}(x(t))=\sum_{n=-\infty}^{\infty} x(t) \cdot w(t-n H) \cdot e^{-j \omega t}$     (3)

In Eq. (3), x(t) is the input audio signal, w(t-nH) denotes the windowing function that isolates a small portion of the signal around time t. The parameter H controls the window size. n is the integer index which shifts the window across the signal. e^-jωt: This is a complex exponential that represents the frequency component at ω.

4.1.4 The pseudocode for data preparation

x(n): Input audio signal

T1, T2, T3, T4: Time masking, shifting, time stretching augmentation functions.

x'(n) = T4(T3(T2(T1(x(n)))))

x''(n) = x'(n)x'(n)·...,x'(n))

y(n) = truncate_or_pad(x''(n), L)

As depicted in Figure 1 the waveform and spectrogram of the audio sample is depicted before applying any type of audio augmentation. The waveform and spectrogram of the audio when audio augmentation techniques are applied is presented in Figure 2 and Figure 3.

2.png

Figure 2. Waveform and spectrogram before audio augmentation

3.png

Figure 3. Waveform and spectrogram after audio augmentation

It is a fundamental stage of audio processing that involves turning unprocessed input into a numerical format. A total of 80 features were taken from audio signals for this project. These characteristics fall into several categories.

The following categories apply to the features that have been added to the system:

• Mel-Frequency Cepstral Coefficients (MFCCs): These characteristics show the audio signal's envelope since they are the most crucial component in characterizing its tonal properties [35].
• Chroma features: These give the matching audio's perceived pitch, which may reveal details about the chords that are present in the sound.
• Spectral features: The overall form of the spectrum is also described by the spectral centroid, bandwidth, and rolloff.
• Features connected to rhythm: When analyzing the audio signal's rhythm patterns, zero-crossing rate, harmonic, and percussive characteristics are crucial.

The systematic feature extraction approach facilitates the analysis of each acoustic property's contribution to classification. This analysis enables identifying and eliminating superfluous features, leading to a more robust and interpretable model. Table 4 delineates the 80 features retrieved from the audio signals. Table 5 presents samples of extracted feature values alongside their corresponding labels.

Table 4. A set of 80 features were used to extract from the audio signals

Table 5. Extracted feature values and labels from the sample records

4.png

Figure 4. Heatmap for all the features

It is a basic level of audio processing which is a process of converting raw data into a numerical form. The study extracted 80 features categorized as follows:

• Mel-Frequency Cepstral Coefficients (MFCCs): MFCC captures the power spectrum of audio signals, mimicking human auditory perception. These features depict the envelope of the audio signal since this is the most important aspect in describing the tonal aspects of the signal. Represent the short-term power spectrum of audio signals. Calculated using:

$\operatorname{MFCC}_k(t)=\sum_{m=1}^M \log \left(S_m\right) \cos \left[\frac{k \pi(m-0.5)}{M}\right]$     (4)

where, S_m is the mel spectrum, and M is the number of mel filters.

• Chroma features: Capture the harmonic structure of sounds by mapping audio frequencies onto 12 chroma bins representing musical pitches.
• Spectral features: Include centroid, bandwidth, and rolloff, which describe the overall shape and energy distribution of the audio spectrum.
• Rhythm-related features: Incorporating zero-crossing rates and harmonic/percussive properties to capture rhythmic patterns.

This systematic feature extraction ensures comprehensive representation of the audio signals, facilitating high-accuracy classification. This analysis enables the identification and potential removal of less relevant features, leading to a more efficient and interpretable model. Table 2 lists the 80 features extracted from the audio signals. Table 3 presents examples of extracted feature values and their corresponding labels. Figure 4 provides a heatmap visualization of the relationships between all features used in the model.

Table 6 represents grouping 80 features into 9 feature set groups allows for more efficient data management and analysis by categorizing similar features together [36]. This approach enhances the interpretability of the data, making it easier to understand and visualize the relationships between different features. In addition, it can improve the performance of machine learning models by reducing redundancy and focusing on the most relevant feature subsets for specific tasks [37].

Table 6. Features are grouped

• df_feature_all this group comprises all extracted features that provide a summary of the audio signal. This category encompasses intricate characteristics and facilitates audio analysis across all levels. Utilizing its capacity to exhibit harmonic and spectral characteristics along with temporal variations, we constructed an extensive audio representation. A holistic approach can improve the efficacy of machine learning models by leveraging many examples and diverse attributes to feed the modelling process.
• df_feature_mel_chroma the integration of harmonic (Chroma CENS) and spectral (Mel-Spectrogram) features yields df_feature_mel_chroma, which is profound, stable, and enhances performance in music genre classification [38].
• df_feature_mfcc_mean comprises the average values of the Mel-Frequency Cepstral Coefficients (MFCCs), which denote the short-term power spectrum of audio. MFCCs are extensively utilized in speech recognition and music categorization due to their capacity to represent the spectrum characteristics of audio signals. This group, favoring mean values, provides an average representation of spectral features, rendering it appropriate for advanced pattern and trend recognition. A collective that enhances the efficacy of the machine learning model by offering a consistent and dependable feature collection that encapsulates essential spectral and spatial data.

$M F C C_k(t)=\sum_{m=0}^{M-1} \log (M[m, t]) \cdot \cos \left(\frac{\pi k}{M}\left(m+\frac{1}{2}\right)\right), k=0,1, \ldots, 12$     (5)

where, MFCC_k(t) denotes the k-th Mel-Frequency Cepstral Coefficient at time frame t, $\sum_{m=0}^{M-1}$ signifies a summation across M Mel filter banks. The $\log (M[m, t])$ denotes the logarithm of the Mel filter bank output at index m and time frame t. The $\left(\frac{\pi k}{M}\left(m+\frac{1}{2}\right)\right)$ represents the cosine function utilized in the Discrete Cosine Transform (DCT) and k=0, 1, …, 12 signifies that the MFCCs are calculated for k values ranging from 0 to 12.

The advantages of amalgamating 13 MFCC features into a singular feature vector are as follows. The methods utilized in generating these audio signal representations provide a more succinct and efficient form of the signal, preserving the most pertinent elements of the spectral content. It fundamentally diminishes the dimensionality of the data, facilitating processing and analysis. Mean values of the MFCCs are also calculated, which provide helpful information about the generic spectral envelope of the signal, enabling discrimination between sounds and extract patterns. The method enables the extraction of stable characteristics from dynamic and transient inputs, such as heartbeats, voice, and music, enhancing model performance in classification tasks. It also assists in reducing noise and variability, leading to a more stable and precise evaluation.

• df_feature_mfcc_std quantifies the variability of the spectral envelope, augmenting the mean values and enriching the feature set for the comprehensive investigation of spectral features.

$\mu_{M F C C i}=\frac{1}{T} \sum t=1^T M F C C_i(t)$ for $i=0,1, \ldots, 12$     (6)

$\mu_{M F C C i}$ calculates the mean value $\mu_{M F C C i}$ of each MFCC coefficient by averaging the MFCC values across all time frames.

• df_feature_mfcc integrates both mean and standard deviation values of the MFCCs, offering a thorough representation that encapsulates both the average and variability of the spectral envelope.

$\begin{gathered}\operatorname{mfccs}_{-} s t d_i=\sqrt{\sum_{t=1}^T\left(\operatorname{MFCC}_i(t)-\text {mfccs_mean}_i\right)^2} \\ \text { for } i=0,1, \ldots, 12\end{gathered}$     (7)

where, (mfccs_std_i) calculates the standard deviation of all Mel-Frequency Cepstral Coefficient (MFCC) values across all time frames is given here. The term $\frac{1}{T}$ is the normalization factor, where $T$ is the total number of time frames, $\sum_{t=1}^T\left(\operatorname{MFCC}_i(t)-\text {mfccs_mean}_i\right)^2$ aggregates the squared differences between the $i$-th MFCC values and their mean (mfccs_mean_i) across all frames.

• df_feature_chroma_mean encapsulates the harmonic and melodic elements of the audio, offering a consistent and dependable feature set for discerning overarching patterns in music information retrieval tasks. It integrates mean and standard deviation data to create a more comprehensive feature set, facilitating an in-depth investigation of spectral features.
• df_feature_chroma_std measures the variability of the harmonic content, complementing the mean values and enhancing the feature set for detailed analysis of the harmonic characteristics and also includes the standard deviation values of the MFCCs, which measure the variability of the spectral envelope.

$\begin{gathered}\text { chroma_mean }_i=\frac{1}{T} \sum_t t=1^T \operatorname{Chroma}_i(t) \\ \text { for } i=0,1, \ldots, 11\end{gathered}$     (8)

• df_feature_chroma amalgamates the mean and standard deviation of the Chroma features, offering a thorough representation that encapsulates both the average and variability of the harmonic content.
• df_feature_csrzhp offers a comprehensive spectral analysis, including spectral centroid, bandwidth, roll-off, zero crossing rate, and harmonic and percussive content, encapsulating the audio signal's intricacies and dynamics. It integrates both temporal and frequency domains, yielding a comprehensive analysis of the sound's spectral attributes, including spectral centroid, bandwidth, roll-off, harmonic and percussive content, and zero crossing rate, which elucidate the subtleties of the audio signal.

$c s r z h p(x)=\left\{C_s(x), B_w(x), R_f(x), Z_r(x), H_p(x)\right\}$     (9)

This csrzhp(x) is the grouping of these spectral feature extractions applied to the input audio signal x.

Where C_s is a Spectral Centroid, B_w is a Spectral Bandwidth, R_f is a Spectral Roll-off, Z_r is a Zero Crossing Rate and H_p is a Harmonic and Percussive Content. This equation integrates multiple spectral and temporal aspects into a unified feature vector, offering an extensive description of the audio signal's attributes.

In this research article, a framework is presented meticulously for comparing the effectiveness of several machine learning algorithms in classifying lung sounds. A complex experimental design was developed, utilizing several classifiers, including Decision Trees, Random Forest, and Support Vector Machines. To assess the impact of feature representation, various feature sets were constructed based on extracted features from audio signals, including Mel-Frequency Cepstral Coefficients, chroma features, and other acoustic features. The model was trained and tested using different training and testing data splits (0.2, 0.3, 0.4, and 0.5) to evaluate model stability.

The feature sets (MFCC Mean, MFCC Mean+Std, and Chroma Mean+Std) also yielded good results; however, the most comprehensive feature set, "All Features, MFCC (Mean+Std), achieved the highest accuracy. Tables 7-10 present the average accuracy for All Features, MFCC Mean, MFCC Mean+Std, and Chroma Mean+Std at a split ratio of 0.20. Table 11 shows the accuracy, F1-score, precision, recall, and ROC scores for these feature groups, with All Features and MFCC (Mean_std) achieving good accuracies.

The combination of MFCC Mean+Std features with KNN and SVM classifiers achieved the highest accuracy (99%). This performance can be attributed to the robust feature representation of MFCCs, which effectively capture spectral properties crucial for distinguishing respiratory sounds.

Compared to state-of-the-art methods:

• This approach demonstrates superior performance while maintaining computational efficiency, making it suitable for deployment on edge devices.

Comparative Analysis:

• CNNs: Achieved 97% accuracy in prior studies but at a higher computational cost.
• Proposed Model: Demonstrated state-of-the-art accuracy while being more efficient.

Table 7. ML models average accuracy with split ratio 0.20

Table 8. ML models average accuracy with split ratio 0.30

Table 9. ML models average accuracy with split ratio 0.40

Table 10. ML models average accuracy with split ratio 0.50

Table 11. ML models average accuracies on All features, MFCC mean, MFCC mean+std, chroma mean+std

Error Analysis

Confusion matrices reveal that misclassifications primarily occurred in overlapping classes (e.g., Crackles+Wheezes). Augmentation mitigated errors in minority classes.

To compare the classification results across different feature sets and machine learning models, a cross-comparison of tables was performed. This matrix compares the performance of different features and models, including Decision Trees and Random Forest, using features such as All Features, MFCC mean, MFCC mean+std, see Figures 5-7. The average accuracy results across different models, features, and split ratios suggest that the best-performing model utilizes all features and a split ratio of 0.20. The study [4] achieved 97% accuracy using CNNs, while our approach achieved 99% with KNN and SVM, demonstrating superior performance with lower computational costs.

5.png

Figure 5. Cross-matrix representation of ML models using all features (df_feature_all) group

6.png

Figure 6. Cross-matrix representation of ML models using MFCC mean features

7.png

Figure 7. Cross-matrix representation of ML models using Chroma(mean+std) features