COVID-19 Detection from Cough Sounds Using XGBoost and LSTM Networks

The SARS-CoV-2 (COVID-19) virus, identified in December 2019, has affected hundreds of millions of individuals worldwide, resulting in a global-scale mortality. According to medical references [1-3], COVID-19 infection is characterized by a variety of symptoms, including fever, fatigue, and a persistent cough [4, 5]. Understanding these symptoms and detecting them early is crucial for accurate diagnosis. Generally, the COVID-19-associated cough is typically dry, without mucus, and can result in intense, recurring coughing episodes that may endure for an extended duration. Hospitals and labs use various testing methods, such as PCR, Chest CT imaging, and antigen tests, which vary in effectiveness, cost, and speed [6, 7]. Utilizing artificial intelligence (AI) for COVID-19 detection through cough analysis is a promising alternative [8]. AI techniques, primarily machine and deep learning-based, are user-friendly, non-invasive, offer rapid detection, and ensuring swift results. In this regard, we will introduce several studies from the literature that focus on the automatic detection of COVID-19 based on cough sounds.

A review article [9] discussed recent research on COVID-19 diagnosis using AI and respiratory sound analysis. The researchers [10] outlined AI-driven efforts for diagnosing COVID-19, offering a valuable automated system that utilizes non-invasive biological signals from both speech and non-speech audio. In their study, using cough signals, Pahar et al. [11] employed two datasets, Coswara and Sarcos, derived from smartphone voice recordings. Employing seven machine learning classifiers with various features, their top-performing classifier, ResNet50, achieved an impressive Area under the ROC Curve (AUC) of 0.98, while the LSTM-based classifier reached an AUC of 0.94. A study by Imran et al. [12] employed a deep transfer learning-based multi-class classifier to diagnose COVID-19 based on cough, achieving an accuracy of 92.64%. MFCC features were extracted from cough recordings and processed through a Convolutional Neural Network (CNN) architecture featuring three pre-trained ResNet50 models and a parallel Poisson biomarker [13]. The results exhibited a remarkable COVID-19 sensitivity of 98.5% with a specificity of 94.2% and an AUC of 0.97. For asymptomatic individuals, the sensitivity reached 100% with a specificity of 83.2%.

Ponomarchuk et al. [14] claim to introduce a new deep learning and signal processing-based approach to detect COVID-19, as well as providing denoising methods, cough detection and classification. A classification was made between a dry cough and wet cough by applying some modern and traditional machine learning methods to the COUGHVID database, where it has been shown that the decision tree models are superior to the rest of the models in terms of overall performance [15]. Based on logistic regression and support vector machines for acoustic data, and decision tree models for symptoms data, the researchers [16] proposed a multi-modal diagnostic for COVID-19. Using Coswara dataset, an AUC of 0.92 has been reached.

To the best of our knowledge, all the examined relevant COVID-19 diagnosis methods utilizing cough signals are vulnerable to environmental noise and undesirable signals, which could compromise the efficiency of AI-driven COVID-19 detection systems. To address this noise issue, this study suggests an automated system referred to as XGBoost-LSTM, which relies on XGBoost and LSTM algorithms for COVID-19 detection using cough audio signals.‏ Specifically,‏ the main contributions of this work are:

(1) Generally, datasets of cough sounds are often contaminated by environmental noise and undesirable parts (laughter, speech, music, and loud instruments), which may affect drastically‏‏ AI system training. Thus, an XGBoost-based signal preprocessing system is used to accurately detect cough segments in recordings. Then, the scheme‏‏ evaluates cough segment quality by estimating the signal-to-noise ratio (SNR) of the audio signal, and effectively preserves records that actually contain cough sounds.

(2) COVID-19 affects males and females differently [17-19], leading to a slight distinction in their cough signals. As a result, two separate classification models have been developed for each gender.

(3) To classify the detected cough segments extracted from the acquired audio recordings into COVID-19+ and COVID-19-, the LSTM neural network, capable of modeling time series variations, is employed along with MFCC features and log energy.

The remainder of the paper is laid out as follows. A general idea of the proposed framework is depicted in section 2. In particular, Cough detection and Cough Classification are described. In Section 3, the experimental results with a detailed discussion are reported. The conclusion is described in Section 4.

3.1 Datasets description

3.1.1 COUGHVID dataset

Over 25,000 crowdsourced cough recordings are available in the COUGHVID collection, covering a wide range of participant ages, genders, geographic areas, and COVID-19 statuses. Between April 1st, 2020, and December 1st, 2020, all of the recordings were collected using a Web application hosted on a private server at the École Polytechnique Fédérale de Lausanne (EPFL) in Switzerland. There are roughly 35 hours of audio samples in the COUGHVID database corresponding to almost 37,000 segmented coughs with a sampling frequency of 48 kHz. More than 2,800 recordings were labeled by four qualified physicians in order to diagnose medical problems from the cough. All publicly available data records, as well as metadata, are stored in the Zenodo repository (https://zenodo.org/record/4498364#.YgBQj-rMJPY) [20], where we can distinguish three types of variables in metadata: (1) context information (timestamp and likelihood that the recording actually contains cough sounds), (2) user-reported information, and (3) labels provided by expert medical annotators about the cough recordings clinical assessment.

3.1.2 VIRUFY dataset

VIRUFY is an open source database (https://github.com/virufy/virufy-data) containing cough sounds for the two categories (COVID-19 positive and negative). This database was collected through the VIRUFY mobile data collection application as a resource of crowdsourcing data by smartphones public users. With the help of VIRUFY clinical researchers and medical advisors [31], the VIRUFY database was attached to information (symptoms, medical history, gender, and age) for all patients along with PCR test results and whether the patient was a smoker or not. VIRUFY contains 16 subjects (10 males and 6 women) containing more than 120 cough sound samples between positive and negative COVID-19 sounds sampled at 48kHz.

3.2 Dataset cleaning

The utilization of crowdsourcing has paved the way for unconventional approaches in contrast to traditional data collection methods. This, in turn, has positioned crowdsourcing as a catalyst for research, given its capacity to facilitate cost-effective and expedited data access within the research community. Nonetheless, integrating crowdsourced data into artificial intelligence applications continues to encounter a range of challenges, with data quality, as highlighted in the study of Lease [32], standing out as a key concern. Crowdsourced data often includes samples unrelated to the content subject matter. Consequently, this study incorporates data cleaning processes to address this issue by using the metadata in order to choose the most reliable data in building the proposed automated system.

To clean the data from the COUGHVID database, we selected records that have a probability greater than 0.8 by the XGBoost cough detection model described earlier which are assumed to contain cough. Then, we selected the subjects whose expert opinion agrees with the self-report of people with COVID-19 or healthy people only, while excluding subjects with other pathologies, since our main objective is the detection of people with COVID-19. After cleaning, the retained recording samples are reported in Table 2.

Concerning the VIRUFY database, depicted in Table 3, it does not require any cleaning due to its high-quality recording.

Table 2. Cough samples retained after cleaning the COUGHVID dataset

Table 3. Cough samples retained after cleaning the VIRUFY dataset

3.3 Results and discussion

The LSTM model performs the classification of cough audio segments transformed into MFCC supported by log energy features at its input. Using the Scikit-Learn random search algorithm with COUGHVID and VIRUFY datasets, the most effective obtained model configuration, adopted in this work, is presented in Table 4. It includes a sequence input layer for 14-dimensional sequences, a BiLSTM layer with 100 hidden units for capturing temporal dependencies, two fully connected layers for information transformation, a sigmoid activation function, and a classification output layer utilizing binary cross-entropy.

During experiments on the COUGHVID and VIRUFY datasets, two LSTM models, one for each sex, were utilized, since physiological differences in the phonatory apparatus has an impact on the two sexes in terms of infection and symptom variance, and therefore may affect the cough classification process [17-19]. Additionally, these two databases provide descriptive information on the sexual status of audio recordings.

In order to evaluate the performance of the proposed system, a set of metrics, namely specificity, recall (sensitivity), precision, accuracy and F1 score, were used:

- Specificity is a measure of how many healthy people can be correctly identified as healthy:

$Specificity=T N /(T N+F P)$       (10)

- Recall (sensitivity) indicates the proportion of genuinely unhealthy individuals correctly predicted:

$Recall =T P /(T P+F N)$      (11)

- Precision is, intuitively, the classifier capacity to avoid identifying a healthy sample as a positive covid-19 sample:

$Precision =T P /(T P+F P)$ (12)

- Accuracy is the proportion of correctly identified individuals to the total number of samples:

$Accuracy =\frac{T P+T N}{T P+T N+F P+F N}$         (13)

- F1-score is defined as an average of the model precision and recall:

$F 1 \_score =\frac{2 \times(\text { Recall } \times \text { Precision })}{(\text { Recall }+ \text { Precision })}$      (14)

where,

- TP (True positive): The classifier detected covid-19 when covid-19 cough was present.

- TN (True negative): The classifier identified healthy cough when the healthy cough was present.

- FP (False positive): The classifier identified covid-19 in the presence of a healthy cough.

- FN (False negative): The classifier detected the healthy cough when covid-19 was present.

Tables 5-7 display the simulation results obtained for the XGBoost-LSTM diagnostic system using the COUGHVID and VIRUFY datasets with different LSTM minibatch sizes. These results were obtained through a 4-fold stratified cross-validation (CV) approach, where the data was divided into four subsets. Three folds were used for training, constituting 75% of the data, while one fold was reserved for testing, representing the remaining 25% of the data. This approach improved the reliability of model evaluation by systematically testing it on various data segments, thereby enhancing its overall performance.

Table 5 and Figure 6 present the results for females in the COUGHVID dataset, where it can be noted that the best results were obtained for SNR ≥15dB with F1-score=92% and sensitivity=94.73%. It is noteworthy that for the case of better recording quality with SNR≥20dB, a lower F1-score=90% was achieved which may be due to the low number of training cough samples, that may result in underfitting of the model.

Table 4. The LSTM configuration

Table 5. Female classification results for COUGHVID dataset dataset (LSTM Minibach=30)

Table 6. Male classification results for COUGHVID dataset (LSTM Minibach=25)

Table 7. Classification results for VIRUFY dataset (LSTM Minibach=27)

6.png

Figure 6. Female classification results for COUGHVID dataset

7.png

Figure 7. Male classification results for COUGHVID dataset

Table 6 and Figure 7 report the results for males in the COUGHVID dataset where the best results were achieved for SNR ≥10dB with F1-score=82% and a sensitivity of about 97%. For the cases of SNR ≥15dB where there is not enough number of samples to train the model properly, the classification system performs poorly with relatively low values of sensitivity and F1-score.

Table 7 and Figure 8 show the results for the VIRUFY dataset for both females and males. Because the recordings in this dataset are of good quality, the SNR level was not taken into account. Female samples reached a percentage of 100% for all metrics, whereas male samples achieved a sensitivity (recall) of 90.9% and an F1-score of 95%. It is important to note that sensitivity and F1-score were the primary metrics used to assess the results because of their significance in evaluating the classifier ability to detect pathological samples. Meanwhile, specificity, accuracy, and precision play a critical role in evaluating the classifier performance in identifying healthy samples and determining the overall classification accuracy.

8.png

Figure 8. Classification results for VIRUFY dataset

Using the VIRUFY dataset, the suggested scheme shows high classification performance for both sexes, with a precision and specificity of 100%, and an average F1-score and accuracy of nearly 98%. It is noteworthy that all the data of this dataset are labeled with COVID-19 PCR test status and are quite accurate due to the fact that they were taken at a hospital under the observation of physicians using standard operating procedures. For the COUGHVID dataset, which is relatively less accurate due to self-reported information provided by the user and the lack of agreement among the experts about the patient health status, the suggested diagnosis system exhibits thoroughly good classification performance in terms of the SNR, and, specifically, a quite high sensitivity (recall) as reported in Tables 5 and 6.

Table 8. A comparative table of our classification results with relevant results in the literature

In order to evaluate the efficiency of the proposed COVID-19 detection technique, Table 8 gives a comparison of the classification achieved results with those reported in the literature based on the same two datasets. It can be considered that the proposed model outperforms the literature results for both datasets. First, from Table 8 it can be noticed that the suggested model performs better for females than for males, achieving an F1-score of 92% for COUGHVID and 100% for VIRUFY datasets. This difference may be attributed to the fact that the female model was trained on an imbalanced dataset with more COVID-19 cases than healthy cases, as shown in Table 5, while the male model was trained on a balanced dataset, as reported in Table 6.

For both datasets the suggested scheme exhibits relatively better sensitivity with 94.73% for females and 97.05% for males with COUGHVID dataset, and 100% for females and 90.9% for males with VIRUFY dataset. Furthermore, it also showcases superior accuracy, with results of 91.89% for females and 84.61% for males in the COUGHVID dataset, and 100% for females and 95.23% for males in the VIRUFY dataset. Additionally, the precision of our proposed scheme is noteworthy, attaining 90.0% for females and 71.7% for males in the COUGHVID dataset, and a 100% for both females and males in the VIRUFY dataset.