2.1.1 Questionnaire design.

A digital well-being questionnaire was developed for the larger study consisting of 72 items in total [64]. The survey was circulated among the students using a Google form. Participants’ voluntary consent was collected prior to the survey being conducted. There were 5 items dedicated to collecting demographic information: gender, age, residence, ethnicity/geographical location, and academic level of study. Two items were targeted to obtain data related to sleeping hours and the amount of time students were connected to the Internet. Two items inquired into student satisfaction with sleep and academic study. The depression component of the questionnaire consists of 7 items. Each of these 7 items was answered on a 4-point rating scale along a continuum from Not at all true (0) to Always true (3).

This study particularly focused on depression for which the DASS-Depression subscale had been adopted. DASS-Depression has been adopted from DASS-21 which was established by Lovibond et al. in 1995 to investigate the depression of a cohort [65]. The 7 items that were utilized for this study are: I could not seem to experience any positive feeling at all; I found it difficult to work up the initiative to do things; I felt that I had nothing to look forward t; I felt downhearted and sad; I was unable to become enthusiastic about anything; I felt I was not worth much as a person; and I felt that life was meaningless.

2.1.2 Ethics statement.

This study is approved by the Committee for Advanced Studies and Research (CASR), Islamic University of Technology (IUT), Bangladesh. The consent was collected from the participants at the beginning of the survey.

2.1.3 DASS-Depression scale.

The Depression-Anxiety-Stress Scale (DASS) is a widely used survey instrument to evaluate depression experienced by the people. There are other numerous scales for depression screening such as Beck Depression Inventory (BDI), Patient Health Questionnaire (PHQ), Hamilton Rating Scale, Geriatric Depression Scale (GDI), etc. However, DASS can be regarded as a benchmark scale to evaluate the depressive mental condition of [66]. The full scale consists of 42 items which are divided into three subscales for measuring depression, anxiety, and stress. The reduced 21-item-based DASS-21 scale also exhibits similar psychometric properties as the larger scale. As discussed earlier, this scale has been validated among different populations of the world and has a good internal consistency among the depression-related items (Cronbach alpha, α = 0.94) [67]. Unlike some other scales, it can differentiate mood disturbances and core depression symptoms such as dysphoria, hopelessness, self-deprecation, inertia, etc. so that this characteristic has the potential to help medical practitioners to gain a better understanding of a pupil’s mental health [67, 68]. The 7-item DASS-Depression scale is one of the three self-report subscales of the DASS-21. The scoring for each item is based on a 4-point severity scale and the sum of the 7 items’ scores is multiplied by 2 to find out the depression severity score of each participant [69]. These five depression severity levels are determined as follows:

• Normal (scores 0–9)–labelled as 0
• Mild (scores 10–13)–labelled as 1
• Moderate (scores 14–20)–labelled as 2
• Severe (scores 21–27)–labelled as 3
• Extremely Severe (scores 28+)–labelled as 4

2.2.1 Survey objective.

The primary objective of this survey was to assess the mental condition of the students due to their increased exposure to digital tools, mobile connectivity and internet activities associated with the hybrid or distance mode of learning. Besides, sleep data and internet connectivity status were also collected to see whether there was an irregular connection between the hybrid mode of learning and physical activity.

As a general guideline, participants were instructed to not think about a question for too long and all the items related to the mental condition were to be filled in based on the experience of the past 4 weeks. At the end of the survey, a total of 444 responses were submitted and the data was then saved into a.csv file for further processing.

2.2.2 Exploratory data analysis.

Of the 444 respondents, 291 were male and 153 were female participants ranging from below 20 years to over 35 years of age. University students usually enroll after the age of 17 years. Thus ‘below 20 years’ implies the range between 17 to 20 years. The survey was primarily focused on the students at the Islamic University of Technology (IUT) situated in Gazipur, Bangladesh. As a result, most of the participants were aged between 21 to 25 years. Besides, since the university operates under the umbrella of an international organization named Organization of Islamic Cooperation (OIC); international students from more than 20 countries such as Iran, Pakistan, Brunei, Indonesia, Uganda, Nigeria, Cameroon, Gambia, Yemen and so on were attending this university under this international banner. These participants particularly have experienced their academic activities being transitioned from remote learning to hybrid mode. Moreover, the participants ranged from undergraduate-first year to postgraduate studies and from several types of contexts (e.g., small towns, big cities), which added to the authenticity of the results obtained through the survey. First-year students were comparatively higher in number than the other four academic levels and the significantly higher number of respondents came from big cities rather than from villages or small towns. The overall data distribution is shown in Tables 1–5.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. The male-to-female ratio of the study group.

https://doi.org/10.1371/journal.pone.0294803.t001

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. The age range of the study group.

https://doi.org/10.1371/journal.pone.0294803.t002

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Ethnicity/geographical location of the study group.

https://doi.org/10.1371/journal.pone.0294803.t003

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Academic level of the study group.

https://doi.org/10.1371/journal.pone.0294803.t004

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Residence of the study group.

https://doi.org/10.1371/journal.pone.0294803.t005

An elaborated analysis with the findings of the relationship between survey data and depression severity is demonstrated in the next section as stated in the research contribution.

2.4.1 Feature selection methods.

Feature selection methods select the most apposite features by disregarding the irrelevant ones so that the ML models do not suffer from mediocre performance on high-dimensional data [75]. It is also more interpretable to the researchers as it reduces the training time. In this study, 30 features out of 61 features were selected using both the Chi-square test and RFE method which are discussed below.

Chi-square test. In the Chi-square (χ²) test, the observations are categorized into classes that are mutually exclusive in conventional applications. The test statistic generated from the observations follows a χ² frequency distribution if the null hypothesis, that there are no significant differences between the classes in the population, is true. A random sample of n observations from a population is divided into k mutually exclusive classes, each with a set of observed numbers, x_i (for i = 1,2,…,k). The dependency of two events is tested by using this formula [75]: (1) where, x_i = observed value, and expected value, m_i = np_i, n being the number of samples. And, (2) Here, high χ² value means the feature set has significant relevance with the output and thus that feature set can be used for the training of the ML models.

Recursive Feature Elimination (RFE). RFE method falls under the filtering category of using a proxy measure to evaluate a feature subset. In place of a clear best feature subset, many filters offer a feature ranking, with cross-validation used to determine the ranking’s cutoff point. RFE is one such method which keeps only the best key features by calculating the weights of each feature for a given output. As a result, the model complexity gets much reduced and can be run efficiently by training on an optimized set of features. The weighted voting scheme follows this equation [76]: (3) where, D(x) = decision function, w = weight vector and μ = is the mean vector over all training patterns. The ranking is determined by using the following equation in an iterative process: (4) where, J = cost function. The iterative algorithm is as follows:

1. Train the classifier first by optimizing the weights w_i with respect to J.
2. Determine the DJ(i) or (w_i)² ranking criterion for each feature.
3. Discard the component with the lowest ranking criterion.

2.4.2 Feature extraction methods.

In the case of feature extraction, a new set of features is used for model training, which are extracted from the existing features. The outcome is similar to feature selection, that is, a reduced and concise set of features capable of predicting the outcome most precisely. The methods also help to save memory and computational power while producing adequate performance. Though feature extraction methods are vastly employed for image processing purposes, discrete valued features can be tested as well to explore their effectiveness [77]. The feature extraction methods applied in this study were PCA and SparsePCA, both of which are discussed below:

Principal Component Analysis (PCA). PCA is a classical statistical method for transforming attributes of a dataset into a new set of uncorrelated attributes called principal components. PCA can be used to reduce the dimensionality of a dataset, while still retaining the variability of the dataset as much as possible. In PCA, a matrix X, which is a N×p matrix, serves as the dataset’s representation. All the observations for one attribute are contained in each column, X_j. The set of inputs X₁,X₂,…,X_n is transformed by PCA into another set of column vectors T₁, T₂,…,T_N. A p×p matrix P specifies this linear transformation of the matrix X such that the transformed variables T are given by: (5) Alternatively, the equation can be rewritten as: (6) where P is known as the loading matrix, and columns of this matrix can be calculated as eigenvectors of the matrix X^TX [77].

Sparse Principal Component Analysis (SparsePCA). By adding sparsity structures to the input variables, SparsePCA expands the traditional PCA approach for reducing the dimensionality of data. The principal components of standard PCA are often linear combinations of all input variables, which is a particular drawback. By identifying linear combinations with a smaller number of input variables, SparsePCA gets around this drawback [78]. The sparse factors that account for the greatest amount of variance can be expressed as follows: (7) where, z∈Rⁿ, ρ = parameter controlling sparsity, Card(z) = the number of non-zero coefficients of z. There are several solutions to this problem, one of which is Greedy Search Algorithm [78]. For this algorithm, the input would be ∑∈R^n×n. The steps are as follows:

1. Preprocessing: Sort variables using diagonal elements in decreasing order, then permute ∑ elements, as required. Perform the Cholesky decomposition calculation ∑ = A^TA.
2. Initialization:
3. Compute
4. Set and compute x_k+1 as the leading eigenvector of .
5. Set k = k+1. If k<n, go back to step 3.

2.5.1 Gaussian Naïve Bayes.

It is a classifier that uses only probabilistic models and the Bayes theorem. The fundamental presumption in this situation is that the features for which an outcome is anticipated are independent of one another [79]. The following is the classifier formula: (8) The probability estimate will be zero if a feature never appears in any categories in the training data [54]. This happens because the probability estimate is related to the number of times a feature’s value occurs.

2.5.2 Logistic Regression.

The "logistic function" is used as the cost function in this algorithm, which is based on the idea of probability. The algorithm’s nature limits this cost function to a range between 0 and 1, and its formula is as follows: (9) To distinguish one class from another in a classification task, a threshold probability value is specified [80]. The same general idea applies to multiclass regression, with the exception that there are M possible outcomes rather than simply two.

2.5.3 Random Forest.

The decision tree concept is used in random forests, which is extended further by incorporating many different individual trees. A weight is assigned to each of these trees’ results, and the sum of those weights serves as the foundation for determining the anticipated class through voting. The bagging formula is as follows: (10) where, f_b = classification tree, B = number of bagging (1, 2…, b), x′ = training set. This algorithm offers more accuracy compared to other classification techniques because it is an ensemble classifier that takes help of multiple decision trees [81].

2.5.4 Support Vector Machine (SVM).

This technique locates a hyper-plane that divides a collection of points in an N-dimensional space depending on features. To secure the classification process for the future, the hyper-plane must maximize the margin between the classes of data. Support vectors are the values closest to the hyper-plane in this case, impacting the margin’s position and orientation. The equation for linear SVM, among other types, is as follows: (11) Here, w^T = normal vector to the hyperplane, and b determines the offset distance from the origin. SVM has the advantage of being less prone to the over-fitting issue than other algorithms [82].

2.5.5 Light Gradient Boosted Machine.

Shortened as LightGBM, this algorithm is developed using a decision tree algorithm and has the ability to reproduce results like a boosting algorithm (XGBoost). Regarding efficiency and memory usage, LightGBM employs a highly efficient decision tree learning technique based on histograms. As an ensemble algorithm, any arbitrary differentiable loss function and the gradient descent optimization procedure are used to fit the models. A few steps of the algorithm Gradient-based One-Side Sampling technique used in LightGBM is given below [83]:

1. Input: training data (I), iterations (d), sampling ratio of large gradient data (a) and small gradient data (b), loss function (L)
2. 
3. 
4. For d iterations, ; where, ,

2.5.6 Multi-layer Perceptron.

Multi-layer Perceptron or MLP performs similarly to a supervised learning method by adding backpropagation during training. Specifically, this feedforward network requires at least three layers of nodes namely, the input layer, hidden layer, and output layer. One of the activation functions utilizing the hyperbolic tangent function is as follows: (12) Here, node output is represented by y and v = weighted sum of the inputs. The node weights that are to be modified in the backpropagation process in accordance with the change that reduces overall output error, follows this equation [84]: (13) where, e_j = degree of error in an output node j while n represents a datapoint.

2.5.7 CatBoost.

Another boosting framework that has been used in this study is CatBoost which is one of the most frequently used ML models in the modality. In contrast to LightGBM which has a probability of losing information for clustering tail categories according to expected target value (target statistics), CatBoost uses the target statistics as new numerical features. These features follow an ordering principle in which a random permutation is introduced for the training samples and the preceded memory is used to calculate target statistics [85]. An example of a training and a testing sample are as follows: (14) where, σ signifies a random permutation of j or k samples to satisfy the target statistics equation.

2.5.8 K-Nearest Neighbor (KNN).

Finding the data points that are close to one another is how the KNN algorithm operates. To compare a new point to k numbered reference points in order to establish how close it is, Euclidean distance is typically utilized. The k value in this case determines the amount of stability the algorithm will maintain in precise predictions [86]. A point in a feature space that should be given to the category of closest neighbors is: (15) where, x = a point in space, and Y₁ = a class the point is assigned to. It is more straightforward to construct compared to other models because no additional suppositions or adjusting are required.

The parameters used for each of the classifiers can be seen in Table 6. Other than the LightGBM and Catboost classifiers, which have their own libraries, scikit-learn was the library to call the parameters and to set the values accordingly. All the features of engineering methods followed these values.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Parameter names and associated values for the ML models.

https://doi.org/10.1371/journal.pone.0294803.t006

2.7.1 Accuracy.

This metric is the most widely used metric to measure the correctness of the ML model, that is, the proportion of accurately anticipated observations to all observations. Although accuracy is one of the mostly used performance metrics in the field of ML, two more metrics namely, F1 and ROC-AUC scores are also adopted by the researchers to evaluate the outcome of an algorithm [88]. The formula for accuracy is: (17) Here, TP and TN denote the number of positive and negative instances that are accurately classified in the multi-class classification of this study. In this type of classification, a threshold is set for each class and any prediction that is out of the threshold falls into a misclassification.

2.7.2 F1-score.

This metric displays the precision and recall weighted average and the F1-score is measured to balance recall and precision scores [89]. Moreover, the value range of the F1-score is between 0 to 1.0 which can be calculated using Eq (18).

(18)

Here,

2.7.3 ROC-AUC.

ROC-AUC is the value of the area under a Receiver Operating Characteristic (ROC) curve which is calculated using True Positive Rate (TPR) and False Positive Rate (TPR) obtained by Eqs (19–20) (19) (20) As opposed to a threshold-based metric like accuracy, the AUC value indicates a classifier’s performance by using ranking, and sometimes, it exhibits a better metric for evaluation [90]. For the multi-class classification task, the performance is measured using this formula [91]: (21) where, is the probability operator, i and j is a pair of classifiers, c is the number of classes. In general, the one vs. all approach can be utilized to plot N number of AUC curves for N number of classes in a multi-class classification problem [92].

Lastly, an overall workflow diagram is shown in Fig 2, and all the findings are exhibited and conscientiously analyzed in the following section.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Overall workflow diagram.

https://doi.org/10.1371/journal.pone.0294803.g002

3.1.1 Gender, age-groups, and ethnicity-based depression.

An in-depth analysis was performed to better understand the connection between depression severity and the different study groups. Here, all the illustrations have been shown using boxplots from the Python matplotlib library and seaborn package. A boxplot is a graphical method of showing the data distribution in terms of range, locality, and skewness by dividing the amount of data into four quartiles [93]. In the graph, the box depicts the data from the first quartile to the third quartile with a whisker (the line protruding from the sides of each box) pointing out the minimum and maximum values, if applicable.

Firstly, there were about twice as many data points for male participants as for female participants, as shown in Table 7. When it comes to depression levels, more than half of the data points from females gathered under the ‘Moderate’ and ‘Extremely Severe’ categories compared to that of the males, as shown by the median line in Fig 3. Overall, no depression was observed among about 44% of the accumulated participants while mild-to-moderate and severe-to-extremely-severe depression were observed among 25% and 31% of the population, respectively. For the age groups, it can be seen from Table 8 that there was a rise in the ‘Extremely Severe’ level for the participants between 21 to 25 years. The median line shown in Fig 4 also demonstrates the relatively high level of depression in this group compared to two other groups which are below 20 years and participants between 26-to-30-year. Lastly, though a large number of participants were inhabitants of Bangladesh, there were international participants whose depression severity can be observed in Tables 9 and 10. South Asian participants from countries other than Bangladesh, Pakistan, and Afghanistan. were in a state of moderate depression compared to mild depression of these countries’ participants. Notable mention here is that most participants from the Southeast Asian countries fell under the ‘Moderate’ to ‘Extremely Severe’ category and as a result, the boxplot (Fig 5) has been shifted up to those levels., The least amount of depression can be seen in the participants from Africa, more than two-thirds of whom were facing no depression.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Gender-based depression severity distribution.

https://doi.org/10.1371/journal.pone.0294803.g003

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Age-based depression severity distribution.

https://doi.org/10.1371/journal.pone.0294803.g004

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Ethnicity-based depression severity distribution.

https://doi.org/10.1371/journal.pone.0294803.g005

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. Result of the data analysis among males and females.

https://doi.org/10.1371/journal.pone.0294803.t007

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. Result of the data analysis among age groups.

https://doi.org/10.1371/journal.pone.0294803.t008

Download:

• PNG
  larger image
• TIFF
  original image

Table 9. Result of the data analysis among ethnicity groups (South Asia and Southeast Asia).

https://doi.org/10.1371/journal.pone.0294803.t009

Download:

• PNG
  larger image
• TIFF
  original image

Table 10. Result of the data analysis among ethnicity groups (Southeast Asia, Africa and Others).

https://doi.org/10.1371/journal.pone.0294803.t010

3.1.2 Academic level-based depression.

The prevalence of depression among the students of different academic levels ranging from first-year students to postgraduate was identified; the results are presented in Figs 6, 7 and Table 11. It was found that a randomly selected student in his/her 2^nd or 3^rd academic year has a larger probability of facing a higher severity of depression than a 1^st year or postgraduate student, as shown in Fig 6. Postgraduate students particularly experienced the least amount of depression during the hybrid mode of learning (median line at ‘Normal’), immediately followed by undergraduate-fourth year students (median line at ‘Mild’). This can be counted as a valid comparison as the percentages of the four academic level students excluding the first year were relatively adjacent to each other (Table 11). Meanwhile, one finding also suggests that if the students were satisfied with their examination results and other academic activities, they fell under the ‘Normal’ category more than the students who were dissatisfied. In fact, half of the students who were not enjoying the academic curriculum were at the ‘Moderate’ depression level contrary to half of the contented students who experienced no depression at all (Fig 7). This is a crucial finding since this study was intended to investigate the applicability of hybrid learning which should be able to cater for a broad student demographic.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Academic level-based depression severity distribution.

https://doi.org/10.1371/journal.pone.0294803.g006

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Satisfaction with academic study-based depression severity distribution.

https://doi.org/10.1371/journal.pone.0294803.g007

Download:

• PNG
  larger image
• TIFF
  original image

Table 11. Result of the data analysis among academic level groups.

https://doi.org/10.1371/journal.pone.0294803.t011

3.1.3 Residence and physical activity-based depression.

Depression was found to be more prevalent in big cities than in villages or small towns, which can be inferred from Table 12. Big city dwellers lived with a ‘Moderate’ median level of depression compared to the ‘Normal’ median level of the small town or village habitants, shown in Fig 8. This finding is backed up by internet connectivity status, empirically, which is more ubiquitous in the cities than in villages. Moreover, the respondents who were continuously connected to the internet were found to be slightly more depressed (Fig 9). The individuals who were suffering from lack of sleep were more depressed than those who were satisfied with the quality of their sleep. In fact, a gradual decline in depression can be seen among the respondents as their sleep satisfaction increased, as arrayed in Fig 10. Finally, those whose sleep duration was between 7–9 hours, were found to be experiencing less depression (median) than those who were oversleeping (more than 9 hours) or under sleeping (less than 7 hours), as shown in Fig 11.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Residence-based depression severity distribution.

https://doi.org/10.1371/journal.pone.0294803.g008

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Internet connectivity-based depression severity distribution.

https://doi.org/10.1371/journal.pone.0294803.g009

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. Sleep satisfaction-based depression severity distribution.

https://doi.org/10.1371/journal.pone.0294803.g010

Download:

• PNG
  larger image
• TIFF
  original image

Fig 11. Sleeping hours-based depression severity distribution.

https://doi.org/10.1371/journal.pone.0294803.g011

Download:

• PNG
  larger image
• TIFF
  original image

Table 12. Result of the data analysis based on residence types.

https://doi.org/10.1371/journal.pone.0294803.t012

3.1.4 Statistical summary.

The overall summary of the statistical data analysis is represented in Fig 12. The correlation between depression severity and the input variables is shown in terms of attribute importance by performing the Chi-squared test. The letters between the parenthesis designate the column in the dataset used in this study. Gender played a key role as a demographic variable when it comes to calculating depression severity scores. Age and residence are also seen to be affecting the severity score. However, the questionnaire items were found to be more crucial when determining a person’s depression severity.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 12. Comparison of accuracy using All features by performing (Chi-squared test).

https://doi.org/10.1371/journal.pone.0294803.g012

3.3.1 Comparison between Chi-square test and RFE.

From the feature selection results in Tables 14 and 15, SVM has achieved the highest score for both methods most of the time in terms of accuracy, F1-score, and AUC. A notable point here is that before tuning, SVM performed inadequately and came out in the last place among classifiers, which has been reversed after tuning. The only exception in which this classifier was outperformed by another classifier is in the Chi-square test by the LR model which achieved the highest accuracy of 88.74%. Having said that, KNN once again exhibited the lowest scores among all three metrics for both methods. Overall, the RFE method comparatively better assisted the ML models than the Chi-square test in this study’s findings as shown in Figs 15 and 16.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 15. Comparison of accuracy between Chi-square test and RFE.

https://doi.org/10.1371/journal.pone.0294803.g015

Download:

• PNG
  larger image
• TIFF
  original image

Fig 16. Comparison of ROC-AUC between Chi-square test and RFE.

https://doi.org/10.1371/journal.pone.0294803.g016

Download:

• PNG
  larger image
• TIFF
  original image

Table 14. ML models performance (Chi-square test).

https://doi.org/10.1371/journal.pone.0294803.t014

Download:

• PNG
  larger image
• TIFF
  original image

Table 15. ML models performance (RFE).

https://doi.org/10.1371/journal.pone.0294803.t015

3.3.2 Comparison between PCA and SparsePCA.

When the ML models were trained with the feature extraction PCA method, the LR model consistently performed better than other models by achieving 90.55%, 90.35%, and 0.9943 for accuracy, F1-score, and ROC-AUC score respectively (Tables 16 and 17). Nevertheless, the SparsePCA method experienced a different model, the CatBoost, to overperform and become the best performing model among all, which was followed by LGBM with a marginal gap. Lastly, KNN again became the most subdued performing classifier for both feature extraction methods. Hence, it can be concluded that SparsePCA showed better performance than the PCA method, and the performance comparison is also depicted in Figs 17 and 18.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 17. Comparison of accuracy between PCA and SparsePCA.

https://doi.org/10.1371/journal.pone.0294803.g017

Download:

• PNG
  larger image
• TIFF
  original image

Fig 18. Comparison of ROC-AUC between PCA and SparsePCA.

https://doi.org/10.1371/journal.pone.0294803.g018

Download:

• PNG
  larger image
• TIFF
  original image

Table 16. ML models performance (PCA).

https://doi.org/10.1371/journal.pone.0294803.t016

Download:

• PNG
  larger image
• TIFF
  original image

Table 17. ML models performance (SparsePCA).

https://doi.org/10.1371/journal.pone.0294803.t017

In addition to the chart comparison, the mean value of the confusion matrices generated from all the tuned ML models when incorporating the SparsePCA feature extraction method have been shown in Tables 18–20.

Download:

• PNG
  larger image
• TIFF
  original image

Table 18. The mean value of confusion matrix for GNB, LR, RF (SparsePCA).

https://doi.org/10.1371/journal.pone.0294803.t018

Download:

• PNG
  larger image
• TIFF
  original image

Table 19. The mean value of confusion matrix for SVM, LGBM, MLP (SparsePCA).

https://doi.org/10.1371/journal.pone.0294803.t019

Download:

• PNG
  larger image
• TIFF
  original image

Table 20. The mean value of confusion matrix for Catboost, KNN (SparsePCA).

https://doi.org/10.1371/journal.pone.0294803.t020