1 Introduction

The COVID-19 pandemic caused an unprecedented global effect on private and public healthcare systems and social and economic impacts. Among the COVID-19 containment measures is vaccination. However, the distribution of COVID-19 vaccines continues to be applied with low vaccination coverage in underdeveloped countries, especially adapted vaccines to prevalent COVID-19 variants. Therefore, mass testing for COVID-19 is still a strategic preventive measure to reduce its transmission.

There is a need for new techniques developed to prevent it from mass propagating again and also to prevent possible pandemic scenarios. Locations such as airports, schools, workplaces, and remote communities would benefit from faster and reagent-free screening techniques.

Vibrational spectroscopy is considered a promising method for biological analyses. It enables label-free extraction of biochemical information and images toward diagnosis and the assessment of the biochemical and molecular composition of biofluids [1]. Vibrational spectroscopy techniques are bioanalytical approaches with great potential to discriminate between healthy and pathological conditions, like Diabetes mellitus [4] and Zika virus [15]. In addition, it allows the analysis of different bio-fluids, such as blood, urine, saliva, and tissues [2, 14, 18].

Among several vibrational spectroscopy technologies, Raman is known by its ability to determine molecular components of various types of materials quickly and without the need for the use of reagents. It uses inelastic light scattering based on a monochromatic laser source. The changes in the spectra represent parts of specific molecules. Each Raman shift is unique in each wavenumber frequency, representing a specific functional group [8].

The application of biofluids in Raman spectroscopy for diagnostic or screening fields permits simple sample preparation without the insertion of reagents and thus can be considered a green technology due to a significant reduction in environmental waste [8]. Furthermore, its usage has been investigated in the diagnosis of several diseases like breast, colon, lung, and prostate cancers [9].

In this paper, we are interested in the analysis of Raman spectra obtained from blood serum samples for the detection of COVID-19. The ongoing COVID-19 pandemic caused an unprecedented overburning of healthcare systems and still represents a global health issue with the emergence of COVID-19 variants.

Due to the relevance of the topic, there are already some related works in the literature. In [18] the authors presented a support vector machine (SVM) model to detect COVID-19 using Raman spectrum of serum samples, and in [19] the authors extended the work of [18] by combining the SVM with an extreme gradient boosting (XGB) technique. Despite the Raman spectroscopy data used in [18, 19] being the same considered in our research, our work goes a step further by investigating deep learning techniques, like convolutional neural networks (CNNs).

Different from most Raman spectra analysis, which adopts simple statistical analysis [6], and also of some machine learning initiatives with Raman, which are focused on traditional classifiers [18, 19], we hypothesize that convolutional neural networks [2] can contribute in the detection of COVID-19 from Raman spectra of serum samples by efficiently learning robust representations of such a problem and also improve model explainability. Despite some few works have already applied 1D CNN to classify breast cancer, minerals, and pancreatic cancer by Raman spectroscopy, e.g., [11, 13, 16], the investigation of such architectures for COVID-19 detection from Raman spectra is yet a barely explored topic in the literature.

Therefore, the work most related to ours in the literature is the study presented in [5], which investigated traditional and deep learning techniques for the diagnosis of COVID-19 using Raman spectroscopy. In that work, the authors evaluated SVM, XGB, Random Forest (RF), and CNN, which achieved good predictive performance over several groups of experiments. However, our work differs from this for some reasons, such as: we consider Raman spectra of serum samples instead of saliva ones; we consider a publicly available Raman data set instead of a private one, which benefits reproducibility aspects; and besides the predictive performance, we also are interested in the explainability analysis.

The main obstacle to differentiate each class based on the Raman spectrum from control and COVID-19 samples is the similarity of both spectra. The complexity of the dataset is a critical feature to solve issues for discriminating samples of each class. Truncation in regions with organic compounds can be performed at specific wavelength regions. The Raman dataset analyses and interpretation require expertise due to the potential overlap of peaks corresponding to distinct chemical components. The Raman spectra differentiation can be especially problematic in complex biofluids with thousands of molecular components, featuring variability which is an issue for discrimination. This spectral variability can impact the accuracy and discrimination associated with high-dimensional data, with a large number of wavelengths contributing to the spectrum. The management of high-dimensional data is a challenge for computational complexity, feature selection, and generalization of the classification models for spectral datasets.

This paper aims to contribute to the development and evaluation of CNNs for the detection of COVID-19 using Raman spectra of serum samples, in order to obtain more accurate models. To be specific, the study presents a one-dimensional (1D) CNN architecture to process Raman spectra as well as an explainability analysis of its salient features. In summary, the main contributions of our work are listed as follows:

  • Development and evaluation of the proposed 1D CNN method for detection of COVID-19 from Raman spectra of serum samples;

  • Comparative evaluation of several traditional and deep learning techniques for the problem;

  • Explainability analysis regarding salient Raman shifts to detect and discriminate COVID-19 and healthy groups.

The remainder of the paper is organized as follows. Section 2 presents the Raman spectroscopy data as well as the steps of data preparation and preprocessing. Section 3 describes the 1D CNN proposed in this paper. Section 4 shows the experimental results and explainability analysis regarding our CNN model. Section 5 concludes the paper.

2 Data Set Description and Preparation

The data adopted in this work was publicly available by [18]. Blood samples were collected between February 10th and May 10th, 2020. Blood samples were taken from COVID-19 patients and suspected cases upon admission between February 10 and May 10, 2020. The serum was isolated by centrifuging at 3000 rpm for 10 min after a one-hour rest. The serum samples were stored at 4\(^\circ \)C and measured within 36 h. For measurement, approximately 0.5 ml of serum sample was prepared in cryopreservation tubes and strictly sealed for Raman scan. Additional spectral data were collected from cryopreservation tubes with saline solution [18].

The data were obtained from blood serum samples in a total of 309 spectra, of which 159 received a positive diagnosis for COVID-19 (COVID-19 group) and 150 received a negative test result for COVID-19 (Healthy group). The serum samples were processed by a Raman spectrometer. For each sample, the equipment is responsible for generating a corresponding spectrum.

Figure 1 shows the average spectrum of the Healthy group (a), the average spectrum of the COVID-19 group (b), the average spectrum of both groups (c) and all spectra from both groups (d). One can see some interesting regions in Figs. 1(c) and 1(d), in which samples from different groups seem better separated.

Fig. 1.
figure 1

(a) Average of healthy group spectra, (b) Average of COVID-19 group spectra, (c) Average of healthy and COVID-19 spectra, (d) Total COVID-19 and healthy spectral plot

Full size image

Regarding data preparation and preprocessing, two steps were conducted in this study: Savitzky-Golay filtering and vector normalization:

  • The Savitzky-Golay (SG) is a filter for smoothing and differentiation that optimally fits a set of data points to a polynomial in the least-squares sense [17]. This process is important to reduce noise and smooth the signal while maintaining higher-order moments of the original spectrum. SG has been used to normalize and preprocess spectral data for FTIR, Raman, and other spectroscopy equipments due to the inherent presence of noises during sample collection, manipulation, and so on.

  • Vector normalization is the process in which the intensity values of all spectra are normalized by the Euclidean norm.

3 Model Description

This section describes the one-dimensional CNN proposed in this study for the detection of COVID-19 using Raman spectra of serum samples. The proposed architecture is illustrated by Fig. 2. There are three main types of layers: convolutional, pooling, and fully connected. In the following, each one of them is briefly described.

Fig. 2.
figure 2

The architecture of the one-dimensional CNN proposed.

Full size image

Convolutional Layer. The convolutional layer aims to learn representative features of the input data. The convolution layer is made up of several convolution kernels, also known as filters, which are used to compute different feature maps.

Specifically, each neuron in a feature map is connected to a region of neighboring neurons in the previous layer. Such a neighborhood is referred to as the receptive field of the neuron.

Mathematically, the feature value at location (i, j) referring to the kth feature map of the lth layer, \(z^l_{i,j,k}\) is calculated as:

$$\begin{aligned} z^l_i = \sigma \left( w^{l^T}_{k} x^l_{i,j} + b^l_k \right) ~, \end{aligned}$$
(1)

in which \(x^l_{i,j}\) is the input centered at position (i, j) referring to the l-th layer, \(w^{l^T}_{k}\) and \(b^l_k\) are respectively the weight vector and the bias term of the k-th filter of the l-th layer, and \(\sigma \) is the activation function which in CNN architecture is usually ReLU (Rectified Linear Units):

$$\begin{aligned} ReLU(x) = max\{0, x\}~. \end{aligned}$$
(2)

To compute a feature map, the process can be divided into two steps: first, the input is convoluted with an already learned filter, and then a non-linear activation function is applied to the convolution results. This non-linearity is not an integral part of the convolutional layer, but in practice, almost all convolutional layers are followed by a non-linearity. These non-linearities are desirable, as they allow the detection of non-linear features [7].

Pooling Layer. The pooling layer aims to achieve displacement invariance, reducing the spatial resolution of feature maps. This spatial reduction is achieved through a non-unitary stride. In this way, only one activation at each n (stride size) position of the input is calculated in each spatial dimension.

Each feature map of a pooling layer is connected to the corresponding feature map of the previous convolutional layer. The output generated by the pooling function for the feature map \(z^l_{i,j,k}\) is represented by the expression:

$$\begin{aligned} p_{m,k} = \underset{n=1}{\overset{N}{\max }} \left( z^{l}_{i-1} \times q + n \right) ~, \end{aligned}$$
(3)

in which N is the cluster size; and q is the slide size, which determines the degree to which the adjacent pool windows overlap.

Fully Connected Layer and Dropout. Fully connected layers are usually located at the end of the network, as shown by our proposed architecture in Fig. 2. In these layers, the features extracted in the previous convolution layers are used to obtain the network classification output.

A fully connected layer takes all the neurons in the previous layer and connects them to every neuron in the current layer. It also establishes which high-level attributes most closely relate to the object’s class.

Dropout is a technique usually adopted in this layer, in which some neurons are randomly turned off, along with their connections, during training only. During test, all neurons are kept active. The reason for doing this is to avoid overfitting in training. In our architecture, we adopted a dropout rate of 20% between the dense layers.

Finally, after the fully connected layers and the dropouts come to the output layer. For the classification problem, it is common to use as many neurons as there are classes to predict and the output has a sigmoid activation function that can show the result of the classification. The sigmoid is represented in the following equation:

$$\begin{aligned} \sigma (x) = \dfrac{1}{1+e^{-x}}~. \end{aligned}$$
(4)

This function only returns values between 0 and 1. It is responsible to classify the output and is located in the final step of our model shown in Fig. 2.

The CNN architecture shown by Fig. 2 receives a Raman spectrum as input. It then extracts features of the spectrum through three one-dimensional convolutional layers. At the end of each convolutional layer, ReLU is adopted as the activation function. Between each convolutional layer, there is a MaxPooling layer. By stacking multiple convolutional and pooling layers, one can gradually extract features with a higher (complex) level of abstraction. After extracting the features of the input, the Flattening step transforms the arrays into an one-dimensional vector to start the classification method. The Dense layers, which are also composed by Dropout functions, learn the weights to accurately adjust the classification of each group, and the Sigmoid function is used to return the prediction.

4 Experimental Results

In this section, we present the experiments and results obtained by our CNN in comparison with other techniques in the problem of COVID-19 detection from Raman spectra. In addition, we also present an explainability analysis regarding the salient features considered by CNN in its classification using Shapley additive explanations (SHAP) [3].

The experiments were conducted using a k-fold cross-validation process. This method partition the dataset samples in k disjoint subsets, with k-1 adopted as a training set and the remaining one as a test set. A total of k executions are performed, which means each subset is adopted as the test set once. The predictive performance of the techniques is then obtained over k executions in which performance metrics like accuracy, sensitivity, and specificity are calculated. In this study, we considered 5-fold cross-validation.

We compared our CNN technique against widely known supervised learning techniques like Naive Bayes (NB), Random Forest (RF), Extreme Gradient Boosting (XGB), Multi-Layer Perceptron (MLP), and Support Vector Machine (SVM) as well as against other deep learning techniques like Fully Convolutional Network (FCN), Multi-Channel Deep Convolutional Neural Network (MCDCNN) and Residual Networks (ResNet). All models were trained using the parameters recommended in [10, 19].

Regarding the computational simulations, the experiments were done using Python on two machines: a laptop with Core i7 9th Gen processor, 32 GB of Ram, and Geforce GTX 1660Ti GPU, and a desktop with Ryzen 9 processor, 32 GB of Ram, and Titan V GPU.

4.1 Predictive Results for COVID-19 Detection

Table 1 shows the predictive results of the techniques under analysis in terms of accuracy, sensitivity, and specificity. As sensitivity and specificity covers different groups, in order to support our analysis we also evaluate the arithmetic and harmonic mean between both metrics, respectively named Mean(Se, Sp) and F1(Se, Sp). The best result obtained by each metric is boldfaced in the table. Among traditional techniques, RF, SVM and MLP achieved the best results. Indeed, SVM and MLP have respectively the highest sensitivity and specificity in the table. However, the best overall results were achieved by the CNN which classified samples correctly more than any other technique. On the other hand, worse results were obtained by NB and FCN. While NB is the unique linear technique under analysis, FCN is a deep learning technique essentially based on the convolutional layer, which means it does not have fully connected layers to help in the learning of complex mapping functions.

Table 1. Predictive performance of the techniques under analysis for the COVID-19 detection from Raman spectra.
Full size table
Fig. 3.
figure 3

Analysis of the predictive performance of the techniques under comparison over the 5-fold cross-validation.

Full size image

Now we analyze each one of the executions of the 5-fold cross-validation in order to better understand the performance of the techniques under analysis. Figure 3 shows the predictive performance of each model in terms of accuracy, specificity, sensitivity, and Mean(Se, Sp). One can see FCN really had troubles in some executions. One can also see that ResNet, a state-of-the-art deep learning technique, had its performance oscillating among the executions. For example, in Fig. 3 ResNet achieved the worst (execution 1) and the best predictive result (execution 3), indicating that the model may be overfitted in some scenarios, which will be subject of our further investigations. Regarding the CNN performance, it was very stable in terms of accuracy, specificity, and Mean(Se, Sp), indicating that such a technique can contribute in the detection of COVID-19 from serum Raman spectroscopy samples. In order to better understand the salient features that contributed to the almost 97% of accuracy, we next investigate the explainability of the CNN model.

Fig. 4.
figure 4

Explainability analysis of the CNN model considering the 5-fold cross-validation for COVID-19 detection.

Full size image

4.2 Explainability Analysis

The SHAP method is an approach to explain the output of machine learning models. It connects optimal credit allocation with local explanations using the classic Shapley values from game theory and their related extensions [12]. With that in mind, to understand a bit more about how the CNN model created its inductive mechanism for COVID-19 detection, some SHAP analysis were conducted in order to provide insights regarding the CNN most relevant features.

Figure 4 shows the SHAP features of the main Raman shift capable to discriminate both groups of samples. The main Raman shifts between 1605–1607 cm\(^{-1}\) can be attributed to nucleoproteins. The Raman shifts between 1448–1450 cm\(^{-1}\) and 1271–1273 cm\(^{-1}\) can be attributed to (CH2/CH3) from lipids and the proline from collagen, respectively. Besides, the Raman shift at 831 cm\(^{-1}\) can be attributed to the amino acid Tyrosine. Furthermore, the Raman shifts between 427–432 cm\(^{-1}\) and 402-410 cm\(^{-1}\) were attributed to Phosphate and the heme deformation of proteins. Lastly, the Raman shift at 1657 cm\(^{-1}\) can be attributed to Amide I in proteins [19] . In general, the main Raman shifts used to discriminate control from COVID-19 samples were related to proteins, indicating changes in the proteomic profile in the serum of COVID-19-infected subjects.

5 Conclusions

This paper investigated deep learning solutions based on CNNs for the detection of COVID-19 using Raman spectra of serum samples in order to obtain more accurate models. To be specific, an 1D CNN was proposed and comparatively evaluated against several traditional and deep learning techniques.

Experimental results showed that the CNN model achieved the highest predictive performance in comparison with other techniques, with 96.8% accuracy, 95.5% sensitivity, and 98.2% of specificity, outperforming even other robust deep learning solutions like ResNet.

In addition, explainability analysis regarding salient Raman shifts to detect COVID-19 and healthy groups were also conducted, revealing some proteins which probably are associated with the spectral differences between the groups.

The investigation conducted here contributes to the Raman spectroscopy field and may also impact the society in the future by providing regular, and financially accessible diagnosis for the population.

The fast and accurate results achieved by the CNN method proposed can make the system suitable for use in screening and routine examinations, for example, contributing to the promotion and regular testing of the population.

Forthcoming works will focuses on the development of even better deep learning architectures as well as in addressing other applications based on spectroscopy technologies.