Electrochemical Biosensing and Deep Learning-Based Approaches in the Diagnosis of COVID-19: A Review

COVID-19 caused by the transmission of SARS-CoV-2 virus taking a huge toll on global health and caused life-threatening medical complications and elevated mortality rates, especially among older adults and people with existing morbidity. Current evidence suggests that the virus spreads primarily through respiratory droplets emitted by infected persons when breathing, coughing, sneezing, or speaking. These droplets can reach another person through their mouth, nose, or eyes, resulting in infection. The “gold standard” for clinical diagnosis of SARS-CoV-2 is the laboratory-based nucleic acid amplification test, which includes the reverse transcription-polymerase chain reaction (RT-PCR) test on nasopharyngeal swab samples. The main concerns with this type of test are the relatively high cost, long processing time, and considerable false-positive or false-negative results. Alternative approaches have been suggested to detect the SARS-CoV-2 virus so that those infected and the people they have been in contact with can be quickly isolated to break the transmission chains and hopefully, control the pandemic. These alternative approaches include electrochemical biosensing and deep learning. In this review, we discuss the current state-of-the-art technology used in both fields for public health surveillance of SARS-CoV-2 and present a comparison of both methods in terms of cost, sampling, timing, accuracy, instrument complexity, global accessibility, feasibility, and adaptability to mutations. Finally, we discuss the issues and potential future research approaches for detecting the SARS-CoV-2 virus utilizing electrochemical biosensing and deep learning.

complex sample matrix, they can be used to detect 84 viruses [17], cancers cells [18], bacteria [19], and small 85 biomolecules such as glucose, dopamine, uric acid, and ascor-86 bic [20], [21]. Electrochemical biosensors have been exten- 87 sively studied for their unique advantages, such as portability, 88 low cost, fast response, and high sensitivity, over other analyt- 89 ical devices [22]. The electrochemical biosensors produce a 90 signal by interacting with receptors/bioreceptors and the par-91 ticular analyte to produce or consume ions or electrons. This 92 causes a change in the electrical properties of the electrolyte 93 solution. The change in the electrical current or potential of 94 the electrolyte solution is measured using functionalized elec-95 trodes, which generate an electrical signal that is correlated 96 to the amount of target analyte in the test sample [23], [24]. 97 Many recent studies have shown that electrochemical biosen-98 sors can be utilized for SARS-CoV-2 diagnosis [25]. In gen-99 eral, deep-throat saliva and nasopharyngeal samples [26] can 100 be tested with or without the extraction of genetic material 101 from SARS-CoV-2, improving the time taken for rapid diag-102 nosis. For example, Raziq et al. vortexed the clinical sam-103 ples from nasopharyngeal in lysis buffer to release proteins 104 and reduce the inferencing species [14]. On the other hand, 105 Beduk et al. developed an electrochemical immunoassay for 106 the detection of SARS-CoV-2 using serum samples without 107 any pre-treatment [27]. 108 There is a high demand for deep learning-based approaches 109 in various research fields such as medical [28] and agri-110 culture [29]. Recently, deep-learning-based models (DLMs) 111 were extensively studied for the diagnosis of the SARS-CoV-112 2 [30], [31], [32], [33]. The developed DLMs consist of 113 data collection, data preparation, feature extraction, and lastly 114 model evaluation [34]. Data collection is the first and crucial 115 step. The quantity and quality of the computed tomography 116 (CT) and X-ray lung images collected are used to validate the 117 success of the developed model. The data preparation stage 118 mostly includes data augmentation, noise removal, and resiz- 119 ing the input image [35]. The processed data are then divided 120 into training, test, and validation sets. The model is developed 121 using the training dataset, and its optimization is checked gen-122 erally using a cross-validation technique [36]. The optimized 123 model is then run on the test set to validate its performance on 124 the unseen data. Feature extraction is the process of reducing 125 the dimensionality in which the initial raw data are processed 126 to more manageable groups by maintaining accuracy and 127 still describing the original dataset. Finally, the developed 128 model is evaluated using various metrics, such as accuracy, 129 confusing matrix, sensitivity, specificity, precision, F1-score, 130 etc. [37]. 131 X-ray and CT are the most widely used imaging modali-132 ties in the field of artificial intelligence (AI) for the accurate 133 diagnosis of SARS-CoV-2 [38]. The manual interpretation of 134 medical images by radiologists is a time-consuming process 135 and it is prone to human errors and bias. Recently, AI technol-136 ogy is evolving in the medical diagnosis of various diseases. 137 Deep learning [15], machine learning [39], data science [16], 138 the internet of things [40], and big data [41] are the main 139 FIGURE 1. Schematic illustration of SARS-CoV-2 structure and its mode of host entry. Adapted with permission from Elsevier, Copyright (2022), 5234941355836 [5].  in biomedical applications by achieving human-level accu-150 racies [42] or even beyond in some cases [43]. 151 This review aims to demonstrate the use of electrochemical limitations of these two techniques are discussed along with 157 ways to minimize the limitations and increase the use of elec-158 trochemical sensors to combat with SARS-CoV-2. Finally, 159 the future research directions in both fields as potential tools 160 to reduce the dependency of RT-PCR tests and help minimize 161 the severity of the pandemic are discussed. Influenza (flu) and Coronaviruses are typically identified by 166 examining their genomes, particularly their ribonucleic acid 167 (RNA) sequences [44]. The PCR process multiplies DNA 168 sequences in a quick period, considerably improves the capa-169 bility of infectious diseases diagnosis. Variants of PCR tech-170 niques, such as end-point PCR, quantitative PCR (qPCR), 171 digital PCR have been developed and employed in diagnos-172 tics since its development [45]. A real-time reverse transcrip-173 tion step precedes the PCR (RT-PCR) for coronaviruses, as it 174 does for other RNA viruses, and transcribes the RNA into 175 cDNA. Due to its sensitivity and specificity, the PCR test 176 has become the method of choice to detect SARS-CoV-2. 177 It is theoretically capable of identifying a single copy 178 of virus, resulting in a shorter diagnostic window than 179 immunoassays.

180
Reliable laboratory diagnosis is vital to slow down of the 181 respiratory diseases. When the COVID-19 pandemic broke 182 out, RT-PCR diagnostics were the first to be developed and 183 widely used. RT-PCR is commonly used to discover causal 184 viruses from respiratory secretions in cases of acute respi-185 ratory illness [46]   To address the need for better sampling methods, noninva-244 sive ways of screening may be an ideal approach for the iden-245 tification of biomarkers in body fluids, including urine, saliva, 246 tears, sweat, and breath for the screening of SARS-CoV-2. 247 For example, Alefeef et al. recently reported a paper-based 248 electrochemical sensor to detect SARS-CoV-2 [57]. In their 249 design ( Figure 3A), gold nanoparticles (AuNPs) were capped 250 with highly specific antisense oligonucleotides (ssDNA) 251 to target viral nucleocapsid phosphoprotein (N-gene). The 252 issue with this prototype is that it requires RNA isolation 253 from SARS-CoV-2, which makes it unsuitable for on-site 254 detection. Yakoh et al. introduced a paper-based electro-255 chemical biosensor using spike protein receptor-binding 256 domain (SP RBD) of SARS-CoV-2 as a recognition group 257 ( Figure 3B) [58]. After immobilization of the SP RBD on the 258 electrode surface, the square-wave voltammetry (SWV) tech-259 nique was utilized for the detection of SARS-CoV-2. Unfor-260 tunately, the detection limit of this electrochemical sensor 261 prevented the detection of SARS-CoV-2 in the actual nasal 262 swab specimens.

263
One of the most promising studies on electrochemical 264 detection of SARS-CoV-2 was carried out in [17]. In their 265 study, the silicon dioxide layer was first placed on a silicon 266 wafer, followed by 25-nm-thick thermally deposited titanium 267 layers, and finally, a 350-nm-thick gold layer deposited via 268 electron-beam assisted gold evaporation and patterned with 269 photolithography to fabricate a platform (a chip). A redox 270 probe (ferrocene) modified DNA was then attached to the 271 chip, followed by the antiSARS-CoV-2 spike S1 antibody 272 linked to the amine-terminated DNA. The as-fabricated chip 273 was tested using chronoamperometry and the results were 274 obtained in minutes. This is the first study that used undiluted 275 saliva samples for the detection of SARS-CoV-2.

276
Besides its short analysis time and easy sampling method, 277 the biosensors have about nine months of shelf-life [59]. 278 In another study, Seo et al. reported a field-effect tran-279 sistor (FET)-based electrochemical biosensor to detect 280 SARS-CoV-2 in clinical samples without requiring sample 281 pretreatment or labeling [60]. As shown in Figure 3C, the FET 282 was coated with a graphene layer and modified with a specific 283 antibody against SARS-CoV-2 spike protein. The biosensor 284 was very sensitive to SARS-CoV-2 antigen protein and could 285 distinguish the virus from the MERS-CoV antigen protein. 286 Raziq et al [14] used a gold-based thin-film electrode as 287 a disposable sensor chip modified by SARS-CoV-2 nucle-288 oprotein (ncovNP) molecularly imprinted polymer (MIP) 289 to form an artificial receptor for the detection of ncovNP 290 ( Figure 3D). The sensor was designed to detect ncovNP 291 which shows a linear response of up to 111 fM with 292 detection and quantification limits of 15 fM. A portable 293 potentiostat was utilized to test the as-fabricated sensor 294 with nasopharyngeal swab samples of COVID-19 positive 295 patients. Although the swab samples had to be vortexed for 296 30 min in a lysis buffer to release the viral protein before 297 each test, the MIP technology is still very attractive due to 298 its rapid, low-cost and sensitive detection capabilities, and 299    Radiologists use images from CT and X-ray scans to diag-311 nose COVID-19 as shown in Figure 4 [63]. X-ray is an inex-312 pensive imaging technique and it poses a low-risk radiation 313 hazard to human health [64]. However, it may be difficult to 314 diagnose the stage of infection just by looking at the X-ray 315 scans. This is due to the similarity of white spots, which may 316 consist of water and pus that are associated with other lung 317 diseases, such as tuberculosis. On the other hand, CT scans 318 offer more precise detection but are more expensive than 319 X-ray imaging [65], [66]. However, the detection accuracy 320 from CT scans is still not satisfactory and defective in the 321 diagnosis of COVID-19. Other techniques, in addition to the 322 CT scan, can help improve the accuracy of the COVID-19 323 diagnosis [67]. Between these two imaging modalities, X-ray 324 imaging is usually preferred since X-ray imaging poses less 325 radiation, is cheaper and more accessible than CT imaging in 326 hospitals.

327
In this section, only the state-of-the-art DLMs using 328 CT and X-ray imaging modality for COVID-19 diagno-329 sis are reviewed. The use of deep learning to assist the 330 diagnosis stage of COVID-19 is divided into three main 331 tasks, namely, classification, detection, and segmentation. 332    [74]. In their model, a heat map of X-ray images 363 was produced and interpreted by radiologists, and it was con-364 cluded that their model could assist radiologists and reduce 365 the clinical workload in hospitals. They had used a limited 366 number of datasets at the time of publication when developing 367 the model, hence the model would require a larger dataset 368 for future use. The researchers effort to limit the severity 369 of the pandemic are affected by the lack of datasets when 370 developing deep learning-based models [75]. The generative 371 adversarial network (GAN) has helped to generate synthetic 372 challenged the accuracy of the deep-learning models in the diagnosis of COVID-19 [76]. Sethy [88]. In their model, around 99% accuracy 433 was achieved over 7996 test images which was the largest 434 dataset used for COVID-19 diagnosis. As the use of transfer 435 learning reduces the chance of the model to overfit, He devel-436 oped a self-trans-based approach in which the contrastive 437 self-supervised learning with transfer learning was integrated 438 synergistically to learn unbiased features which achieved 439 86% accuracy with DenseNet-169 using CT images [89]. 440 The most popular Deep learning architectures for 441 COVID-19 detection are shown in Table 2 The early and fast diagnosis of COVID-19 is crucial to com-451 bat the rapid spread of COVID-19 globally. In this section, the 452 state-of-the-art in deep learning and electrochemical biosen-453 sors will be discussed in terms of cost, sampling, timing, 454 accuracy, instrument complexity, global accessibility, feasi-455 bility, and adaptability to mutations.

457
A PCR test kit consists of a combination of chemicals, nucleic 458 acid extraction kits, and other elements. A PCR test usu-459 ally costs about $60 for patients [90]. This cost can vary 460 greatly from country to country depending on their econ-461 omy, resources, and capabilities. Furthermore, due to the 462 COVID-19 pandemic, many countries face a supply shortage 463 of these kits and are unable to acquire them from the manu-464 facturing countries [91]. 465 As for using deep learning models, the cost of developing 466 these models for COVID-19 diagnosis comes from the pur-467 chase of a GPU-powered PC and the data collection process. 468 The initial development cost varies based on the training size, 469 model size, and training volume. Once the model is developed 470 and ready-to-deploy, there are no extra costs to the hospitals 471 or clinics to maintain the system. The variable cost to the 472 patient is the CT and X-ray scans. The costs of CT and X-ray 473 scans vary based on the geographical location of the hospital. 474 For example, in the USA alone, a chest CT scan can range 475 from 1, 072upto3,509 while an X-ray scan can range from 476 82 to 417, depending on the hospital location [92]. It has to 477 be noted that a CT machine is not often available in small or 478 rural hospitals, as compared to a medical X-ray machine.

479
Despite numerous advancements in biosensor technology, 480 glucose biosensor is considered the first biosensor and con-481 tinues to dominate and accounts for roughly 85% of a $5 482 VOLUME 10, 2022 billion market [93]. Due to an existing market in the area of  During the scan, the exam table will shift, causing the X-ray 520 beam to generate a series of images from various angles.  During an X-ray scan, small amounts of radiation are used 529 to produce images of the body's organs, tissues, and bones.

530
It can detect abnormalities in the airways, blood vessels, 531 and lungs when focused on the chest. Getting a chest X-ray 532 does not require much planning on the part of the patient. 533 The X-ray is performed in a special room equipped with a 534 movable X-ray camera mounted on a long metal arm, where 535 the patient is positioned next to a ''plate.'' This plate would 536 contain an X-ray film or a special sensor that captures the 537 images and saves them to a computer to be analyzed by the 538 radiologist.

539
On the other hand, nasopharyngeal, and oropharyn-540 geal swabs are routinely taken and studied to detect the 541 SARS-CoV-2 virus. Urine, feces, sputum, plasma serum, and 542 whole blood are also studied, but not as widely due to the 543 difficulty of sampling for both patients and health care work-544 ers [96]. In most studies, the sampling principle relies on the 545 extraction of viral RNA from the samples. This is considered 546 a more reliable approach compared to CT and X-ray imaging 547 because viral RNA can be detected 2-3 days before symp-548 toms appear and can last for up to 25-50 days depending on 549 the severity of the disease [97]. induced lung cell dysfunction and by tracing ROS in spu-560 tum samples, SARS-CoV-2 can be detected with more than 561 97% accuracy. PCR tests samples are usually taken from 562 Nasopharyngeal, and oropharyngeal swabs, which may cause 563 discomfort to the patient, as well as suffer from false posi-564 tives and false negatives. In DLMs, the patients are exposed 565 to radiations during CT-scans or X-rays. However, electro-566 chemical sensors can be designed for nasopharyngeal, and 567 oropharyngeal swabs urine, feces, and whole blood. These 568 sensors are much more flexible and harmless, compared to 569 PCR and DLMs.  DLMs are run through a host computer that receives images 642 from either CT or X-ray machines. Hence, the models devel-643 oped by deep learning techniques do not add any complex-644 ity to the existing systems. However, the imaging modality 645 used within the system is a significant parameter that has 646 a direct impact on the complexity of the whole diagnosis 647 procedure. An X-ray machine mainly consists of an X-ray 648 generator and an image detector. The main parts of the X-ray 649 generator are tube, high voltage generator, control console, 650 and the cooling system. A CT scanner mainly consists of four 651 main components. gantry (frame) houses the X-ray source, 652 detectors, patient port (a large opening in the middle), subject 653 table, and a computer system that gathers all data from the 654 detectors. Finally images are produced based on the captured 655 data [102]. A CT scanner must move around the patient being 656 scanned; hence, an X-ray equipment is much smaller and less 657 complicated than a CT machine.

658
Instruments used for electrochemical biosensors are less 659 complex with very low cost compared to CT and X-ray instru-660 ments which costs between 15, 000−90,000. For accurate 661 electrochemical biosensing of SARS-CoV-2, the potentio-662 stat is required to ensure signal processing and cell con-663 ditioning. Potentiostat devices can be either bench-top or 664 portable. Portable devices (miniaturized potentiostat) are usu-665 ally equipped with a portable mobile device for read-out 666 which is a great feature to monitor public health surveillance 667 of SARS-CoV-2 on-site [95], [96]. On the other hand, bench-668 top models are not that user-friendly and require skilled per-669 sonnel to operate.

671
To date, there is no U.S. Food and Drug Administration 672 (FDA)-approved system using deep-learning techniques with 673 CT or X-ray imaging modality to diagnose COVID-19. Apart 674 from FDA, the employment of DLM for COVID-19 diagnosis 675 also depends on the acceptance of radiologists and clinicians. 676 However, there are already FDA-approved software using 677 DLM, such as the OsteoDetect to analyze X-ray images for 678 wrist fracture [103], which demonstrates the potential reli-679 ability and feasibility of DLM for other types of medical 680 devices, such as COVID-19 diagnosis.

681
Global accessibility and feasibility of the DLMs will 682 depend on access to good imaging facilities in the hospitals. 683 Therefore, access to CT and X-ray machines is one of the 684 important concerns when it comes to the feasibility of using 685 the deep learning models in hospitals. As COVID-19 is 686 The SARS-CoV-2 genome structure was sequenced for the 739 first time in Wuhan, China in January 2020. Understanding 740 SARS-CoV-2 genome sequencing is important to interpret 741 the virus's nature and mutation rate, as well as successful 742 prevention strategies such as vaccines and drugs. Several 743 investigations have noted the new coronavirus's genetic 744 diversity and rapid evolution. Some mutations may have little 745 or no impact, whereas others may affect the virus's properties, 746 such as increased transmissibility. The significant mutations 747 of Sars-CoV-2 presented in Figure 5 indicate the origin, date 748 of the first detection, and the main concerns of the variants of 749 SARS-CoV-2.

750
The procedure for collection, preservation, storage and 751 processing of the samples affects the accuracy of RT-PCR. 752 The lack of proofreading ability in the viral RNA polymerases 753 results in a high rate of mutation. As a consequence, if the 754 virus mutates in the targeting genomic region, which happens 755 often, the precision of these diagnostic methods is adversely 756 . However, more attention is required in 774 the area of diagnosis considering the high mutation rate of 775 the SARS-CoV-2. Apart from developing DLMs to iden-776 tify the different variants, estimating the mutation rate using 777 DLMs has also drawn great attention from researchers world-778 wide [114]. This is because by knowing the mutation rate, 779 scientists will be able to illustrate the risk of emergent 780 SARS-CoV-2 infection [115]. DLMs can be developed to 781 detect the mutations and are adaptable to new variants of 782 SARS-CoV-2 as long as the datasets are accessible [116]. 783 Since the development of the DLMs relies on the dataset 784 used during the training procedure for the diagnosis of 785 SARS-CoV-2, the developed models can easily be adapted 786 to new variants of SARS-CoV-2. This procedure requires an 787 update on the datasets used during the training process of the 788 deep learning models.

789
A comparison of PCR, electrochemical biosensors, and 790 deep-learning-based COVID-19 diagnostic tests are provided 791 in Table 3 in terms of cost, sampling, the time required for 792 diagnosis, accuracy, instrument complexity, global accessi-793 bility and feasibility, and adaptability to mutations. It is noted 794 that the cost is given in USD throughout this study.   would be a tremendous step forward [118], [119], [120]. 818 Since breath analysis has the added advantage of real-time 819 and point-of-care analysis [121], there has been strong inter-   University of Social Sciences, Singapore. He has 1420 published more than 550 papers in refereed inter-1421 national SCI-IF journals (500), international conference proceedings (42), 1422 books (17) with more than 50,000 citations in Google Scholar (with an 1423 H-index of 113). In addition, he has worked on various funded projects with 1424 grants worth more than six million SGD. According to the Essential Science 1425 Indicators of Thomson, he has been ranked in the top 1% of the Highly 1426 Cited Researchers for the last six consecutive years (2016-2021) in computer 1427 science. His major research interests include biomedical signal processing, 1428 biomedical imaging, artificial intelligence, visualization, and biophysics for 1429 better healthcare design, delivery, and therapy. He is on the editorial board 1430 of many journals and has served as a guest editor for many journals.