Selective Electrochemical Detection of SARS-CoV-2 Using Deep Learning

Abstract

COVID-19 has been in the headlines for the past two years. Diagnosing this infection with minimal false rates is still an issue even with the advent of multiple rapid antigen tests. Enormous data are being collected every day that could provide insight into reducing the false diagnosis. Machine learning (ML) and deep learning (DL) could be the way forward to process these data and reduce the false diagnosis rates. In this study, ML and DL approaches have been applied to the data set collected using an ultra-fast COVID-19 diagnostic sensor (UFC-19). The ability of ML and DL to specifically detect SARS-CoV-2 signals against SARS-CoV, MERS-CoV, Human CoV, and Influenza was investigated. UFC-19 is an electrochemical sensor that was used to test these virus samples and the obtained current response dataset was used to diagnose SARS-CoV-2 using different algorithms. Our results indicate that the convolution neural networks algorithm could diagnose SARS-CoV-2 samples with a sensitivity of 96.15%, specificity of 98.17%, and accuracy of 97.20%. Combining this DL model with the existing UFC-19 could selectively identify SARS-CoV-2 presence within two minutes.

Keywords: COVID deep learning; COVID-19 diagnosis; differential diagnosis; electrochemical SARS-CoV-2 detection; electrochemical biosensor.

Figure 1

Example current response of a SARS-CoV-2 positive sample with 0.1 cp/mL concentration, baseline (true negative), and sample–baseline difference. The SARS-CoV-2 sample current response is higher than the baseline sample.

Figure 2

Univariate feature scores of the futures. The bar chart shows the informative and non-informative features. The most informative features are F0, F6, and F3 and the least informative features are F5, F11, and F14.

Figure 3

(a) Signature analysis of 400 SARS-CoV-2 positive samples. All current response difference points for 400 samples are above the threshold line. Meaning all samples were diagnosed correctly as SARS-CoV-2 positive. (b) Signature analysis of 400 blank samples. All current response difference points for 400 samples are below the threshold line. Meaning all samples were diagnosed correctly as SARS-CoV-2 negative.

Figure 4

(a) Signature analysis of 400 SARS-CoV samples. The 242 out of 400 samples were misclassified as SARS-CoV-2 positive and the rest of the samples falls below the threshold line (b) Signature analysis of 400 HCoV-OC43 samples. Most of the samples were correctly classified as SARS-CoV-2 negative. Only 9 samples were classified as SARS-CoV-2 positive.

Figure 5

(a) Signature analysis of 400 MERS-CoV samples, while 356 samples were above the threshold and classified as SARS-CoV-2, 44 samples were under the threshold. (b) Signature analysis of 400 Influenza virus samples, only 14 samples are identified as SARS-CoV-2 positive.

Figure 6

Confusion matrix of manual data analysis results with 2% difference threshold rule. The true-positive rate is 100% since the threshold is set for a 100% sensitivity rate. However, due to the high false-positive rate, the overall accuracy is 74.1%, precision is 39.1% and F1 score is 56.3%.

Figure 7

Machine learning algorithm result comparison for different sets of features of ABC, DTC, MLPC, and SVC algorithms. The DTC algorithm outperformed all other algorithms with the feature set of F0-F2-F3-F6-F10-F13 by achieving 96.6% overall accuracy.

Figure 8

Confusion matrix results for DTC algorithm with an overall accuracy of 96.6%.

Figure 9

The CNN network parameters are tuned for optimal performance.

Figure 10

Accuracy and standard deviation results of CNN algorithm with the different time window of the data. The results showed that the highest accuracy with the lowest variation was achieved by using the 0–50 ms portion of the signals.

Figure 11

Confusion matrix results for the CNN algorithm with an overall accuracy of 97.20%, specificity of 98.17%, and sensitivity of 96.15%.

Figure 12

Performance metric comparison between DTC and CNN to diagnose SARS-CoV-2 with their best performing parameters. While the CNN algorithm outperforms the DTC algorithm in accuracy, precision, sensitivity, specificity, and F1 score.