Machine Learning Techniques for Effective Pathogen Detection based on Resonant Biosensors

1. Introduction

The global COVID pandemic has caused huge impact on world health and economy [1]. The fast-spreading virus SARS-CoV-2 is the main culprit, and detection of the virus in human population is crucial for curbing the epidemic [2]. Traditional detection approaches are mainly Nucleic Acid Amplification Test (NAAT) [3] and antigen detection [4] techniques. Currently, the mainstream is quantitative Polymerase Chain Reaction (qPCR) [5] which is a kind of NAAT that has high sensitivity and specificity, but requires clean environment, bulky and expensive equipment, and trained personnel. Therefore, qPCR is not suitable for onsite, fast turnaround detection, or population scale screening, which are often required in pandemic control scenarios [6]. To complement qPCR, antigen detection based on lateral flow [7] has also been employed in home use and self-test. However, antigen detection is limited in detection sensitivity and specificity, hindering its efficacy in fighting a pandemic [8]. There still lacks rapid, accurate and low cost detection techniques that can be deployed onsite for population scale epidemic screening and/or surveillance [9], especially for regions of limited resources [10].

Biosensors have been proposed for detection of SARS-CoV-2 [11]. Biosensor technologies have the advantages of high sensitivity, good specificity, fast turnaround, ease of operation, low cost, and onsite deployment capability [12, 13]. We have previously proposed a photonic biosensor for fast onsite detection of SARS-CoV-2 with high sensitivity and specificity [14, 15]. The biosensor is based on nanoporous silicon material fabricated via CMOS-compatible silicon process, and nanophotonic working principles of Localized Surface Plasmon Resonance (LSPR) [16] and Tamm Plasmon Polariton (TPP) [17, 18]. The measurement of the biosensor is based on reflection spectroscopy [14].

We have also developed handheld and high throughput detection systems [19] that can collect the refection spectrum of biosensors and process the spectral data to determine the detection results efficiently. The high throughput detection system is suitable for populations scale screening of infection, and the handheld detection system is for home use or self-test. The spectral data processing algorithm works by recognizing the characteristic resonant valleys in the reflection spectrum of the biosensor and determines the detection result by judging if there is spectral red shift of the characteristic resonant valleys. This is the often used and so called “find peaks” technique, with its name originating from the MATLAB function find_peaks(). This technique can also be implemented on Field Programmable Gate Array (FPGA) for fast and efficient processing of signals from array of biosensors [20]. In addition, researchers have also proposed Interferogram Average over Wavelength (IAW) technique to process signals of optical biosensors that depend on spectral shift of characteristic resonant features, which can achieve sensitivity enhancement compared with spectral shift detection [21]. Detection of change in reflection intensity due to shift of spectral features in spectrum has also been used to detect biomolecules in real time [22]. However, both IAW and light intensity measurement techniques are subject to spectral amplitude fluctuations, and thus requires highly stable spectroscopy systems, such as stable light source and high signal-to-noise ratio spectrometer.

In this work, we demonstrate that it is advantageous to utilize artificial intelligence technology, more specifically machine learning (ML) algorithm to process the spectral data of the biosensor [23]. Instead of depending on programming, its algorithm is learnt from big volume of data [24]. Machine learning has been used for computer vision [25], face recognition [26], autonomous driving [27], auxiliary decision-making [28], brain-machine interface [29], and games [30]. It includes supervised learning, unsupervised learning, and reinforcement learning [31]. Supervised learning (SL) is an algorithm that learns from massive, labeled data and generates prediction models that can work to generate labels for new dataset. SL includes Support Vector Machine (SVM) [32], Multilayer Perceptron (MLP) [33], Linear Regression [34], Linear Discriminant Analysis [35], K-nearest Neighbor [36], Decision Tree [37], and Naïve Bayes [38]. In this work, we demonstrate that SVM and MLP can be used for processing of the photonic biosensor signal and dataset. Compared with previously proposed techniques, ML technique has the advanteges of : 1) no need to find appropriate parameters of the algorithm, e.g. the find_peaks() function, in a try-and-error way to guarantee accurate recognition of spectral feature; 2) no need to discriminate between redshift or blueshift which can be an extra issue in algorithm design; 3) not sensitive to spectral amplitude fluctuations so that requirements on stable and expensive hardware are relaxed; 4) generalizable to all kinds of sensors with salient features in response signal which serve as the basis of discriminating between positive and negatibe responses.

Data visualization approach can help us to understand the distribution of dataset and find out the distinguishability of the dataset. T-distributed Stochastic Neighbor Embedding (t-SNE) is a prevalent approach to map high-dimensional data to low-dimensional embedding [39]. In this contribution, we also implemented t-SNE approach on the SARS-CoV-2 detection dataset to clarify the distinguishability of the biosensor dataset so that a better understanding of the data processing and ML prediction performances can be obtained.

2. Materials and Methods

2.1. Biosensor Working Principal and Measurement Setup

As shown in Figure 1(a), the biosensor is basically a porous silicon microcavity consisting of two Bragg reflectors and one resonant cavity [14, 40].

One Bragg reflector is six periods of alternating low porosity (LP) and high porosity (HP) porous silicon (PSi) thin films of thickness equal to quarter resonant wavelength. On top of the porous silicon is deposited noble metal thin film. Because of the nanoporous structure of the porous silicon material, the conformally deposited noble metal thin film is also porous. When light is incident on the surface of the biosensor, some of its energy will couple into Localized Surface Plasmon Resonance (LSPR) supported [40] by the nanostructures of the noble metal thin film. In addition, some of its energy will also couple into Tamm Plasmon Polariton (TPP) supported by the interface between the top Bragg reflector and the noble metal thin film [41]. Therefore, LSPR and TPP are excited by the incident light simultaneously, coupled with each other, and form strong field confinement around the noble metal thin film. If specific antibodies are immobilized beforehand on the surface of the noble metal, they can capture SARS-CoV-2 virus specifically. Such binding events will cause addition of biomaterials around the noble metal thin film, and the added biomaterial will interact strongly with the LSPR and TPP coupled field. This is the working mechanism of the biosensor for sensitive detection of virus. As shown in Figure 1(a), in order to measure the signals of the biosensor, reflection spectroscopy is used. White light source provides incident light, which goes through the Y-shape fiber and shine vertically onto the biosensor surface. The light reflected from the biosensor surface is collected by the Y-shape fiber and goes into the spectrometer for data analysis. The Y-shape fiber consists of six circumferential fibers guiding incident light, and one central fiber guiding reflected light.

Figure 1(b) and 1(c) show representative reflection spectra of the biosensor. They have characteristic resonant valleys in the spectral range of 600-800 nm in wavelength. If there are viruses binding with antibodies on the biosensor surface, the binding events will cause shift of the spectral features to longer wavelength, which is also called “redshift”. For example, Figure 1(b) shows such a case where virus binds with antibodies, redshift occurs and the detection result is determined to be positive. On the other hand, if there is no virus binding with antibodies on the biosensor surface, there will be no shift of the spectral features, i.e., almost overlapping spectra for both before and after binding reaction. In the third case, there could also appear shift of the resonant features to shorter wavelength, which is also called “blueshift”. In such case, the detection result is determined to be negative. Figure 1(c) shows an example wherein there appears blueshift. In summary, the principle of the biosensor is based on interactions between biomaterials and photonic energy, and the detection result is determined based on shift of spectral features in optical spectrum collected from reflection spectroscopy measurement.

2.2. Data Preprocessing

The dataset was obtained from detection experiments of inactivated SARS-CoV-2 in clinical swab specimens, with virus concentration as low as 1 TCID₅₀/ml [14]. Figure 1 shows example spectra of the biosensor for positive and negative detection results. For positive result, there is spectral redshift; and for negative result, there is no spectral shift or there is spectral blueshift. The experimental data is collected by reflection spectroscopy with corresponding spectra for before and after applying specimens on biosensor surface. Each spectral data contains 2048 data points representing reflection intensities, with data-to-data spacing of 0.48 nm in the wavelength range of 200-1200 nm. We usually need to carry out preprocessing of the spectral data before data analysis, such as normalization and artifacts removal. Furthermore, normalization was implemented on each data sample for the purpose of training convenience. Spectral data of both before and after adding specimens are combined as a single sample, so that the size of the reformed sample is 2×2048, or 4096. Each detection experiment is regarded as a sample for either training or test purpose. After several outliers were removed to clean the dataset, there are 486 negative samples and 108 positive samples left in total for the classification model training and prediction test.

2.3. Feature Engineering

As shown in Figure 2, the input to the model could be the data of size 4096. This requires 4096 input neuron nodes which could be a computation burden. In addition to this raw data approach, the input could also be features extracted from the data. We propose feature engineering methods comprising three different approaches – wavelet transform, Fourier transform, and spectral difference. As for wavelet domain, we used the wavelet transform with scales of 30 and take average of each scale, which generates 30 features for each spectral curve, two curves (before and after virus) would generate 60 wavelet-based features. In terms of Fourier domain, we have found that most information appears in the low frequency range (< 50Hz), so that we took average of each 5Hz in order from 0 to 50Hz, so that 10 features for each spectral curve and 20 features for spectra pairs would be obtained in Fourier domain. For spectral difference, we utilized the difference between the spectral data of before and after binding reaction on biosensor, instead of two separate spectra. There are three features selected from spectral difference, they are mean, variance and sign change rate.

Eventually, for each training sample containing spectral data of before and after reaction, wavelet transform and Fourier transform domain features need spectra of both before and after reaction, and spectral difference features only need difference between spectra of before and after reaction. Therefore, there are 83 (60 wavelet domain + 20 Fourier domain + 3 spectral difference) features selected for the classification experiments.

2.4. Classification Models

All samples were randomly shuffled and separated as 70% for training and 30% for test. This allocation ratio is a practical standard for benchmark performance. Multilayer Perceptron (MLP) and Support Vector Machine (SVM) models were used since they are usually considered as efficient ML models capable of achieving baseline performance. As shown in Figure 3, in terms of MLP model, two hidden layers with 100 and 50 neuron nodes are implemented, optimizer is stochastic gradient decent solver, learning rate and epoch is set as 0.1 and 30, respectively. As for SVM model, we set the gamma parameter of radial basis kernel function as 1.

2.5. Control Experiments

For control experiments, we detected SARS-CoV-2 specimens with photonic biosensors which do not have specific antibodies immobilized on biosensor surface beforehand. There are in total 732 data samples of detecting SARS-CoV-2 virus specimens of various concentrations, and 36 data samples of detecting specimens containing no SARS-CoV-2 viruses. This new dataset is processed by the SVM and MLP models already trained as shown in Figure 3.

2.6. Dataset Distinguishability Analysis

Nowadays, data visualization approach can help understand the distribution of the dataset and intuitively investigate whether the dataset is distinguishable or not. T-distributed Stochastic Neighbor Embedding (t-SNE) is a tool to visualize high-dimensional data. It converts similarities between data points to joint probabilities and tries to minimize the Kullback-Leibler divergence [42] between the joint probabilities of the low-dimensional embedding and the high-dimensional data.

We implemented t-SNE tool on specimen detection dataset to interpret the distinguishability of the dataset. The data distribution patterns can help interpret performances of models on the dataset. Both raw dataset and features extracted from raw data are considered of their distinguishability. We also investigated whether extracted features have distributions different from that of raw data.

3. Results and Discussion

In terms of the experiments, we used SVM and MLP models to test the raw data processing and feature engineering method. Two performance metrics are considered in the experiments: sensitivity (SEN) and specificity (SPE) which are defined as

$$\mathrm{S}\mathrm{E}\mathrm{N}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}$$

(1)

$$\mathrm{S}\mathrm{P}\mathrm{E}=\frac{\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}}$$

(2)

where TP, FN, TN, FP stand for true positive, false negative, true negative and false positive, respectively.

Table 1 shows the performances of ML model predictions. We can see that perfect performances are achieved for both raw data and feature engineering methods, combined with either SVM or MLP model. The last row in Table 1 shows the performance of the models in processing the control experiment dataset. The performance is very poor, and this is due to the fact that the biosensors have not been functionalized with specific antibodies and thus, cannot detect SARS-CoV-2 virus effectively.

Figure 4 (a) shows data distribution of raw dataset in 2D space by t-SNE data visualization approach. We can see that the positive and negative samples from dataset of valid detection experiments are clustered without any overlapping. Thus, the valid experimental dataset is distinguishable. Figure 4 (b) shows data distribution of features extracted from the dataset of Figure 4 (a). The extracted features change the data distribution, while maintaining the distinguishability because the samples are separated into different clusters. Figure 4(c) shows the data distribution of dataset obtained from control experiments wherein biosensors are not functionalized with specific antibodies. Negative samples are overlaping with positive samples, and the dataset is indistinguishable according to the visualization results. Figure 4 (d) shows data distribution of features extracted from the dataset of Figure 4 (c). The distribution of features’ dataset is still mixed up, so that feature engineering cannot help the dataset to be classified effectively. These dataset distribution results could serve to interpret the performance comparisons demonstrated in Table 1.

Table 2 demonstrates the advantages of ML data processing technique when compared with other techniques. It can be seen that the general advtantages of ML are valid, in addition to eased hardware requirement.

To verify the efficacy of the ML data processing technique for biosensors, detection experiments of inactivated SARS-CoV-2 in vaccination sites of Hangzhou Center for Disease Control and Prevention (CDC) were carried out and the detection results are compared with the gold standard-reverse transcription qPCR technique. The envrionmental specimens were collected from various locations in different vaccination sites, delivered to Hangzhou CDC within 4 hours,, and were simultaneously analyzed by both techniques. Table 3 shows that biosensors, together with ML data processing, generate detection results that are consistent with qPCR results. Note that qPCR provides semi-quantitative results dependent on the Ct value [5], while ML processing of biosensor data only provides qualitative results. This comparative study demonstrates that the ML technique is an effective tool for biosensor signal and data processing,

4. Conclusion

In this work, machine learning techniques have been used to process the signals and dataset of photonic biosensors. Both SVM and MLP have been used to process raw data and future engineering data, and perfect results have been obtained to distinguish between negative and positive detections. Control experiments have also been carried out wherein biosensors not functionalized with specific antibodies are used to detect SARS-CoV-2 virus. Both SVM and MLP models trained with valid experimental data cannot distinguish between negative and positive detections in control experiments. To demonstrate the distinguishability of the raw data and the future engineering data for both valid experiments and control experiments, we implemented t-SNE data visualization approach. Results show that the valid experimental dataset is distinguishable, and the control experimental dataset is indistinguishable according to both raw data and features engineering methods. The results are consistent with the data processing performances of machine learning techniques achieved for valid experimental dataset and control experimental dataset. Future research will focus on ML techniques for determination of quantitative detection results so that the quantity of target biospecies in specimen can be obtained. ML can be a powerful tool in processing signals and dataset of biosensors for which there are salient features in the response signals of such biosensors, such as optical, electrochemical, thermal, and mechanical biosensors.