*5.3. Spectroscopy*

Of the spectroscopic biosensing techniques, surface-enhanced Raman spectroscopy (SERS) has shown grea<sup>t</sup> success [151,152]. SERS is a vibrational surface sensing technique that enhances Raman scattering based on surface characteristics. Briefly, SERS utilizes incident laser light to induce inelastic scattering (Raman scattering) from the target analyte. The intensity of the Raman scattering is enhanced by interaction with the conduction electrons of metal nanostructures (SERS substrate). The enhancement of the Raman scattering is what makes SERS so sensitive. Researchers have reported enhancement factors of up to ten or eleven orders of magnitude [153]. Figure 11 illustrates a SERS sensor for the analysis of breath volatile organic compound (VOC) biomarkers [154]. Due to the complex nature of the obtained spectral signal, various machine learning algorithms have been used to process SERS data in multiple contexts [28].

**Figure 11.** SERS sensor for analysis of breath VOC biomarkers utilizing AuNPs. Reprinted with permission from [154] without modification. Copyright 2016 American Chemical Society.

Although bioreceptors may be used to allow for specific binding of the target analyte to the SERS sensing surface [155,156], direct detection is also possible. Robust classification and regression algorithms can bring specificity and sensitivity to these biosensors. A simple ye<sup>t</sup> effective method for SERS based quantification is partial least squares regression (PLSR). PLSR has been used for a variety of quantification applications such as biofilm formation monitoring [69], blood serum methotrexate concentration [63], aquaculture toxins [62], and food antiseptics [66]. PLSR has the advantage of model simplicity with well-defined parameters, but it may be insufficient in modeling data with significant sources of noise.

Since the spectra have high dimensionality, dimension reduction is a frequent preprocessing step (Figure 12). PCA is again popularly used as a dimension reduction or feature extraction step [60,61,64,65,68,70,71,73], or for exploratory analysis [62,72,157]. Once the spectra are remapped using PCA, a classifier or regression model is employed such as an extreme learning machine (ELM) [71], LDA [68], SVM [60,64,73], PLSR [65], or ANN [70]. An alternative to dimension reduction is utilizing the high dimensionality spectral data directly with a node-based algorithm such as ANN [72,158,159] and CNN [160,161].

**Figure 12.** PCA results using the spectral range of 400–1700 cm<sup>−</sup><sup>1</sup> of 112 average SERS spectra from 14 different commercially available pollen species. Loadings of the first four PCs (**a**) as well as the scores of the first and second (**b**), first and third (**c**), and first and fourth PC (**d**) are shown. PCA was done with standardized first derivatives of the mean spectra of 500 vector-normalized spectra. Reprinted with permission from [72] without modification. Copyright 2016 John Wiley and Sons.

The reusability and generalizability of the trained models are often limited. Spectral response is affected not just by analyte presence but surface structure. Therefore, for the model to be reused on a new SERS biosensing dataset, the surface characteristics must be very similar. In terms of transfer learning, this is an issue of changes in the underlying data distributions. However, if the surface structure methods are well documented and reproducible, transfer learning could be employed on a spectral library [28]. Ideally, researchers could contribute to this library in an open-access manner and use these spectra for model training. In this case, the quality of the attached metadata would be a crucial factor.

Clearly, machine learning has been used extensively in the context of SERS sensors. The most common pipeline is to perform unsupervised dimensionality reduction/feature extraction for which PCA is generally the preferred method. Less consistency is seen in the algorithms used for classification and regression. Alternatively, ANNs can be used directly on the data, and the advantage of one approach over the other is not clearly illustrated

in the literature. We anticipate, however, that like in the case of electrochemical sensors, node-based models would allow for more efficient transfer learning to accommodate target task change.

#### *5.4. Summary of Optical Bioreceptor-Free Biosensing*

A variety of optical sensing methods have benefited from machine learning techniques, with the preferred method being dependent on the data type. For image type data, CNN is the most obvious choice for its ability to detect features as well as reconstruct images obtained by lensless systems. For spectral data, the approach is similar to spectral data obtained with electrochemical sensors. In those instances, dimensionality reduction coupled with a classification/regression algorithm may perform nearly as well as node-based methods. Indeed, they may be preferable in instances where the quantity of training data is small.

#### **6. Considerations and Future Perspectives**

Biosensor research has shown grea<sup>t</sup> success and promise. For both systems with and without bioreceptor, ML has demonstrated huge success in going from large, complex sensor datasets to getting meaningful measurements and classification of analytes. However, in many of these systems, a key challenge is consistency in device manufacturing. This manifests itself regarding sensor reproducibility for Enose and Etongue, or as substrate reproducibility for SERS. Since the models used to process these data often rely on subtle signals in the data, even small changes in sensor response characteristics can lead to poor performance. These issues have effectively limited widespread commercial adoption of these technologies. There has been some success in accommodating these inconsistencies through computational methods, notably with transfer learning for Enose. More work, both from a manufacturing and computational standpoint, needs to be done before many of these systems are robust enough for widespread adoption.

One area in which these systems have pushed to increase commercial potential is through miniaturization and modularity. There have been efforts with several of the methods presented here to develop compact standalone devices that rival their bulkier counterparts in terms of performance [16,47,162–166]. We believe that cloud computing may be a key element to the success of these endeavors. Some of the models in use, especially for image-based sensors, are computationally expensive. By offloading the computational work to cloud computing, the device footprint imposed by processing and memory needs is greatly reduced.

A central question is what the relative advantages and disadvantages are between systems that utilize a bioreceptor and those that do not. A key advantage of those that eliminate the bioreceptor addresses one of the barriers to commercialization—manufacture variability. By eliminating the bioreceptor, device manufacture is simplified, and may decrease manufacture variability. Additionally, sensor longevity is generally improved because the long-term stability of the bioreceptor is often limited [6]. However, to match LOD and specificity of bioreceptors, improvements must be made. Nanomaterials show promise for improving device performance [167].

There have been studies that attempt to gain the advantages of both systems by creating artificial bioreceptors, notably nanomaterials with enzymatic properties referred to as nanozymes [168,169]. While exciting progress has been made in this field, current nanozyme-based biosensors have inferior catalytic activity and specificity to their biological alternatives [170,171]. Nanozyme catalytic activity is also currently limited to oxidase-like activity [171]. If researchers can broaden nanozyme activity and improve selectivity, these biosensors may become a competitive alternative for biological bioreceptors.

In addition to device considerations, there are computational challenges to consider. Although some ML algorithms have been in use for decades such as PCA and SVM, the field of ML is advancing rapidly with new algorithms being described frequently. While many areas are quick to adopt the new methods, improper usage is common and certainly not limited to biosensing. Some common mistakes are inappropriate data splitting, hidden variables serving as bad predictors, and mistaking the objective of the model [172]. Great emphasis must be placed on the importance of reporting appropriate performance metrics. A grea<sup>t</sup> example of a misleading metric is reporting accuracy on highly imbalanced data such as in Durante et al. [41]. It can often be difficult to determine if the proper preprocessing and model assumption checks are being performed. This may be centering and re-scaling prior to PCA, or normality checks for LDA.

Some of these issues can be solved with better methods reporting, especially regarding computational methods. Certain key details are frequently left out, making critical evaluation difficult and reproducibility impossible [173]. One of the most striking examples from the literature described herein is reporting classification metrics, without reporting what classifier was used on PCA processed data [49]. Perhaps the best way to make methods clear and reproducible is to release all associated code, preferably publicly.

Increased availability in general can greatly improve this field. More open access repositories of training sets may allow researchers to improve model robustness by exposing them to more diverse datasets [16]. Some examples currently exist such as the gas sensor drift dataset [115] and the EIS breast tissue dataset [103], both available in the UCI Machine Learning repository [104]. One vision would be to have large repositories of gas sensor responses to many analytes under various experimental conditions. Models could be trained on such repositories to improve generalizability. Ideally, with such repositories and improved manufacturing consistency, trained models could be shared directly and need only minimal recalibration.
